Machine Learning (cs.LG)

Abstract: In this paper, we study the setting in which data owners train machine learning models collaboratively under a privacy notion called joint differential privacy [Kearns et al., 2018]. In this setting, the model trained for each data owner $j$ uses $j$'s data without privacy consideration and other owners' data with differential privacy guarantees. This setting was initiated in [Jain et al., 2021] with a focus on linear regressions. In this paper, we study this setting for stochastic convex optimization (SCO). We present an algorithm that is a variant of DP-SGD [Song et al., 2013; Abadi et al., 2016] and provides theoretical bounds on its population loss. We compare our algorithm to several baselines and discuss for what parameter setups our algorithm is more preferred. We also empirically study joint differential privacy in the multi-class classification problem over two public datasets. Our empirical findings are well-connected to the insights from our theoretical results.

Abstract: Several recent studies have elucidated why knowledge distillation (KD) improves model performance. However, few have researched the other advantages of KD in addition to its improving model performance. In this study, we have attempted to show that KD enhances the interpretability as well as the accuracy of models. We measured the number of concept detectors identified in network dissection for a quantitative comparison of model interpretability. We attributed the improvement in interpretability to the class-similarity information transferred from the teacher to student models. First, we confirmed the transfer of class-similarity information from the teacher to student model via logit distillation. Then, we analyzed how class-similarity information affects model interpretability in terms of its presence or absence and degree of similarity information. We conducted various quantitative and qualitative experiments and examined the results on different datasets, different KD methods, and according to different measures of interpretability. Our research showed that KD models by large models could be used more reliably in various fields.

Abstract: Explaining the decisions made by machine learning models for high-stakes applications is critical for increasing transparency and guiding improvements to these decisions. This is particularly true in the case of models for graphs, where decisions often depend on complex patterns combining rich structural and attribute data. While recent work has focused on designing so-called post-hoc explainers, the question of what constitutes a good explanation remains open. One intuitive property is that explanations should be sufficiently informative to enable humans to approximately reproduce the predictions given the data. However, we show that post-hoc explanations do not achieve this goal as their explanations are highly dependent on fixed model parameters (e.g., learned GNN weights). To address this challenge, this paper proposes RAGE (Robust Ante-hoc Graph Explainer), a novel and flexible ante-hoc explainer designed to discover explanations for a broad class of graph neural networks using bilevel optimization. RAGE is able to efficiently identify explanations that contain the full information needed for prediction while still enabling humans to rank these explanations based on their influence. Our experiments, based on graph classification and regression, show that RAGE explanations are more robust than existing post-hoc and ante-hoc approaches and often achieve similar or better accuracy than state-of-the-art models.

Abstract: Graph Neural Networks (GNNs) are widely used for graph representation learning in many application domains. The expressiveness of vanilla GNNs is upper-bounded by 1-dimensional Weisfeiler-Leman (1-WL) test as they operate on rooted subtrees through iterative message passing. In this paper, we empower GNNs by injecting neighbor-connectivity information extracted from a new type of substructure. We first investigate different kinds of connectivities existing in a local neighborhood and identify a substructure called union subgraph, which is able to capture the complete picture of the 1-hop neighborhood of an edge. We then design a shortest-path-based substructure descriptor that possesses three nice properties and can effectively encode the high-order connectivities in union subgraphs. By infusing the encoded neighbor connectivities, we propose a novel model, namely Union Subgraph Neural Network (UnionSNN), which is proven to be strictly more powerful than 1-WL in distinguishing non-isomorphic graphs. Additionally, the local encoding from union subgraphs can also be injected into arbitrary message-passing neural networks (MPNNs) and Transformer-based models as a plugin. Extensive experiments on 17 benchmarks of both graph-level and node-level tasks demonstrate that UnionSNN outperforms state-of-the-art baseline models, with competitive computational efficiency. The injection of our local encoding to existing models is able to boost the performance by up to 11.09%.

Abstract: Time series prediction is a prevalent issue across various disciplines, such as meteorology, traffic surveillance, investment, and energy production and consumption. Many statistical and machine-learning strategies have been developed to tackle this problem. However, these approaches either lack explainability or exhibit less satisfactory performance when the prediction horizon increases. To this end, we propose a novel plan for the designing of networks' architecture based on transformations, possessing the potential to achieve an enhanced receptive field in learning which brings benefits to fuse features across scales. In this context, we introduce four different transformation mechanisms as bases to construct the learning model including Fourier Transform (FT), Singular Value Decomposition (SVD), matrix multiplication and Conv block. Hence, we develop four learning models based on the above building blocks, namely, FT-Matrix, FT-SVD, FT-Conv, and Conv-SVD. Note that the FT and SVD blocks are capable of learning global information, while the Conv blocks focus on learning local information. The matrix block is sparsely designed to learn both global and local information simultaneously. The above Transformation Learning Networks (TLNets) have been extensively tested and compared with multiple baseline models based on several real-world datasets and showed clear potential in long-range time-series forecasting.

Abstract: Attention mechanisms have greatly improved the performance of deep-learning models on visual, NLP, and multimodal tasks while also providing tools to aid in the model's interpretability. In particular, attention scores over input regions or concrete image features can be used to measure how much the attended elements contribute to the model inference. The recently proposed Concept Transformer (CT) generalizes the Transformer attention mechanism from such low-level input features to more abstract, intermediate-level latent concepts that better allow human analysts to more directly assess an explanation for the reasoning of the model about any particular output classification. However, the concept learning employed by CT implicitly assumes that across every image in a class, each image patch makes the same contribution to concepts that characterize membership in that class. Instead of using the CT's image-patch-centric concepts, object-centric concepts could lead to better classification performance as well as better explainability. Thus, we propose Concept-Centric Transformers (CCT), a new family of concept transformers that provides more robust explanations and performance by integrating a novel concept-extraction module based on object-centric learning. We test our proposed CCT against the CT and several other existing approaches on classification problems for MNIST (odd/even), CIFAR100 (super-classes), and CUB-200-2011 (bird species). Our experiments demonstrate that CCT not only achieves significantly better classification accuracy than all selected benchmark classifiers across all three of our test problems, but it generates more consistent concept-based explanations of classification output when compared to CT.

Abstract: Weakly supervised learning aims to empower machine learning when the perfect supervision is unavailable, which has drawn great attention from researchers. Among various types of weak supervision, one of the most challenging cases is to learn from multiple unlabeled (U) datasets with only a little knowledge of the class priors, or U$^m$ learning for short. In this paper, we study the problem of building an AUC (area under ROC curve) optimization model from multiple unlabeled datasets, which maximizes the pairwise ranking ability of the classifier. We propose U$^m$-AUC, an AUC optimization approach that converts the U$^m$ data into a multi-label AUC optimization problem, and can be trained efficiently. We show that the proposed U$^m$-AUC is effective theoretically and empirically.

Abstract: We introduce a differentiable clustering method based on minimum-weight spanning forests, a variant of spanning trees with several connected components. Our method relies on stochastic perturbations of solutions of linear programs, for smoothing and efficient gradient computations. This allows us to include clustering in end-to-end trainable pipelines. We show that our method performs well even in difficult settings, such as datasets with high noise and challenging geometries. We also formulate an ad hoc loss to efficiently learn from partial clustering data using this operation. We demonstrate its performance on several real world datasets for supervised and semi-supervised tasks.

Abstract: Ensembling is among the most popular tools in machine learning (ML) due to its effectiveness in minimizing variance and thus improving generalization. Most ensembling methods for black-box base learners fall under the umbrella of "stacked generalization," namely training an ML algorithm that takes the inferences from the base learners as input. While stacking has been widely applied in practice, its theoretical properties are poorly understood. In this paper, we prove a novel result, showing that choosing the best stacked generalization from a (finite or finite-dimensional) family of stacked generalizations based on cross-validated performance does not perform "much worse" than the oracle best. Our result strengthens and significantly extends the results in Van der Laan et al. (2007). Inspired by the theoretical analysis, we further propose a particular family of stacked generalizations in the context of probabilistic forecasting, each one with a different sensitivity for how much the ensemble weights are allowed to vary across items, timestamps in the forecast horizon, and quantiles. Experimental results demonstrate the performance gain of the proposed method.

Abstract: Graph neural networks (GNNs) have achieved remarkable success in various tasks, however, their vulnerability to adversarial attacks raises concerns for the real-world applications. Existing defense methods can resist some attacks, but suffer unbearable performance degradation under other unknown attacks. This is due to their reliance on either limited observed adversarial examples to optimize (adversarial training) or specific heuristics to alter graph or model structures (graph purification or robust aggregation). In this paper, we propose an Invariant causal DEfense method against adversarial Attacks (IDEA), providing a new perspective to address this issue. The method aims to learn causal features that possess strong predictability for labels and invariant predictability across attacks, to achieve graph adversarial robustness. Through modeling and analyzing the causal relationships in graph adversarial attacks, we design two invariance objectives to learn the causal features. Extensive experiments demonstrate that our IDEA significantly outperforms all the baselines under both poisoning and evasion attacks on five benchmark datasets, highlighting the strong and invariant predictability of IDEA. The implementation of IDEA is available at https://anonymous.4open.science/r/IDEA_repo-666B.

Abstract: In recent years, numerous screening methods have been published for ultrahigh-dimensional data that contain hundreds of thousands of features; however, most of these features cannot handle data with thousands of classes. Prediction models built to authenticate users based on multichannel biometric data result in this type of problem. In this study, we present a novel method known as random forest-based multiround screening (RFMS) that can be effectively applied under such circumstances. The proposed algorithm divides the feature space into small subsets and executes a series of partial model builds. These partial models are used to implement tournament-based sorting and the selection of features based on their importance. To benchmark RFMS, a synthetic biometric feature space generator known as BiometricBlender is employed. Based on the results, the RFMS is on par with industry-standard feature screening methods while simultaneously possessing many advantages over these methods.

Abstract: Exceptional text-to-image (T2I) generation results of Stable Diffusion models (SDMs) come with substantial computational demands. To resolve this issue, recent research on efficient SDMs has prioritized reducing the number of sampling steps and utilizing network quantization. Orthogonal to these directions, this study highlights the power of classical architectural compression for general-purpose T2I synthesis by introducing block-removed knowledge-distilled SDMs (BK-SDMs). We eliminate several residual and attention blocks from the U-Net of SDMs, obtaining over a 30% reduction in the number of parameters, MACs per sampling step, and latency. We conduct distillation-based pretraining with only 0.22M LAION pairs (fewer than 0.1% of the full training pairs) on a single A100 GPU. Despite being trained with limited resources, our compact models can imitate the original SDM by benefiting from transferred knowledge and achieve competitive results against larger multi-billion parameter models on the zero-shot MS-COCO benchmark. Moreover, we demonstrate the applicability of our lightweight pretrained models in personalized generation with DreamBooth finetuning.

Abstract: Accurate estimation of multiple quality variables is critical for building industrial soft sensor models, which have long been confronted with data efficiency and negative transfer issues. Methods sharing backbone parameters among tasks address the data efficiency issue; however, they still fail to mitigate the negative transfer problem. To address this issue, a balanced Mixture-of-Experts (BMoE) is proposed in this work, which consists of a multi-gate mixture of experts (MMoE) module and a task gradient balancing (TGB) module. The MoE module aims to portray task relationships, while the TGB module balances the gradients among tasks dynamically. Both of them cooperate to mitigate the negative transfer problem. Experiments on the typical sulfur recovery unit demonstrate that BMoE models task relationship and balances the training process effectively, and achieves better performance than baseline models significantly.

Abstract: A successful tactic that is followed by the scientific community for advancing AI is to treat games as problems, which has been proven to lead to various breakthroughs. We adapt this strategy in order to study Rocket League, a widely popular but rather under-explored 3D multiplayer video game with a distinct physics engine and complex dynamics that pose a significant challenge in developing efficient and high-performance game-playing agents. In this paper, we present Lucy-SKG, a Reinforcement Learning-based model that learned how to play Rocket League in a sample-efficient manner, outperforming by a notable margin the two highest-ranking bots in this game, namely Necto (2022 bot champion) and its successor Nexto, thus becoming a state-of-the-art agent. Our contributions include: a) the development of a reward analysis and visualization library, b) novel parameterizable reward shape functions that capture the utility of complex reward types via our proposed Kinesthetic Reward Combination (KRC) technique, and c) design of auxiliary neural architectures for training on reward prediction and state representation tasks in an on-policy fashion for enhanced efficiency in learning speed and performance. By performing thorough ablation studies for each component of Lucy-SKG, we showed their independent effectiveness in overall performance. In doing so, we demonstrate the prospects and challenges of using sample-efficient Reinforcement Learning techniques for controlling complex dynamical systems under competitive team-based multiplayer conditions.

Abstract: Heterogeneous Graph Neural Networks (HGNNs) are a class of powerful deep learning methods widely used to learn representations of heterogeneous graphs. Despite the fast development of HGNNs, they still face some challenges such as over-smoothing, and non-robustness. Previous studies have shown that these problems can be reduced by using gradient regularization methods. However, the existing gradient regularization methods focus on either graph topology or node features. There is no universal approach to integrate these features, which severely affects the efficiency of regularization. In addition, the inclusion of gradient regularization into HGNNs sometimes leads to some problems, such as an unstable training process, increased complexity and insufficient coverage regularized information. Furthermore, there is still short of a complete theoretical analysis of the effects of gradient regularization on HGNNs. In this paper, we propose a novel gradient regularization method called Grug, which iteratively applies regularization to the gradients generated by both propagated messages and the node features during the message-passing process. Grug provides a unified framework integrating graph topology and node features, based on which we conduct a detailed theoretical analysis of their effectiveness. Specifically, the theoretical analyses elaborate the advantages of Grug: 1) Decreasing sample variance during the training process (Stability); 2) Enhancing the generalization of the model (Universality); 3) Reducing the complexity of the model (Simplicity); 4) Improving the integrity and diversity of graph information utilization (Diversity). As a result, Grug has the potential to surpass the theoretical upper bounds set by DropMessage (AAAI-23 Distinguished Papers). In addition, we evaluate Grug on five public real-world datasets with two downstream tasks...

Abstract: Deep Neural Networks (DNNs) generalization is known to be closely related to the flatness of minima, leading to the development of Sharpness-Aware Minimization (SAM) for seeking flatter minima and better generalization. In this paper, we revisit the loss of SAM and propose a more general method, called WSAM, by incorporating sharpness as a regularization term. We prove its generalization bound through the combination of PAC and Bayes-PAC techniques, and evaluate its performance on various public datasets. The results demonstrate that WSAM achieves improved generalization, or is at least highly competitive, compared to the vanilla optimizer, SAM and its variants. The code is available at https://github.com/intelligent-machine-learning/dlrover/tree/master/atorch/atorch/optimizers.

Abstract: Graph Neural Networks (GNNs) have emerged as a powerful tool for semi-supervised node classification tasks. However, recent studies have revealed various biases in GNNs stemming from both node features and graph topology. In this work, we uncover a new bias - label position bias, which indicates that the node closer to the labeled nodes tends to perform better. We introduce a new metric, the Label Proximity Score, to quantify this bias, and find that it is closely related to performance disparities. To address the label position bias, we propose a novel optimization framework for learning a label position unbiased graph structure, which can be applied to existing GNNs. Extensive experiments demonstrate that our proposed method not only outperforms backbone methods but also significantly mitigates the issue of label position bias in GNNs.

Abstract: EXplainable Artificial Intelligence (XAI) aims to help users to grasp the reasoning behind the predictions of an Artificial Intelligence (AI) system. Many XAI approaches have emerged in recent years. Consequently, a subfield related to the evaluation of XAI methods has gained considerable attention, with the aim to determine which methods provide the best explanation using various approaches and criteria. However, the literature lacks a comparison of the evaluation metrics themselves, that one can use to evaluate XAI methods. This work aims to fill this gap by comparing 14 different metrics when applied to nine state-of-the-art XAI methods and three dummy methods (e.g., random saliency maps) used as references. Experimental results show which of these metrics produces highly correlated results, indicating potential redundancy. We also demonstrate the significant impact of varying the baseline hyperparameter on the evaluation metric values. Finally, we use dummy methods to assess the reliability of metrics in terms of ranking, pointing out their limitations.

Abstract: The generalization of neural networks is a central challenge in machine learning, especially concerning the performance under distributions that differ from training ones. Current methods, mainly based on the data-driven paradigm such as data augmentation, adversarial training, and noise injection, may encounter limited generalization due to model non-smoothness. In this paper, we propose to investigate generalization from a Partial Differential Equation (PDE) perspective, aiming to enhance it directly through the underlying function of neural networks, rather than focusing on adjusting input data. Specifically, we first establish the connection between neural network generalization and the smoothness of the solution to a specific PDE, namely ``transport equation''. Building upon this, we propose a general framework that introduces adaptive distributional diffusion into transport equation to enhance the smoothness of its solution, thereby improving generalization. In the context of neural networks, we put this theoretical framework into practice as PDE+ (\textbf{PDE} with \textbf{A}daptive \textbf{D}istributional \textbf{D}iffusion) which diffuses each sample into a distribution covering semantically similar inputs. This enables better coverage of potentially unobserved distributions in training, thus improving generalization beyond merely data-driven methods. The effectiveness of PDE+ is validated in extensive settings, including clean samples and various corruptions, demonstrating its superior performance compared to SOTA methods.

Abstract: This work presents a novel approach to tabular data prediction leveraging graph structure learning and graph neural networks. Despite the prevalence of tabular data in real-world applications, traditional deep learning methods often overlook the potentially valuable associations between data instances. Such associations can offer beneficial insights for classification tasks, as instances may exhibit similar patterns of correlations among features and target labels. This information can be exploited by graph neural networks, necessitating robust graph structures. However, existing studies primarily focus on improving graph structure from noisy data, largely neglecting the possibility of deriving graph structures from tabular data. We present a novel solution, Tabular Graph Structure Learning (TabGSL), to enhance tabular data prediction by simultaneously learning instance correlation and feature interaction within a unified framework. This is achieved through a proposed graph contrastive learning module, along with transformer-based feature extractor and graph neural network. Comprehensive experiments conducted on 30 benchmark tabular datasets demonstrate that TabGSL markedly outperforms both tree-based models and recent deep learning-based tabular models. Visualizations of the learned instance embeddings further substantiate the effectiveness of TabGSL.

Abstract: Dropout is a widely utilized regularization technique in the training of neural networks, nevertheless, its underlying mechanism and its impact on achieving good generalization abilities remain poorly understood. In this work, we derive the stochastic modified equations for analyzing the dynamics of dropout, where its discrete iteration process is approximated by a class of stochastic differential equations. In order to investigate the underlying mechanism by which dropout facilitates the identification of flatter minima, we study the noise structure of the derived stochastic modified equation for dropout. By drawing upon the structural resemblance between the Hessian and covariance through several intuitive approximations, we empirically demonstrate the universal presence of the inverse variance-flatness relation and the Hessian-variance relation, throughout the training process of dropout. These theoretical and empirical findings make a substantial contribution to our understanding of the inherent tendency of dropout to locate flatter minima.

Abstract: Off-policy learning (OPL) aims at finding improved policies from logged bandit data, often by minimizing the inverse propensity scoring (IPS) estimator of the risk. In this work, we investigate a smooth regularization for IPS, for which we derive a two-sided PAC-Bayes generalization bound. The bound is tractable, scalable, interpretable and provides learning certificates. In particular, it is also valid for standard IPS without making the assumption that the importance weights are bounded. We demonstrate the relevance of our approach and its favorable performance through a set of learning tasks. Since our bound holds for standard IPS, we are able to provide insight into when regularizing IPS is useful. Namely, we identify cases where regularization might not be needed. This goes against the belief that, in practice, clipped IPS often enjoys favorable performance than standard IPS in OPL.

Abstract: Electronic health records (EHR) often contain different rates of representation of certain subpopulations (SP). Factors like patient demographics, clinical condition prevalence, and medical center type contribute to this underrepresentation. Consequently, when training machine learning models on such datasets, the models struggle to generalize well and perform poorly on underrepresented SPs. To address this issue, we propose a novel ensemble framework that utilizes generative models. Specifically, we train a GAN-based synthetic data generator for each SP and incorporate synthetic samples into each SP training set. Ultimately, we train SP-specific prediction models. To properly evaluate this method, we design an evaluation pipeline with 2 real-world use case datasets, queried from the MIMIC database. Our approach shows increased model performance over underrepresented SPs. Our code and models are given as supplementary and will be made available on a public repository.

Abstract: In this work, we present GAROM, a new approach for reduced order modelling (ROM) based on generative adversarial networks (GANs). GANs have the potential to learn data distribution and generate more realistic data. While widely applied in many areas of deep learning, little research is done on their application for ROM, i.e. approximating a high-fidelity model with a simpler one. In this work, we combine the GAN and ROM framework, by introducing a data-driven generative adversarial model able to learn solutions to parametric differential equations. The latter is achieved by modelling the discriminator network as an autoencoder, extracting relevant features of the input, and applying a conditioning mechanism to the generator and discriminator networks specifying the differential equation parameters. We show how to apply our methodology for inference, provide experimental evidence of the model generalisation, and perform a convergence study of the method.

Abstract: Domain generalization (DG) is a prevalent problem in real-world applications, which aims to train well-generalized models for unseen target domains by utilizing several source domains. Since domain labels, i.e., which domain each data point is sampled from, naturally exist, most DG algorithms treat them as a kind of supervision information to improve the generalization performance. However, the original domain labels may not be the optimal supervision signal due to the lack of domain heterogeneity, i.e., the diversity among domains. For example, a sample in one domain may be closer to another domain, its original label thus can be the noise to disturb the generalization learning. Although some methods try to solve it by re-dividing domains and applying the newly generated dividing pattern, the pattern they choose may not be the most heterogeneous due to the lack of the metric for heterogeneity. In this paper, we point out that domain heterogeneity mainly lies in variant features under the invariant learning framework. With contrastive learning, we propose a learning potential-guided metric for domain heterogeneity by promoting learning variant features. Then we notice the differences between seeking variance-based heterogeneity and training invariance-based generalizable model. We thus propose a novel method called Heterogeneity-based Two-stage Contrastive Learning (HTCL) for the DG task. In the first stage, we generate the most heterogeneous dividing pattern with our contrastive metric. In the second stage, we employ an invariance-aimed contrastive learning by re-building pairs with the stable relation hinted by domains and classes, which better utilizes generated domain labels for generalization learning. Extensive experiments show HTCL better digs heterogeneity and yields great generalization performance.

Abstract: Given samples from two joint distributions, we consider the problem of Optimal Transportation (OT) between the corresponding distributions conditioned on a common variable. The objective of this work is to estimate the associated transport cost (Wasserstein distance) as well as the transport plan between the conditionals as a function of the conditioned value. Since matching conditional distributions is at the core of supervised training of discriminative models and (implicit) conditional-generative models, OT between conditionals has the potential to be employed in diverse machine learning applications. However, since the conditionals involved in OT are implicitly specified via the joint samples, it is challenging to formulate this problem, especially when (i) the variable conditioned on is continuous and (ii) the marginal of this variable in the two distributions is different. We overcome these challenges by employing a specific kernel MMD (Maximum Mean Discrepancy) based regularizer that ensures the marginals of our conditional transport plan are close to the conditionals specified via the given joint samples. Under mild conditions, we prove that our estimator for this regularized transport cost is statistically consistent and derive finite-sample bounds on the estimation error. Application-specific details for parameterizing our conditional transport plan are also presented. Furthermore, we empirically evaluate our methodology on benchmark datasets in applications like classification, prompt learning for few-shot classification, and conditional-generation in the context of predicting cell responses to cancer treatment.

Abstract: In this paper, we studied two identically-trained neural networks (i.e. networks with the same architecture, trained on the same dataset using the same algorithm, but with different initialization) and found that their outputs discrepancy on the training dataset exhibits a "double descent" phenomenon. We demonstrated through extensive experiments across various tasks, datasets, and network architectures that this phenomenon is prevalent. Leveraging this phenomenon, we proposed a new early stopping criterion and developed a new method for data quality assessment. Our results show that a phenomenon-driven approach can benefit deep learning research both in theoretical understanding and practical applications.

Abstract: We examine the characteristic activation values of individual ReLU units in neural networks. We refer to the corresponding set for such characteristic activation values in the input space as the characteristic activation set of a ReLU unit. We draw an explicit connection between the characteristic activation set and learned features in ReLU networks. This connection leads to new insights into why various neural network normalization techniques used in modern deep learning architectures regularize and stabilize SGD optimization. Utilizing these insights, we propose a geometric approach to parameterize ReLU networks for improved feature learning. We empirically verify its usefulness with less carefully chosen initialization schemes and larger learning rates. We report improved optimization stability, faster convergence speed, and better generalization performance.

Abstract: Unsupervised disentanglement is a long-standing challenge in representation learning. Recently, self-supervised techniques achieved impressive results in the sequential setting, where data is time-dependent. However, the latter methods employ modality-based data augmentations and random sampling or solve auxiliary tasks. In this work, we propose to avoid that by generating, sampling, and comparing empirical distributions from the underlying variational model. Unlike existing work, we introduce a self-supervised sequential disentanglement framework based on contrastive estimation with no external signals, while using common batch sizes and samples from the latent space itself. In practice, we propose a unified, efficient, and easy-to-code sampling strategy for semantically similar and dissimilar views of the data. We evaluate our approach on video, audio, and time series benchmarks. Our method presents state-of-the-art results in comparison to existing techniques. The code is available at https://github.com/azencot-group/SPYL.

Abstract: Estimating the parameters of a probabilistic directed graphical model from incomplete data remains a long-standing challenge. This is because, in the presence of latent variables, both the likelihood function and posterior distribution are intractable without further assumptions about structural dependencies or model classes. While existing learning methods are fundamentally based on likelihood maximization, here we offer a new view of the parameter learning problem through the lens of optimal transport. This perspective licenses a framework that operates on many directed graphs without making unrealistic assumptions on the posterior over the latent variables or resorting to black-box variational approximations. We develop a theoretical framework and support it with extensive empirical evidence demonstrating the flexibility and versatility of our approach. Across experiments, we show that not only can our method recover the ground-truth parameters but it also performs competitively on downstream applications, notably the non-trivial task of discrete representation learning.

Abstract: Meta-Bayesian optimisation (meta-BO) aims to improve the sample efficiency of Bayesian optimisation by leveraging data from related tasks. While previous methods successfully meta-learn either a surrogate model or an acquisition function independently, joint training of both components remains an open challenge. This paper proposes the first end-to-end differentiable meta-BO framework that generalises neural processes to learn acquisition functions via transformer architectures. We enable this end-to-end framework with reinforcement learning (RL) to tackle the lack of labelled acquisition data. Early on, we notice that training transformer-based neural processes from scratch with RL is challenging due to insufficient supervision, especially when rewards are sparse. We formalise this claim with a combinatorial analysis showing that the widely used notion of regret as a reward signal exhibits a logarithmic sparsity pattern in trajectory lengths. To tackle this problem, we augment the RL objective with an auxiliary task that guides part of the architecture to learn a valid probabilistic model as an inductive bias. We demonstrate that our method achieves state-of-the-art regret results against various baselines in experiments on standard hyperparameter optimisation tasks and also outperforms others in the real-world problems of mixed-integer programming tuning, antibody design, and logic synthesis for electronic design automation.

Abstract: We present a novel perspective and algorithm for learning directed acyclic graphs (DAGs) from data generated by a linear structural equation model (SEM). First, we show that a linear SEM can be viewed as a linear transform that, in prior work, computes the data from a dense input vector of random valued root causes (as we will call them) associated with the nodes. Instead, we consider the case of (approximately) few root causes and also introduce noise in the measurement of the data. Intuitively, this means that the DAG data is produced by few data-generating events whose effect percolates through the DAG. We prove identifiability in this new setting and show that the true DAG is the global minimizer of the $L^0$-norm of the vector of root causes. For data with few root causes, with and without noise, we show superior performance compared to prior DAG learning methods.

Abstract: Some of the most successful knowledge graph embedding (KGE) models for link prediction -- CP, RESCAL, TuckER, ComplEx -- can be interpreted as energy-based models. Under this perspective they are not amenable for exact maximum-likelihood estimation (MLE), sampling and struggle to integrate logical constraints. This work re-interprets the score functions of these KGEs as circuits -- constrained computational graphs allowing efficient marginalisation. Then, we design two recipes to obtain efficient generative circuit models by either restricting their activations to be non-negative or squaring their outputs. Our interpretation comes with little or no loss of performance for link prediction, while the circuits framework unlocks exact learning by MLE, efficient sampling of new triples, and guarantee that logical constraints are satisfied by design. Furthermore, our models scale more gracefully than the original KGEs on graphs with millions of entities.

Abstract: Online learning holds the promise of enabling efficient long-term credit assignment in recurrent neural networks. However, current algorithms fall short of offline backpropagation by either not being scalable or failing to learn long-range dependencies. Here we present a high-performance online learning algorithm that merely doubles the memory and computational requirements of a single inference pass. We achieve this by leveraging independent recurrent modules in multi-layer networks, an architectural motif that has recently been shown to be particularly powerful. Experiments on synthetic memory problems and on the challenging long-range arena benchmark suite reveal that our algorithm performs competitively, establishing a new standard for what can be achieved through online learning. This ability to learn long-range dependencies offers a new perspective on learning in the brain and opens a promising avenue in neuromorphic computing.

Abstract: Despite the increasing relevance of explainable AI, assessing the quality of explanations remains a challenging issue. Due to the high costs associated with human-subject experiments, various proxy metrics are often used to approximately quantify explanation quality. Generally, one possible interpretation of the quality of an explanation is its inherent value for teaching a related concept to a student. In this work, we extend artificial simulatability studies to the domain of graph neural networks. Instead of costly human trials, we use explanation-supervisable graph neural networks to perform simulatability studies to quantify the inherent usefulness of attributional graph explanations. We perform an extensive ablation study to investigate the conditions under which the proposed analyses are most meaningful. We additionally validate our methods applicability on real-world graph classification and regression datasets. We find that relevant explanations can significantly boost the sample efficiency of graph neural networks and analyze the robustness towards noise and bias in the explanations. We believe that the notion of usefulness obtained from our proposed simulatability analysis provides a dimension of explanation quality that is largely orthogonal to the common practice of faithfulness and has great potential to expand the toolbox of explanation quality assessments, specifically for graph explanations.

Abstract: Existing heterogeneous treatment effects learners, also known as conditional average treatment effects (CATE) learners, lack a general mechanism for end-to-end inter-treatment information sharing, and data have to be split among potential outcome functions to train CATE learners which can lead to biased estimates with limited observational datasets. To address this issue, we propose a novel deep learning-based framework to train CATE learners that facilitates dynamic end-to-end information sharing among treatment groups. The framework is based on \textit{soft weight sharing} of \textit{hypernetworks}, which offers advantages such as parameter efficiency, faster training, and improved results. The proposed framework complements existing CATE learners and introduces a new class of uncertainty-aware CATE learners that we refer to as \textit{HyperCATE}. We develop HyperCATE versions of commonly used CATE learners and evaluate them on IHDP, ACIC-2016, and Twins benchmarks. Our experimental results show that the proposed framework improves the CATE estimation error via counterfactual inference, with increasing effectiveness for smaller datasets.

Abstract: We present an approach for analyzing message passing graph neural networks (MPNNs) based on an extension of graphon analysis to a so called graphon-signal analysis. A MPNN is a function that takes a graph and a signal on the graph (a graph-signal) and returns some value. Since the input space of MPNNs is non-Euclidean, i.e., graphs can be of any size and topology, properties such as generalization are less well understood for MPNNs than for Euclidean neural networks. We claim that one important missing ingredient in past work is a meaningful notion of graph-signal similarity measure, that endows the space of inputs to MPNNs with a regular structure. We present such a similarity measure, called the graphon-signal cut distance, which makes the space of all graph-signals a dense subset of a compact metric space -- the graphon-signal space. Informally, two deterministic graph-signals are close in cut distance if they ``look like'' they were sampled from the same random graph-signal model. Hence, our cut distance is a natural notion of graph-signal similarity, which allows comparing any pair of graph-signals of any size and topology. We prove that MPNNs are Lipschitz continuous functions over the graphon-signal metric space. We then give two applications of this result: 1) a generalization bound for MPNNs, and, 2) the stability of MPNNs to subsampling of graph-signals. Our results apply to any regular enough MPNN on any distribution of graph-signals, making the analysis rather universal.

Abstract: We propose SING (StabIlized and Normalized Gradient), a plug-and-play technique that improves the stability and generalization of the Adam(W) optimizer. SING is straightforward to implement and has minimal computational overhead, requiring only a layer-wise standardization of the gradients fed to Adam(W) without introducing additional hyper-parameters. We support the effectiveness and practicality of the proposed approach by showing improved results on a wide range of architectures, problems (such as image classification, depth estimation, and natural language processing), and in combination with other optimizers. We provide a theoretical analysis of the convergence of the method, and we show that by virtue of the standardization, SING can escape local minima narrower than a threshold that is inversely proportional to the network's depth.

Abstract: Multivariate long-term time series forecasting is of great application across many domains, such as energy consumption and weather forecasting. With the development of transformer-based methods, the performance of multivariate long-term time series forecasting has been significantly improved, however, the study of spatial features extracting in transformer-based model is rare and the consistency of different prediction periods is unsatisfactory due to the large span. In this work, we propose a complete solution to address these problems in terms of feature extraction and target prediction. For extraction, we design an efficient spatio-temporal encoding extractor including a semi-adaptive graph to acquire sufficient spatio-temporal information. For prediction, we propose a Cascaded Decoding Predictor (CDP) to strengthen the correlation between different intervals, which can also be utilized as a generic component to improve the performance of transformer-based methods. The proposed method, termed as Spatio-temporal Encoding Cascaded Transformer (Stecformer), achieving a notable gap over the baseline model and is comparable with the state-of-the-art performance of transformer-based methods on five benchmark datasets. We hope our attempt will serve as a regular configuration in multivariate long-term time series forecasting in the future.

Abstract: Adversarial detection aims to determine whether a given sample is an adversarial one based on the discrepancy between natural and adversarial distributions. Unfortunately, estimating or comparing two data distributions is extremely difficult, especially in high-dimension spaces. Recently, the gradient of log probability density (a.k.a., score) w.r.t. the sample is used as an alternative statistic to compute. However, we find that the score is sensitive in identifying adversarial samples due to insufficient information with one sample only. In this paper, we propose a new statistic called expected perturbation score (EPS), which is essentially the expected score of a sample after various perturbations. Specifically, to obtain adequate information regarding one sample, we perturb it by adding various noises to capture its multi-view observations. We theoretically prove that EPS is a proper statistic to compute the discrepancy between two samples under mild conditions. In practice, we can use a pre-trained diffusion model to estimate EPS for each sample. Last, we propose an EPS-based adversarial detection (EPS-AD) method, in which we develop EPS-based maximum mean discrepancy (MMD) as a metric to measure the discrepancy between the test sample and natural samples. We also prove that the EPS-based MMD between natural and adversarial samples is larger than that among natural samples. Extensive experiments show the superior adversarial detection performance of our EPS-AD.

Abstract: The $L_{2}$-regularized loss of Deep Linear Networks (DLNs) with more than one hidden layers has multiple local minima, corresponding to matrices with different ranks. In tasks such as matrix completion, the goal is to converge to the local minimum with the smallest rank that still fits the training data. While rank-underestimating minima can easily be avoided since they do not fit the data, gradient descent might get stuck at rank-overestimating minima. We show that with SGD, there is always a probability to jump from a higher rank minimum to a lower rank one, but the probability of jumping back is zero. More precisely, we define a sequence of sets $B_{1}\subset B_{2}\subset\cdots\subset B_{R}$ so that $B_{r}$ contains all minima of rank $r$ or less (and not more) that are absorbing for small enough ridge parameters $\lambda$ and learning rates $\eta$: SGD has prob. 0 of leaving $B_{r}$, and from any starting point there is a non-zero prob. for SGD to go in $B_{r}$.

Abstract: Collaborative learning techniques have significantly advanced in recent years, enabling private model training across multiple organizations. Despite this opportunity, firms face a dilemma when considering data sharing with competitors -- while collaboration can improve a company's machine learning model, it may also benefit competitors and hence reduce profits. In this work, we introduce a general framework for analyzing this data-sharing trade-off. The framework consists of three components, representing the firms' production decisions, the effect of additional data on model quality, and the data-sharing negotiation process, respectively. We then study an instantiation of the framework, based on a conventional market model from economic theory, to identify key factors that affect collaboration incentives. Our findings indicate a profound impact of market conditions on the data-sharing incentives. In particular, we find that reduced competition, in terms of the similarities between the firms' products, and harder learning tasks foster collaboration.

Abstract: Clustering is a common task in machine learning, but clusters of unlabelled data can be hard to quantify. The application of clustering algorithms in chemistry is often dependant on material representation. Ascertaining the effects of different representations, clustering algorithms, or data transformations on the resulting clusters is difficult due to the dimensionality of these data. We present a thorough analysis of measures for isotropy of a cluster, including a novel implantation based on an existing derivation. Using fractional anisotropy, a common method used in medical imaging for comparison, we then expand these measures to examine the average isotropy of a set of clusters. A use case for such measures is demonstrated by quantifying the effects of kernel approximation functions on different representations of the Inorganic Crystal Structure Database. Broader applicability of these methods is demonstrated in analysing learnt embedding of the MNIST dataset. Random clusters are explored to examine the differences between isotropy measures presented, and to see how each method scales with the dimensionality. Python implementations of these measures are provided for use by the community.

Abstract: Most reinforcement learning algorithms treat the context under which they operate as a stationary, isolated and undisturbed environment. However, in the real world, the environment is constantly changing due to a variety of external influences. To address this problem, we study Markov Decision Processes (MDP) under the influence of an external temporal process. We formalize this notion and discuss conditions under which the problem becomes tractable with suitable solutions. We propose a policy iteration algorithm to solve this problem and theoretically analyze its performance.

Abstract: While the world has been combating COVID-19 for over three years, an ongoing "Infodemic" due to the spread of fake news regarding the pandemic has also been a global issue. The existence of the fake news impact different aspect of our daily lives, including politics, public health, economic activities, etc. Readers could mistake fake news for real news, and consequently have less access to authentic information. This phenomenon will likely cause confusion of citizens and conflicts in society. Currently, there are major challenges in fake news research. It is challenging to accurately identify fake news data in social media posts. In-time human identification is infeasible as the amount of the fake news data is overwhelming. Besides, topics discussed in fake news are hard to identify due to their similarity to real news. The goal of this paper is to identify fake news on social media to help stop the spread. We present Deep Learning approaches and an ensemble approach for fake news detection. Our detection models achieved higher accuracy than previous studies. The ensemble approach further improved the detection performance. We discovered feature differences between fake news and real news items. When we added them into the sentence embeddings, we found that they affected the model performance. We applied a hybrid method and built models for recognizing topics from posts. We found half of the identified topics were overlapping in fake news and real news, which could increase confusion in the population.

Abstract: Circuit representation learning aims to obtain neural representations of circuit elements and has emerged as a promising research direction that can be applied to various EDA and logic reasoning tasks. Existing solutions, such as DeepGate, have the potential to embed both circuit structural information and functional behavior. However, their capabilities are limited due to weak supervision or flawed model design, resulting in unsatisfactory performance in downstream tasks. In this paper, we introduce DeepGate2, a novel functionality-aware learning framework that significantly improves upon the original DeepGate solution in terms of both learning effectiveness and efficiency. Our approach involves using pairwise truth table differences between sampled logic gates as training supervision, along with a well-designed and scalable loss function that explicitly considers circuit functionality. Additionally, we consider inherent circuit characteristics and design an efficient one-round graph neural network (GNN), resulting in an order of magnitude faster learning speed than the original DeepGate solution. Experimental results demonstrate significant improvements in two practical downstream tasks: logic synthesis and Boolean satisfiability solving. The code is available at https://github.com/cure-lab/DeepGate2

Abstract: We consider a combinatorial multi-armed bandit problem for maximum value reward function under maximum value and index feedback. This is a new feedback structure that lies in between commonly studied semi-bandit and full-bandit feedback structures. We propose an algorithm and provide a regret bound for problem instances with stochastic arm outcomes according to arbitrary distributions with finite supports. The regret analysis rests on considering an extended set of arms, associated with values and probabilities of arm outcomes, and applying a smoothness condition. Our algorithm achieves a $O((k/\Delta)\log(T))$ distribution-dependent and a $\tilde{O}(\sqrt{T})$ distribution-independent regret where $k$ is the number of arms selected in each round, $\Delta$ is a distribution-dependent reward gap and $T$ is the horizon time. Perhaps surprisingly, the regret bound is comparable to previously-known bound under more informative semi-bandit feedback. We demonstrate the effectiveness of our algorithm through experimental results.

Abstract: This paper investigates the relationship between the universal approximation property of deep neural networks and topological characteristics of datasets. Our primary contribution is to introduce data topology-dependent upper bounds on the network width. Specifically, we first show that a three-layer neural network, applying a ReLU activation function and max pooling, can be designed to approximate an indicator function over a compact set, one that is encompassed by a tight convex polytope. This is then extended to a simplicial complex, deriving width upper bounds based on its topological structure. Further, we calculate upper bounds in relation to the Betti numbers of select topological spaces. Finally, we prove the universal approximation property of three-layer ReLU networks using our topological approach. We also verify that gradient descent converges to the network structure proposed in our study.

Abstract: Machine learning systems such as large scale recommendation systems or natural language processing systems are usually trained on billions of training points and are associated with hundreds of billions or trillions of parameters. Improving the learning process in such a way that both the training load is reduced and the model accuracy improved is highly desired. In this paper we take a first step toward solving this problem, studying influence functions from the perspective of simplifying the computations they involve. We discuss assumptions, under which influence computations can be performed on significantly fewer parameters. We also demonstrate that the sign of the influence value can indicate whether a training point is to memorize, as opposed to generalize upon. For this purpose we formally define what memorization means for a training point, as opposed to generalization. We conclude that influence functions can be made practical, even for large scale machine learning systems, and that influence values can be taken into account by algorithms that selectively remove training points, as part of the learning process.

Abstract: In this paper, we propose a novel centralized Asynchronous Federated Learning (FL) framework, FAVAS, for training Deep Neural Networks (DNNs) in resource-constrained environments. Despite its popularity, ``classical'' federated learning faces the increasingly difficult task of scaling synchronous communication over large wireless networks. Moreover, clients typically have different computing resources and therefore computing speed, which can lead to a significant bias (in favor of ``fast'' clients) when the updates are asynchronous. Therefore, practical deployment of FL requires to handle users with strongly varying computing speed in communication/resource constrained setting. We provide convergence guarantees for FAVAS in a smooth, non-convex environment and carefully compare the obtained convergence guarantees with existing bounds, when they are available. Experimental results show that the FAVAS algorithm outperforms current methods on standard benchmarks.

Abstract: Oversmoothing in Graph Neural Networks (GNNs) refers to the phenomenon where increasing network depth leads to homogeneous node representations. While previous work has established that Graph Convolutional Networks (GCNs) exponentially lose expressive power, it remains controversial whether the graph attention mechanism can mitigate oversmoothing. In this work, we provide a definitive answer to this question through a rigorous mathematical analysis, by viewing attention-based GNNs as nonlinear time-varying dynamical systems and incorporating tools and techniques from the theory of products of inhomogeneous matrices and the joint spectral radius. We establish that, contrary to popular belief, the graph attention mechanism cannot prevent oversmoothing and loses expressive power exponentially. The proposed framework extends the existing results on oversmoothing for symmetric GCNs to a significantly broader class of GNN models. In particular, our analysis accounts for asymmetric, state-dependent and time-varying aggregation operators and a wide range of common nonlinear activation functions, such as ReLU, LeakyReLU, GELU and SiLU.

Abstract: Due to the unsupervised nature of anomaly detection, the key to fueling deep models is finding supervisory signals. Different from current reconstruction-guided generative models and transformation-based contrastive models, we devise novel data-driven supervision for tabular data by introducing a characteristic -- scale -- as data labels. By representing varied sub-vectors of data instances, we define scale as the relationship between the dimensionality of original sub-vectors and that of representations. Scales serve as labels attached to transformed representations, thus offering ample labeled data for neural network training. This paper further proposes a scale learning-based anomaly detection method. Supervised by the learning objective of scale distribution alignment, our approach learns the ranking of representations converted from varied subspaces of each data instance. Through this proxy task, our approach models inherent regularities and patterns within data, which well describes data "normality". Abnormal degrees of testing instances are obtained by measuring whether they fit these learned patterns. Extensive experiments show that our approach leads to significant improvement over state-of-the-art generative/contrastive anomaly detection methods.

Abstract: Incremental learning (IL) suffers from catastrophic forgetting of old tasks when learning new tasks. This can be addressed by replaying previous tasks' data stored in a memory, which however is usually prone to size limits and privacy leakage. Recent studies store only class centroids as prototypes and augment them with Gaussian noises to create synthetic data for replay. However, they cannot effectively avoid class interference near their margins that leads to forgetting. Moreover, the injected noises distort the rich structure between real data and prototypes, hence even detrimental to IL. In this paper, we propose YONO that You Only Need to replay One condensed prototype per class, which for the first time can even outperform memory-costly exemplar-replay methods. To this end, we develop a novel prototype learning method that (1) searches for more representative prototypes in high-density regions by an attentional mean-shift algorithm and (2) moves samples in each class to their prototype to form a compact cluster distant from other classes. Thereby, the class margins are maximized, which effectively reduces interference causing future forgetting. In addition, we extend YONO to YONO+, which creates synthetic replay data by random sampling in the neighborhood of each prototype in the representation space. We show that the synthetic data can further improve YONO. Extensive experiments on IL benchmarks demonstrate the advantages of YONO/YONO+ over existing IL methods in terms of both accuracy and forgetting.

Abstract: We propose Convex Constraint Learning for Reinforcement Learning (CoCoRL), a novel approach for inferring shared constraints in a Constrained Markov Decision Process (CMDP) from a set of safe demonstrations with possibly different reward functions. While previous work is limited to demonstrations with known rewards or fully known environment dynamics, CoCoRL can learn constraints from demonstrations with different unknown rewards without knowledge of the environment dynamics. CoCoRL constructs a convex safe set based on demonstrations, which provably guarantees safety even for potentially sub-optimal (but safe) demonstrations. For near-optimal demonstrations, CoCoRL converges to the true safe set with no policy regret. We evaluate CoCoRL in tabular environments and a continuous driving simulation with multiple constraints. CoCoRL learns constraints that lead to safe driving behavior and that can be transferred to different tasks and environments. In contrast, alternative methods based on Inverse Reinforcement Learning (IRL) often exhibit poor performance and learn unsafe policies.

Abstract: Particle-based deep generative models, such as gradient flows and score-based diffusion models, have recently gained traction thanks to their striking performance. Their principle of displacing particle distributions by differential equations is conventionally seen as opposed to the previously widespread generative adversarial networks (GANs), which involve training a pushforward generator network. In this paper, we challenge this interpretation and propose a novel framework that unifies particle and adversarial generative models by framing generator training as a generalization of particle models. This suggests that a generator is an optional addition to any such generative model. Consequently, integrating a generator into a score-based diffusion model and training a GAN without a generator naturally emerge from our framework. We empirically test the viability of these original models as proofs of concepts of potential applications of our framework.

Abstract: We formalize and study a phenomenon called feature collapse that makes precise the intuitive idea that entities playing a similar role in a learning task receive similar representations. As feature collapse requires a notion of task, we leverage a simple but prototypical NLP task to study it. We start by showing experimentally that feature collapse goes hand in hand with generalization. We then prove that, in the large sample limit, distinct words that play identical roles in this NLP task receive identical local feature representations in a neural network. This analysis reveals the crucial role that normalization mechanisms, such as LayerNorm, play in feature collapse and in generalization.

Abstract: Since the control of the Lipschitz constant has a great impact on the training stability, generalization, and robustness of neural networks, the estimation of this value is nowadays a real scientific challenge. In this paper we introduce a precise, fast, and differentiable upper bound for the spectral norm of convolutional layers using circulant matrix theory and a new alternative to the Power iteration. Called the Gram iteration, our approach exhibits a superlinear convergence. First, we show through a comprehensive set of experiments that our approach outperforms other state-of-the-art methods in terms of precision, computational cost, and scalability. Then, it proves highly effective for the Lipschitz regularization of convolutional neural networks, with competitive results against concurrent approaches. Code is available at https://github.com/blaisedelattre/lip4conv.

Abstract: Latent Graph Inference (LGI) relaxed the reliance of Graph Neural Networks (GNNs) on a given graph topology by dynamically learning it. However, most of LGI methods assume to have a (noisy, incomplete, improvable, ...) input graph to rewire and can solely learn regular graph topologies. In the wake of the success of Topological Deep Learning (TDL), we study Latent Topology Inference (LTI) for learning higher-order cell complexes (with sparse and not regular topology) describing multi-way interactions between data points. To this aim, we introduce the Differentiable Cell Complex Module (DCM), a novel learnable function that computes cell probabilities in the complex to improve the downstream task. We show how to integrate DCM with cell complex message passing networks layers and train it in a end-to-end fashion, thanks to a two-step inference procedure that avoids an exhaustive search across all possible cells in the input, thus maintaining scalability. Our model is tested on several homophilic and heterophilic graph datasets and it is shown to outperform other state-of-the-art techniques, offering significant improvements especially in cases where an input graph is not provided.

Abstract: In this paper, we find and analyze that we can easily drop the double descent by only adding one dropout layer before the fully-connected linear layer. The surprising double-descent phenomenon has drawn public attention in recent years, making the prediction error rise and drop as we increase either sample or model size. The current paper shows that it is possible to alleviate these phenomena by using optimal dropout in the linear regression model and the nonlinear random feature regression, both theoretically and empirically. % ${y}=X{\beta}^0+{\epsilon}$ with $X\in\mathbb{R}^{n\times p}$. We obtain the optimal dropout hyperparameter by estimating the ground truth ${\beta}^0$ with generalized ridge typed estimator $\hat{{\beta}}=(X^TX+\alpha\cdot\mathrm{diag}(X^TX))^{-1}X^T{y}$. Moreover, we empirically show that optimal dropout can achieve a monotonic test error curve in nonlinear neural networks using Fashion-MNIST and CIFAR-10. Our results suggest considering dropout for risk curve scaling when meeting the peak phenomenon. In addition, we figure out why previous deep learning models do not encounter double-descent scenarios -- because we already apply a usual regularization approach like the dropout in our models. To our best knowledge, this paper is the first to analyze the relationship between dropout and double descent.

Abstract: What can be learned about causality and experimentation from passive data? This question is salient given recent successes of passively-trained language models in interactive domains such as tool use. Passive learning is inherently limited. However, we show that purely passive learning can in fact allow an agent to learn generalizable strategies for determining and using causal structures, as long as the agent can intervene at test time. We formally illustrate that learning a strategy of first experimenting, then seeking goals, can allow generalization from passive learning in principle. We then show empirically that agents trained via imitation on expert data can indeed generalize at test time to infer and use causal links which are never present in the training data; these agents can also generalize experimentation strategies to novel variable sets never observed in training. We then show that strategies for causal intervention and exploitation can be generalized from passive data even in a more complex environment with high-dimensional observations, with the support of natural language explanations. Explanations can even allow passive learners to generalize out-of-distribution from perfectly-confounded training data. Finally, we show that language models, trained only on passive next-word prediction, can generalize causal intervention strategies from a few-shot prompt containing examples of experimentation, together with explanations and reasoning. These results highlight the surprising power of passive learning of active causal strategies, and may help to understand the behaviors and capabilities of language models.

Abstract: Data augmentation (DA) is a crucial technique for enhancing the sample efficiency of visual reinforcement learning (RL) algorithms. Notably, employing simple observation transformations alone can yield outstanding performance without extra auxiliary representation tasks or pre-trained encoders. However, it remains unclear which attributes of DA account for its effectiveness in achieving sample-efficient visual RL. To investigate this issue and further explore the potential of DA, this work conducts comprehensive experiments to assess the impact of DA's attributes on its efficacy and provides the following insights and improvements: (1) For individual DA operations, we reveal that both ample spatial diversity and slight hardness are indispensable. Building on this finding, we introduce Random PadResize (Rand PR), a new DA operation that offers abundant spatial diversity with minimal hardness. (2) For multi-type DA fusion schemes, the increased DA hardness and unstable data distribution result in the current fusion schemes being unable to achieve higher sample efficiency than their corresponding individual operations. Taking the non-stationary nature of RL into account, we propose a RL-tailored multi-type DA fusion scheme called Cycling Augmentation (CycAug), which performs periodic cycles of different DA operations to increase type diversity while maintaining data distribution consistency. Extensive evaluations on the DeepMind Control suite and CARLA driving simulator demonstrate that our methods achieve superior sample efficiency compared with the prior state-of-the-art methods.

Abstract: Unsupervised source separation involves unraveling an unknown set of source signals recorded through a mixing operator, with limited prior knowledge about the sources, and only access to a dataset of signal mixtures. This problem is inherently ill-posed and is further challenged by the variety of time-scales exhibited by sources in time series data. Existing methods typically rely on a preselected window size that limits their capacity to handle multi-scale sources. To address this issue, instead of operating in the time domain, we propose an unsupervised multi-scale clustering and source separation framework by leveraging wavelet scattering covariances that provide a low-dimensional representation of stochastic processes, capable of distinguishing between different non-Gaussian stochastic processes. Nested within this representation space, we develop a factorial Gaussian-mixture variational autoencoder that is trained to (1) probabilistically cluster sources at different time-scales and (2) independently sample scattering covariance representations associated with each cluster. Using samples from each cluster as prior information, we formulate source separation as an optimization problem in the wavelet scattering covariance representation space, resulting in separated sources in the time domain. When applied to seismic data recorded during the NASA InSight mission on Mars, our multi-scale nested approach proves to be a powerful tool for discriminating between sources varying greatly in time-scale, e.g., minute-long transient one-sided pulses (known as ``glitches'') and structured ambient noises resulting from atmospheric activities that typically last for tens of minutes. These results provide an opportunity to conduct further investigations into the isolated sources related to atmospheric-surface interactions, thermal relaxations, and other complex phenomena.

Abstract: Explainability techniques are crucial in gaining insights into the reasons behind the predictions of deep learning models, which have not yet been applied to chemical language models. We propose an explainable AI technique that attributes the importance of individual atoms towards the predictions made by these models. Our method backpropagates the relevance information towards the chemical input string and visualizes the importance of individual atoms. We focus on self-attention Transformers operating on molecular string representations and leverage a pretrained encoder for finetuning. We showcase the method by predicting and visualizing solubility in water and organic solvents. We achieve competitive model performance while obtaining interpretable predictions, which we use to inspect the pretrained model.

Abstract: For automotive applications, the Graph Attention Network (GAT) is a prominently used architecture to include relational information of a traffic scenario during feature embedding. As shown in this work, however, one of the most popular GAT realizations, namely GATv2, has potential pitfalls that hinder an optimal parameter learning. Especially for small and sparse graph structures a proper optimization is problematic. To surpass limitations, this work proposes architectural modifications of GATv2. In controlled experiments, it is shown that the proposed model adaptions improve prediction performance in a node-level regression task and make it more robust to parameter initialization. This work aims for a better understanding of the attention mechanism and analyzes its interpretability of identifying causal importance.

Abstract: State-of-the-art approaches for training Differentially Private (DP) Deep Neural Networks (DNN) faces difficulties to estimate tight bounds on the sensitivity of the network's layers, and instead rely on a process of per-sample gradient clipping. This clipping process not only biases the direction of gradients but also proves costly both in memory consumption and in computation. To provide sensitivity bounds and bypass the drawbacks of the clipping process, our theoretical analysis of Lipschitz constrained networks reveals an unexplored link between the Lipschitz constant with respect to their input and the one with respect to their parameters. By bounding the Lipschitz constant of each layer with respect to its parameters we guarantee DP training of these networks. This analysis not only allows the computation of the aforementioned sensitivities at scale but also provides leads on to how maximize the gradient-to-noise ratio for fixed privacy guarantees. To facilitate the application of Lipschitz networks and foster robust and certifiable learning under privacy guarantees, we provide a Python package that implements building blocks allowing the construction and private training of such networks.

Abstract: Many real-world decision-making tasks, such as safety-critical scenarios, cannot be fully described in a single-objective setting using the Markov Decision Process (MDP) framework, as they include hard constraints. These can instead be modeled with additional cost functions within the Constrained Markov Decision Process (CMDP) framework. Even though CMDPs have been extensively studied in the Reinforcement Learning literature, little attention has been given to sampling-based planning algorithms such as MCTS for solving them. Previous approaches use Monte Carlo cost estimates to avoid constraint violations. However, these suffer from high variance which results in conservative performance with respect to costs. We propose Constrained MCTS (C-MCTS), an algorithm that estimates cost using a safety critic. The safety critic training is based on Temporal Difference learning in an offline phase prior to agent deployment. This critic limits the exploration of the search tree and removes unsafe trajectories within MCTS during deployment. C-MCTS satisfies cost constraints but operates closer to the constraint boundary, achieving higher rewards compared to previous work. As a nice byproduct, the planner is more efficient requiring fewer planning steps. Most importantly, we show that under model mismatch between the planner and the real world, our approach is less susceptible to cost violations than previous work.

Abstract: Score distillation sampling (SDS) has shown great promise in text-to-3D generation by distilling pretrained large-scale text-to-image diffusion models, but suffers from over-saturation, over-smoothing, and low-diversity problems. In this work, we propose to model the 3D parameter as a random variable instead of a constant as in SDS and present variational score distillation (VSD), a principled particle-based variational framework to explain and address the aforementioned issues in text-to-3D generation. We show that SDS is a special case of VSD and leads to poor samples with both small and large CFG weights. In comparison, VSD works well with various CFG weights as ancestral sampling from diffusion models and simultaneously improves the diversity and sample quality with a common CFG weight (i.e., $7.5$). We further present various improvements in the design space for text-to-3D such as distillation time schedule and density initialization, which are orthogonal to the distillation algorithm yet not well explored. Our overall approach, dubbed ProlificDreamer, can generate high rendering resolution (i.e., $512\times512$) and high-fidelity NeRF with rich structure and complex effects (e.g., smoke and drops). Further, initialized from NeRF, meshes fine-tuned by VSD are meticulously detailed and photo-realistic. Project page: https://ml.cs.tsinghua.edu.cn/prolificdreamer/

Abstract: Many machine learning approaches for decision making, such as reinforcement learning, rely on simulators or predictive models to forecast the time-evolution of quantities of interest, e.g., the state of an agent or the reward of a policy. Forecasts of such complex phenomena are commonly described by highly nonlinear dynamical systems, making their use in optimization-based decision-making challenging. Koopman operator theory offers a beneficial paradigm for addressing this problem by characterizing forecasts via linear dynamical systems. This makes system analysis and long-term predictions simple -- involving only matrix multiplications. However, the transformation to a linear system is generally non-trivial and unknown, requiring learning-based approaches. While there exists a variety of approaches, they usually lack crucial learning-theoretic guarantees, such that the behavior of the obtained models with increasing data and dimensionality is often unclear. We address the aforementioned by deriving a novel reproducing kernel Hilbert space (RKHS) that solely spans transformations into linear dynamical systems. The resulting Koopman Kernel Regression (KKR) framework enables the use of statistical learning tools from function approximation for novel convergence results and generalization risk bounds under weaker assumptions than existing work. Our numerical experiments indicate advantages over state-of-the-art statistical learning approaches for Koopman-based predictors.

Abstract: This study focuses on the topic of offline preference-based reinforcement learning (PbRL), a variant of conventional reinforcement learning that dispenses with the need for online interaction or specification of reward functions. Instead, the agent is provided with pre-existing offline trajectories and human preferences between pairs of trajectories to extract the dynamics and task information, respectively. Since the dynamics and task information are orthogonal, a naive approach would involve using preference-based reward learning followed by an off-the-shelf offline RL algorithm. However, this requires the separate learning of a scalar reward function, which is assumed to be an information bottleneck. To address this issue, we propose the offline preference-guided policy optimization (OPPO) paradigm, which models offline trajectories and preferences in a one-step process, eliminating the need for separately learning a reward function. OPPO achieves this by introducing an offline hindsight information matching objective for optimizing a contextual policy and a preference modeling objective for finding the optimal context. OPPO further integrates a well-performing decision policy by optimizing the two objectives iteratively. Our empirical results demonstrate that OPPO effectively models offline preferences and outperforms prior competing baselines, including offline RL algorithms performed over either true or pseudo reward function specifications. Our code is available at https://github.com/bkkgbkjb/OPPO .

Abstract: The success of many machine learning (ML) methods depends crucially on having large amounts of labeled data. However, obtaining enough labeled data can be expensive, time-consuming, and subject to ethical constraints for many applications. One approach that has shown tremendous value in addressing this challenge is semi-supervised learning (SSL); this technique utilizes both labeled and unlabeled data during training, often with much less labeled data than unlabeled data, which is often relatively easy and inexpensive to obtain. In fact, SSL methods are particularly useful in applications where the cost of labeling data is especially expensive, such as medical analysis, natural language processing (NLP), or speech recognition. A subset of SSL methods that have achieved great success in various domains involves algorithms that integrate graph-based techniques. These procedures are popular due to the vast amount of information provided by the graphical framework and the versatility of their applications. In this work, we propose an algebraic topology-based semi-supervised method called persistent Laplacian-enhanced graph MBO (PL-MBO) by integrating persistent spectral graph theory with the classical Merriman-Bence- Osher (MBO) scheme. Specifically, we use a filtration procedure to generate a sequence of chain complexes and associated families of simplicial complexes, from which we construct a family of persistent Laplacians. Overall, it is a very efficient procedure that requires much less labeled data to perform well compared to many ML techniques, and it can be adapted for both small and large datasets. We evaluate the performance of the proposed method on data classification, and the results indicate that the proposed technique outperforms other existing semi-supervised algorithms.

Abstract: We provide a new non-asymptotic analysis of distributed temporal difference learning with linear function approximation. Our approach relies on ``one-shot averaging,'' where $N$ agents run identical local copies of the TD(0) method and average the outcomes only once at the very end. We demonstrate a version of the linear time speedup phenomenon, where the convergence time of the distributed process is a factor of $N$ faster than the convergence time of TD(0). This is the first result proving benefits from parallelism for temporal difference methods.

Abstract: This paper studies the online node classification problem under a transductive learning setting. Current methods either invert a graph kernel matrix with $\mathcal{O}(n^3)$ runtime and $\mathcal{O}(n^2)$ space complexity or sample a large volume of random spanning trees, thus are difficult to scale to large graphs. In this work, we propose an improvement based on the \textit{online relaxation} technique introduced by a series of works (Rakhlin et al.,2012; Rakhlin and Sridharan, 2015; 2017). We first prove an effective regret $\mathcal{O}(\sqrt{n^{1+\gamma}})$ when suitable parameterized graph kernels are chosen, then propose an approximate algorithm FastONL enjoying $\mathcal{O}(k\sqrt{n^{1+\gamma}})$ regret based on this relaxation. The key of FastONL is a \textit{generalized local push} method that effectively approximates inverse matrix columns and applies to a series of popular kernels. Furthermore, the per-prediction cost is $\mathcal{O}(\text{vol}({\mathcal{S}})\log 1/\epsilon)$ locally dependent on the graph with linear memory cost. Experiments show that our scalable method enjoys a better tradeoff between local and global consistency.

Abstract: Collaborative learning techniques have the potential to enable training machine learning models that are superior to models trained on a single entity's data. However, in many cases, potential participants in such collaborative schemes are competitors on a downstream task, such as firms that each aim to attract customers by providing the best recommendations. This can incentivize dishonest updates that damage other participants' models, potentially undermining the benefits of collaboration. In this work, we formulate a game that models such interactions and study two learning tasks within this framework: single-round mean estimation and multi-round SGD on strongly-convex objectives. For a natural class of player actions, we show that rational clients are incentivized to strongly manipulate their updates, preventing learning. We then propose mechanisms that incentivize honest communication and ensure learning quality comparable to full cooperation. Lastly, we empirically demonstrate the effectiveness of our incentive scheme on a standard non-convex federated learning benchmark. Our work shows that explicitly modeling the incentives and actions of dishonest clients, rather than assuming them malicious, can enable strong robustness guarantees for collaborative learning.

Abstract: Learning from human feedback has been shown to improve text-to-image models. These techniques first learn a reward function that captures what humans care about in the task and then improve the models based on the learned reward function. Even though relatively simple approaches (e.g., rejection sampling based on reward scores) have been investigated, fine-tuning text-to-image models with the reward function remains challenging. In this work, we propose using online reinforcement learning (RL) to fine-tune text-to-image models. We focus on diffusion models, defining the fine-tuning task as an RL problem, and updating the pre-trained text-to-image diffusion models using policy gradient to maximize the feedback-trained reward. Our approach, coined DPOK, integrates policy optimization with KL regularization. We conduct an analysis of KL regularization for both RL fine-tuning and supervised fine-tuning. In our experiments, we show that DPOK is generally superior to supervised fine-tuning with respect to both image-text alignment and image quality.

Abstract: This paper proposes a new easy-to-implement parameter-free gradient-based optimizer: DoWG (Distance over Weighted Gradients). We prove that DoWG is efficient -- matching the convergence rate of optimally tuned gradient descent in convex optimization up to a logarithmic factor without tuning any parameters, and universal -- automatically adapting to both smooth and nonsmooth problems. While popular algorithms such as AdaGrad, Adam, or DoG compute a running average of the squared gradients, DoWG maintains a new distance-based weighted version of the running average, which is crucial to achieve the desired properties. To our best knowledge, DoWG is the first parameter-free, efficient, and universal algorithm that does not require backtracking search procedures. It is also the first parameter-free AdaGrad style algorithm that adapts to smooth optimization. To complement our theory, we also show empirically that DoWG trains at the edge of stability, and validate its effectiveness on practical machine learning tasks. This paper further uncovers the underlying principle behind the success of the AdaGrad family of algorithms by presenting a novel analysis of Normalized Gradient Descent (NGD), that shows NGD adapts to smoothness when it exists, with no change to the stepsize. This establishes the universality of NGD and partially explains the empirical observation that it trains at the edge of stability in a much more general setup compared to standard gradient descent. The latter might be of independent interest to the community.

Abstract: Sharpness-aware minimization (SAM) is a recently proposed method that minimizes the sharpness of the training loss of a neural network. While its generalization improvement is well-known and is the primary motivation, we uncover an additional intriguing effect of SAM: reduction of the feature rank which happens at different layers of a neural network. We show that this low-rank effect occurs very broadly: for different architectures such as fully-connected networks, convolutional networks, vision transformers and for different objectives such as regression, classification, language-image contrastive training. To better understand this phenomenon, we provide a mechanistic understanding of how low-rank features arise in a simple two-layer network. We observe that a significant number of activations gets entirely pruned by SAM which directly contributes to the rank reduction. We confirm this effect theoretically and check that it can also occur in deep networks, although the overall rank reduction mechanism can be more complex, especially for deep networks with pre-activation skip connections and self-attention layers. We make our code available at https://github.com/tml-epfl/sam-low-rank-features.

Abstract: Stochastic Gradient Descent (SGD) is arguably the most important single algorithm in modern machine learning. Although SGD with unbiased gradient estimators has been studied extensively over at least half a century, SGD variants relying on biased estimators are rare. Nevertheless, there has been an increased interest in this topic in recent years. However, existing literature on SGD with biased estimators (BiasedSGD) lacks coherence since each new paper relies on a different set of assumptions, without any clear understanding of how they are connected, which may lead to confusion. We address this gap by establishing connections among the existing assumptions, and presenting a comprehensive map of the underlying relationships. Additionally, we introduce a new set of assumptions that is provably weaker than all previous assumptions, and use it to present a thorough analysis of BiasedSGD in both convex and non-convex settings, offering advantages over previous results. We also provide examples where biased estimators outperform their unbiased counterparts or where unbiased versions are simply not available. Finally, we demonstrate the effectiveness of our framework through experimental results that validate our theoretical findings.

Abstract: Communication compression is a common technique in distributed optimization that can alleviate communication overhead by transmitting compressed gradients and model parameters. However, compression can introduce information distortion, which slows down convergence and incurs more communication rounds to achieve desired solutions. Given the trade-off between lower per-round communication costs and additional rounds of communication, it is unclear whether communication compression reduces the total communication cost. This paper explores the conditions under which unbiased compression, a widely used form of compression, can reduce the total communication cost, as well as the extent to which it can do so. To this end, we present the first theoretical formulation for characterizing the total communication cost in distributed optimization with communication compression. We demonstrate that unbiased compression alone does not necessarily save the total communication cost, but this outcome can be achieved if the compressors used by all workers are further assumed independent. We establish lower bounds on the communication rounds required by algorithms using independent unbiased compressors to minimize smooth convex functions, and show that these lower bounds are tight by refining the analysis for ADIANA. Our results reveal that using independent unbiased compression can reduce the total communication cost by a factor of up to $\Theta(\sqrt{\min\{n, \kappa\}})$, where $n$ is the number of workers and $\kappa$ is the condition number of the functions being minimized. These theoretical findings are supported by experimental results.

Abstract: It is well-known that real-world changes constituting distribution shift adversely affect model performance. How to characterize those changes in an interpretable manner is poorly understood. Existing techniques to address this problem take the form of shift explanations that elucidate how to map samples from the original distribution toward the shifted one by reducing the disparity between these two distributions. However, these methods can introduce group irregularities, leading to explanations that are less feasible and robust. To address these issues, we propose Group-aware Shift Explanations (GSE), a method that produces interpretable explanations by leveraging worst-group optimization to rectify group irregularities. We demonstrate how GSE not only maintains group structures, such as demographic and hierarchical subpopulations, but also enhances feasibility and robustness in the resulting explanations in a wide range of tabular, language, and image settings.

Abstract: Diffusion models are powerful generative models but suffer from slow sampling, often taking 1000 sequential denoising steps for one sample. As a result, considerable efforts have been directed toward reducing the number of denoising steps, but these methods hurt sample quality. Instead of reducing the number of denoising steps (trading quality for speed), in this paper we explore an orthogonal approach: can we run the denoising steps in parallel (trading compute for speed)? In spite of the sequential nature of the denoising steps, we show that surprisingly it is possible to parallelize sampling via Picard iterations, by guessing the solution of future denoising steps and iteratively refining until convergence. With this insight, we present ParaDiGMS, a novel method to accelerate the sampling of pretrained diffusion models by denoising multiple steps in parallel. ParaDiGMS is the first diffusion sampling method that enables trading compute for speed and is even compatible with existing fast sampling techniques such as DDIM and DPMSolver. Using ParaDiGMS, we improve sampling speed by 2-4x across a range of robotics and image generation models, giving state-of-the-art sampling speeds of 0.2s on 100-step DiffusionPolicy and 16s on 1000-step StableDiffusion-v2 with no measurable degradation of task reward, FID score, or CLIP score.

Abstract: We derive a sound positive semi-definite approximation of the Hessian of deep models for which Hessian-vector products are easily computable. This enables us to provide an adaptive SGD learning rate strategy based on the minimization of the local quadratic approximation, which requires just twice the computation of a single SGD run, but performs comparably with grid search on SGD learning rates on different model architectures (CNN with and without residual connections) on classification tasks. We also compare the novel approximation with the Gauss-Newton approximation.

Abstract: We discuss a federated learned compression problem, where the goal is to learn a compressor from real-world data which is scattered across clients and may be statistically heterogeneous, yet share a common underlying representation. We propose a distributed source model that encompasses both characteristics, and naturally suggests a compressor architecture that uses analysis and synthesis transforms shared by clients. Inspired by personalized federated learning methods, we employ an entropy model that is personalized to each client. This allows for a global latent space to be learned across clients, and personalized entropy models that adapt to the clients' latent distributions. We show empirically that this strategy outperforms solely local methods, which indicates that learned compression also benefits from a shared global representation in statistically heterogeneous federated settings.

Abstract: When machine learning models are trained continually on a sequence of tasks, they are liable to forget what they learned on previous tasks -- a phenomenon known as catastrophic forgetting. Proposed solutions to catastrophic forgetting tend to involve storing information about past tasks, meaning that memory usage is a chief consideration in determining their practicality. This paper proposes a memory-efficient solution to catastrophic forgetting, improving upon an established algorithm known as orthogonal gradient descent (OGD). OGD utilizes prior model gradients to find weight updates that preserve performance on prior datapoints. However, since the memory cost of storing prior model gradients grows with the runtime of the algorithm, OGD is ill-suited to continual learning over arbitrarily long time horizons. To address this problem, this paper proposes SketchOGD. SketchOGD employs an online sketching algorithm to compress model gradients as they are encountered into a matrix of a fixed, user-determined size. In contrast to existing memory-efficient variants of OGD, SketchOGD runs online without the need for advance knowledge of the total number of tasks, is simple to implement, and is more amenable to analysis. We provide theoretical guarantees on the approximation error of the relevant sketches under a novel metric suited to the downstream task of OGD. Experimentally, we find that SketchOGD tends to outperform current state-of-the-art variants of OGD given a fixed memory budget.

Abstract: This work bridges two important concepts: the Neural Tangent Kernel (NTK), which captures the evolution of deep neural networks (DNNs) during training, and the Neural Collapse (NC) phenomenon, which refers to the emergence of symmetry and structure in the last-layer features of well-trained classification DNNs. We adopt the natural assumption that the empirical NTK develops a block structure aligned with the class labels, i.e., samples within the same class have stronger correlations than samples from different classes. Under this assumption, we derive the dynamics of DNNs trained with mean squared (MSE) loss and break them into interpretable phases. Moreover, we identify an invariant that captures the essence of the dynamics, and use it to prove the emergence of NC in DNNs with block-structured NTK. We provide large-scale numerical experiments on three common DNN architectures and three benchmark datasets to support our theory.

Abstract: In this paper, we consider the problem of learning a linear regression model on a data domain of interest (target) given few samples. To aid learning, we are provided with a set of pre-trained regression models that are trained on potentially different data domains (sources). Assuming a representation structure for the data generating linear models at the sources and the target domains, we propose a representation transfer based learning method for constructing the target model. The proposed scheme is comprised of two phases: (i) utilizing the different source representations to construct a representation that is adapted to the target data, and (ii) using the obtained model as an initialization to a fine-tuning procedure that re-trains the entire (over-parameterized) regression model on the target data. For each phase of the training method, we provide excess risk bounds for the learned model compared to the true data generating target model. The derived bounds show a gain in sample complexity for our proposed method compared to the baseline method of not leveraging source representations when achieving the same excess risk, therefore, theoretically demonstrating the effectiveness of transfer learning for linear regression.

Abstract: Existing neural architecture search (NAS) methods typically rely on pre-specified super deep neural networks (super-networks) with handcrafted search spaces beforehand. Such requirements make it challenging to extend them onto general scenarios without significant human expertise and manual intervention. To overcome the limitations, we propose the third generation of Only-Train-Once (OTOv3). OTOv3 is perhaps the first automated system that trains general super-networks and produces high-performing sub-networks in the one shot manner without pretraining and fine-tuning. Technologically, OTOv3 delivers three noticeable contributions to minimize human efforts: (i) automatic search space construction for general super-networks; (ii) a Hierarchical Half-Space Projected Gradient (H2SPG) that leverages the dependency graph to ensure the network validity during optimization and reliably produces a solution with both high performance and hierarchical group sparsity; and (iii) automatic sub-network construction based on the super-network and the H2SPG solution. Numerically, we demonstrate the effectiveness of OTOv3 on a variety of super-networks, including RegNet, StackedUnets, SuperResNet, and DARTS, over benchmark datasets such as CIFAR10, Fashion-MNIST, ImageNet, STL-10, and SVNH. The sub-networks computed by OTOv3 achieve competitive even superior performance compared to the super-networks and other state-of-the-arts. The library will be released at https://github.com/tianyic/only_train_once.

Abstract: Statistical divergences quantify the difference between probability distributions finding multiple uses in machine-learning. However, a fundamental challenge is to estimate divergence from empirical samples since the underlying distributions of the data are usually unknown. In this work, we propose the representation Jensen-Shannon Divergence, a novel divergence based on covariance operators in reproducing kernel Hilbert spaces (RKHS). Our approach embeds the data distributions in an RKHS and exploits the spectrum of the covariance operators of the representations. We provide an estimator from empirical covariance matrices by explicitly mapping the data to an RKHS using Fourier features. This estimator is flexible, scalable, differentiable, and suitable for minibatch-based optimization problems. Additionally, we provide an estimator based on kernel matrices without having an explicit mapping to the RKHS. We show that this quantity is a lower bound on the Jensen-Shannon divergence, and we propose a variational approach to estimate it. We applied our divergence to two-sample testing outperforming related state-of-the-art techniques in several datasets. We used the representation Jensen-Shannon divergence as a cost function to train generative adversarial networks which intrinsically avoids mode collapse and encourages diversity.

Abstract: Voltage control is crucial to large-scale power system reliable operation, as timely reactive power support can help prevent widespread outages. However, there is currently no built in mechanism for power systems to ensure that the voltage control objective to maintain reliable operation will survive or sustain the uncertainty caused under adversary presence. Hence, this work introduces a Bayesian Reinforcement Learning (BRL) approach for power system control problems, with focus on sustained voltage control under uncertainty in a cyber-adversarial environment. This work proposes a data-driven BRL-based approach for automatic voltage control by formulating and solving a Partially-Observable Markov Decision Problem (POMDP), where the states are partially observable due to cyber intrusions. The techniques are evaluated on the WSCC and IEEE 14 bus systems. Additionally, BRL techniques assist in automatically finding a threshold for exploration and exploitation in various RL techniques.

Abstract: This paper introduces FairDP, a novel mechanism designed to simultaneously ensure differential privacy (DP) and fairness. FairDP operates by independently training models for distinct individual groups, using group-specific clipping terms to assess and bound the disparate impacts of DP. Throughout the training process, the mechanism progressively integrates knowledge from group models to formulate a comprehensive model that balances privacy, utility, and fairness in downstream tasks. Extensive theoretical and empirical analyses validate the efficacy of FairDP, demonstrating improved trade-offs between model utility, privacy, and fairness compared with existing methods.

Abstract: We provide several new results on the sample complexity of vector-valued linear predictors (parameterized by a matrix), and more generally neural networks. Focusing on size-independent bounds, where only the Frobenius norm distance of the parameters from some fixed reference matrix $W_0$ is controlled, we show that the sample complexity behavior can be surprisingly different than what we may expect considering the well-studied setting of scalar-valued linear predictors. This also leads to new sample complexity bounds for feed-forward neural networks, tackling some open questions in the literature, and establishing a new convex linear prediction problem that is provably learnable without uniform convergence.

Abstract: This paper considers a class of reinforcement learning problems, which involve systems with two types of states: stochastic and pseudo-stochastic. In such systems, stochastic states follow a stochastic transition kernel while the transitions of pseudo-stochastic states are deterministic given the stochastic states/transitions. We refer to such systems as mixed systems, which are widely used in various applications, including manufacturing systems, communication networks, and queueing networks. We propose a sample efficient RL method that accelerates learning by generating augmented data samples. The proposed algorithm is data-driven and learns the policy from data samples from both real and augmented samples. This method significantly improves learning by reducing the sample complexity such that the dataset only needs to have sufficient coverage of the stochastic states. We analyze the sample complexity of the proposed method under Fitted Q Iteration (FQI) and demonstrate that the optimality gap decreases as $\tilde{\mathcal{O}}(\sqrt{{1}/{n}}+\sqrt{{1}/{m}}),$ where $n$ is the number of real samples and $m$ is the number of augmented samples per real sample. It is important to note that without augmented samples, the optimality gap is $\tilde{\mathcal{O}}(1)$ due to insufficient data coverage of the pseudo-stochastic states. Our experimental results on multiple queueing network applications confirm that the proposed method indeed significantly accelerates learning in both deep Q-learning and deep policy gradient.

Abstract: In Continual Learning (CL), a model is required to learn a stream of tasks sequentially without significant performance degradation on previously learned tasks. Current approaches fail for a long sequence of tasks from diverse domains and difficulties. Many of the existing CL approaches are difficult to apply in practice due to excessive memory cost or training time, or are tightly coupled to a single device. With the intuition derived from the widely applied mini-batch training, we propose Batch Model Consolidation ($\textbf{BMC}$) to support more realistic CL under conditions where multiple agents are exposed to a range of tasks. During a $\textit{regularization}$ phase, BMC trains multiple $\textit{expert models}$ in parallel on a set of disjoint tasks. Each expert maintains weight similarity to a $\textit{base model}$ through a $\textit{stability loss}$, and constructs a $\textit{buffer}$ from a fraction of the task's data. During the $\textit{consolidation}$ phase, we combine the learned knowledge on 'batches' of $\textit{expert models}$ using a $\textit{batched consolidation loss}$ in $\textit{memory}$ data that aggregates all buffers. We thoroughly evaluate each component of our method in an ablation study and demonstrate the effectiveness on standardized benchmark datasets Split-CIFAR-100, Tiny-ImageNet, and the Stream dataset composed of 71 image classification tasks from diverse domains and difficulties. Our method outperforms the next best CL approach by 70% and is the only approach that can maintain performance at the end of 71 tasks; Our benchmark can be accessed at https://github.com/fostiropoulos/stream_benchmark

Abstract: The well-established practice of time series analysis involves estimating deterministic, non-stationary trend and seasonality components followed by learning the residual stochastic, stationary components. Recently, it has been shown that one can learn the deterministic non-stationary components accurately using multivariate Singular Spectrum Analysis (mSSA) in the absence of a correlated stationary component; meanwhile, in the absence of deterministic non-stationary components, the Autoregressive (AR) stationary component can also be learnt readily, e.g. via Ordinary Least Squares (OLS). However, a theoretical underpinning of multi-stage learning algorithms involving both deterministic and stationary components has been absent in the literature despite its pervasiveness. We resolve this open question by establishing desirable theoretical guarantees for a natural two-stage algorithm, where mSSA is first applied to estimate the non-stationary components despite the presence of a correlated stationary AR component, which is subsequently learned from the residual time series. We provide a finite-sample forecasting consistency bound for the proposed algorithm, SAMoSSA, which is data-driven and thus requires minimal parameter tuning. To establish theoretical guarantees, we overcome three hurdles: (i) we characterize the spectra of Page matrices of stable AR processes, thus extending the analysis of mSSA; (ii) we extend the analysis of AR process identification in the presence of arbitrary bounded perturbations; (iii) we characterize the out-of-sample or forecasting error, as opposed to solely considering model identification. Through representative empirical studies, we validate the superior performance of SAMoSSA compared to existing baselines. Notably, SAMoSSA's ability to account for AR noise structure yields improvements ranging from 5% to 37% across various benchmark datasets.

Abstract: Anomaly detection tools and methods present a key capability in modern cyberphysical and failure prediction systems. Despite the fast-paced development in deep learning architectures for anomaly detection, model optimization for a given dataset is a cumbersome and time consuming process. Neuroevolution could be an effective and efficient solution to this problem, as a fully automated search method for learning optimal neural networks, supporting both gradient and non-gradient fine tuning. However, existing methods mostly focus on optimizing model architectures without taking into account feature subspaces and model weights. In this work, we propose Anomaly Detection Neuroevolution (AD-NEv) - a scalable multi-level optimized neuroevolution framework for multivariate time series anomaly detection. The method represents a novel approach to synergically: i) optimize feature subspaces for an ensemble model based on the bagging technique; ii) optimize the model architecture of single anomaly detection models; iii) perform non-gradient fine-tuning of network weights. An extensive experimental evaluation on widely adopted multivariate anomaly detection benchmark datasets shows that the models extracted by AD-NEv outperform well-known deep learning architectures for anomaly detection. Moreover, results show that AD-NEv can perform the whole process efficiently, presenting high scalability when multiple GPUs are available.

Abstract: Imitation learning methods seek to learn from an expert either through behavioral cloning (BC) of the policy or inverse reinforcement learning (IRL) of the reward. Such methods enable agents to learn complex tasks from humans that are difficult to capture with hand-designed reward functions. Choosing BC or IRL for imitation depends on the quality and state-action coverage of the demonstrations, as well as additional access to the Markov decision process. Hybrid strategies that combine BC and IRL are not common, as initial policy optimization against inaccurate rewards diminishes the benefit of pretraining the policy with BC. This work derives an imitation method that captures the strengths of both BC and IRL. In the entropy-regularized ('soft') reinforcement learning setting, we show that the behaviour-cloned policy can be used as both a shaped reward and a critic hypothesis space by inverting the regularized policy update. This coherency facilities fine-tuning cloned policies using the reward estimate and additional interactions with the environment. This approach conveniently achieves imitation learning through initial behaviour cloning, followed by refinement via RL with online or offline data sources. The simplicity of the approach enables graceful scaling to high-dimensional and vision-based tasks, with stable learning and minimal hyperparameter tuning, in contrast to adversarial approaches.

Abstract: We study the fundamental mistake bound and sample complexity in the strategic classification, where agents can strategically manipulate their feature vector up to an extent in order to be predicted as positive. For example, given a classifier determining college admission, student candidates may try to take easier classes to improve their GPA, retake SAT and change schools in an effort to fool the classifier. Ball manipulations are a widely studied class of manipulations in the literature, where agents can modify their feature vector within a bounded radius ball. Unlike most prior work, our work considers manipulations to be personalized, meaning that agents can have different levels of manipulation abilities (e.g., varying radii for ball manipulations), and unknown to the learner. We formalize the learning problem in an interaction model where the learner first deploys a classifier and the agent manipulates the feature vector within their manipulation set to game the deployed classifier. We investigate various scenarios in terms of the information available to the learner during the interaction, such as observing the original feature vector before or after deployment, observing the manipulated feature vector, or not seeing either the original or the manipulated feature vector. We begin by providing online mistake bounds and PAC sample complexity in these scenarios for ball manipulations. We also explore non-ball manipulations and show that, even in the simplest scenario where both the original and the manipulated feature vectors are revealed, the mistake bounds and sample complexity are lower bounded by $\Omega(|\mathcal{H}|)$ when the target function belongs to a known class $\mathcal{H}$.

Abstract: Self-paced reinforcement learning (RL) aims to improve the data efficiency of learning by automatically creating sequences, namely curricula, of probability distributions over contexts. However, existing techniques for self-paced RL fail in long-horizon planning tasks that involve temporally extended behaviors. We hypothesize that taking advantage of prior knowledge about the underlying task structure can improve the effectiveness of self-paced RL. We develop a self-paced RL algorithm guided by reward machines, i.e., a type of finite-state machine that encodes the underlying task structure. The algorithm integrates reward machines in 1) the update of the policy and value functions obtained by any RL algorithm of choice, and 2) the update of the automated curriculum that generates context distributions. Our empirical results evidence that the proposed algorithm achieves optimal behavior reliably even in cases in which existing baselines cannot make any meaningful progress. It also decreases the curriculum length and reduces the variance in the curriculum generation process by up to one-fourth and four orders of magnitude, respectively.

Abstract: We present a PTAS for learning random constant-depth networks. We show that for any fixed $\epsilon>0$ and depth $i$, there is a poly-time algorithm that for any distribution on $\sqrt{d} \cdot \mathbb{S}^{d-1}$ learns random Xavier networks of depth $i$, up to an additive error of $\epsilon$. The algorithm runs in time and sample complexity of $(\bar{d})^{\mathrm{poly}(\epsilon^{-1})}$, where $\bar d$ is the size of the network. For some cases of sigmoid and ReLU-like activations the bound can be improved to $(\bar{d})^{\mathrm{polylog}(\epsilon^{-1})}$, resulting in a quasi-poly-time algorithm for learning constant depth random networks.

Abstract: A multivariate time series refers to observations of two or more variables taken from a device or a system simultaneously over time. There is an increasing need to monitor multivariate time series and detect anomalies in real time to ensure proper system operation and good service quality. It is also highly desirable to have a lightweight anomaly detection system that considers correlations between different variables, adapts to changes in the pattern of the multivariate time series, offers immediate responses, and provides supportive information regarding detection results based on unsupervised learning and online model training. In the past decade, many multivariate time series anomaly detection approaches have been introduced. However, they are unable to offer all the above-mentioned features. In this paper, we propose RoLA, a real-time online lightweight anomaly detection system for multivariate time series based on a divide-and-conquer strategy, parallel processing, and the majority rule. RoLA employs multiple lightweight anomaly detectors to monitor multivariate time series in parallel, determine the correlations between variables dynamically on the fly, and then jointly detect anomalies based on the majority rule in real time. To demonstrate the performance of RoLA, we conducted an experiment based on a public dataset provided by the FerryBox of the One Ocean Expedition. The results show that RoLA provides satisfactory detection accuracy and lightweight performance.

Abstract: Sliding window sums are widely used for string indexing, hashing and time series analysis. We have developed a family of the generic vectorized sliding sum algorithms that provide speedup of O(P/w) for window size $w$ and number of processors P. For a sum with a commutative operator the speedup is improved to O(P/log(w)). Even more important, our algorithms exhibit efficient memory access patterns. In this paper we study the application of the sliding sum algorithms to the training and inference of the Deep Neural Networks. We demonstrate how both pooling and convolution primitives could be expressed as sliding sums and evaluated by the compute kernels with the shared structure. We show that the sliding sum convolution kernels are more efficient than the commonly used GEMM kernels on the CPU, and could even outperform their GPU counterparts.

Abstract: Deep Reinforcement Learning (DRL) has demonstrated promising capability in solving complex control problems. However, DRL applications in safety-critical systems are hindered by the inherent lack of robust verification techniques to assure their performance in such applications. One of the key requirements of the verification process is the development of effective techniques to explain the system functionality, i.e., why the system produces specific results in given circumstances. Recently, interpretation methods based on the Counterfactual (CF) explanation approach have been proposed to address the problem of explanation in DRLs. This paper proposes a novel CF explanation framework to explain the decisions made by a black-box DRL. To evaluate the efficacy of the proposed explanation framework, we carried out several experiments in the domains of automated driving systems and Atari Pong game. Our analysis demonstrates that the proposed framework generates plausible and meaningful explanations for various decisions made by deep underlying DRLs. Source codes are available at: \url{https://github.com/Amir-Samadi/Counterfactual-Explanation}

Abstract: Contrastive learning (CL) has emerged as a powerful technique for representation learning, with or without label supervision. However, supervised CL is prone to collapsing representations of subclasses within a class by not capturing all their features, and unsupervised CL may suppress harder class-relevant features by focusing on learning easy class-irrelevant features; both significantly compromise representation quality. Yet, there is no theoretical understanding of \textit{class collapse} or \textit{feature suppression} at \textit{test} time. We provide the first unified theoretically rigorous framework to determine \textit{which} features are learnt by CL. Our analysis indicate that, perhaps surprisingly, bias of (stochastic) gradient descent towards finding simpler solutions is a key factor in collapsing subclass representations and suppressing harder class-relevant features. Moreover, we present increasing embedding dimensionality and improving the quality of data augmentations as two theoretically motivated solutions to {feature suppression}. We also provide the first theoretical explanation for why employing supervised and unsupervised CL together yields higher-quality representations, even when using commonly-used stochastic gradient methods.

Abstract: We propose the first theoretical and methodological framework for Gaussian process regression subject to privacy constraints. The proposed method can be used when a data owner is unwilling to share a high-fidelity supervised learning model built from their data with the public due to privacy concerns. The key idea of the proposed method is to add synthetic noise to the data until the predictive variance of the Gaussian process model reaches a prespecified privacy level. The optimal covariance matrix of the synthetic noise is formulated in terms of semi-definite programming. We also introduce the formulation of privacy-aware solutions under continuous privacy constraints using kernel-based approaches, and study their theoretical properties. The proposed method is illustrated by considering a model that tracks the trajectories of satellites.