Machine Learning (cs.LG)

Abstract: Adversarial Training (AT) has been found to substantially improve the robustness of deep learning classifiers against adversarial attacks. AT involves obtaining robustness by including adversarial examples in training a classifier. Most variants of AT algorithms treat every training example equally. However, recent works have shown that better performance is achievable by treating them unequally. In addition, it has been observed that AT exerts an uneven influence on different classes in a training set and unfairly hurts examples corresponding to classes that are inherently harder to classify. Consequently, various reweighting schemes have been proposed that assign unequal weights to robust losses of individual examples in a training set. In this work, we propose a novel instance-wise reweighting scheme. It considers the vulnerability of each natural example and the resulting information loss on its adversarial counterpart occasioned by adversarial attacks. Through extensive experiments, we show that our proposed method significantly improves over existing reweighting schemes, especially against strong white and black-box attacks.

Abstract: The widespread application of Reinforcement Learning (RL) in real-world situations is yet to come to fruition, largely as a result of its failure to satisfy the essential safety demands of such systems. Existing safe reinforcement learning (SafeRL) methods, employing cost functions to enhance safety, fail to achieve zero-cost in complex scenarios, including vision-only tasks, even with comprehensive data sampling and training. To address this, we introduce Safe DreamerV3, a novel algorithm that integrates both Lagrangian-based and planning-based methods within a world model. Our methodology represents a significant advancement in SafeRL as the first algorithm to achieve nearly zero-cost in both low-dimensional and vision-only tasks within the Safety-Gymnasium benchmark. Our project website can be found in: https://sites.google.com/view/safedreamerv3.

Abstract: Even nowadays, where Deep Learning (DL) has achieved state-of-the-art performance in a wide range of research domains, accelerating training and building robust DL models remains a challenging task. To this end, generations of researchers have pursued to develop robust methods for training DL architectures that can be less sensitive to weight distributions, model architectures and loss landscapes. However, such methods are limited to adaptive learning rate optimizers, initialization schemes, and clipping gradients without investigating the fundamental rule of parameters update. Although multiplicative updates have contributed significantly to the early development of machine learning and hold strong theoretical claims, to best of our knowledge, this is the first work that investigate them in context of DL training acceleration and robustness. In this work, we propose an optimization framework that fits to a wide range of optimization algorithms and enables one to apply alternative update rules. To this end, we propose a novel multiplicative update rule and we extend their capabilities by combining it with a traditional additive update term, under a novel hybrid update method. We claim that the proposed framework accelerates training, while leading to more robust models in contrast to traditionally used additive update rule and we experimentally demonstrate their effectiveness in a wide range of task and optimization methods. Such tasks ranging from convex and non-convex optimization to difficult image classification benchmarks applying a wide range of traditionally used optimization methods and Deep Neural Network (DNN) architectures.

Abstract: Load forecasting is of great significance in the power industry as it can provide a reference for subsequent tasks such as power grid dispatch, thus bringing huge economic benefits. However, there are many differences between load forecasting and traditional time series forecasting. On the one hand, load forecasting aims to minimize the cost of subsequent tasks such as power grid dispatch, rather than simply pursuing prediction accuracy. On the other hand, the load is largely influenced by many external factors, such as temperature or calendar variables. In addition, the scale of predictions (such as building-level loads and aggregated-level loads) can also significantly impact the predicted results. In this paper, we provide a comprehensive load forecasting archive, which includes load domain-specific feature engineering to help forecasting models better model load data. In addition, different from the traditional loss function which only aims for accuracy, we also provide a method to customize the loss function based on the forecasting error, integrating it into our forecasting framework. Based on this, we conducted extensive experiments on load data at different levels, providing a reference for researchers to compare different load forecasting models.

Abstract: Controlling nonlinear dynamical systems using machine learning allows to not only drive systems into simple behavior like periodicity but also to more complex arbitrary dynamics. For this, it is crucial that a machine learning system can be trained to reproduce the target dynamics sufficiently well. On the example of forcing a chaotic parametrization of the Lorenz system into intermittent dynamics, we show first that classical reservoir computing excels at this task. In a next step, we compare those results based on different amounts of training data to an alternative setup, where next-generation reservoir computing is used instead. It turns out that while delivering comparable performance for usual amounts of training data, next-generation RC significantly outperforms in situations where only very limited data is available. This opens even further practical control applications in real world problems where data is restricted.

Abstract: Adversarial examples derived from deliberately crafted perturbations on visual inputs can easily harm decision process of deep neural networks. To prevent potential threats, various adversarial training-based defense methods have grown rapidly and become a de facto standard approach for robustness. Despite recent competitive achievements, we observe that adversarial vulnerability varies across targets and certain vulnerabilities remain prevalent. Intriguingly, such peculiar phenomenon cannot be relieved even with deeper architectures and advanced defense methods. To address this issue, in this paper, we introduce a causal approach called Adversarial Double Machine Learning (ADML), which allows us to quantify the degree of adversarial vulnerability for network predictions and capture the effect of treatments on outcome of interests. ADML can directly estimate causal parameter of adversarial perturbations per se and mitigate negative effects that can potentially damage robustness, bridging a causal perspective into the adversarial vulnerability. Through extensive experiments on various CNN and Transformer architectures, we corroborate that ADML improves adversarial robustness with large margins and relieve the empirical observation.

Abstract: Learning with expert advice and multi-armed bandit are two classic online decision problems which differ on how the information is observed in each round of the game. We study a family of problems interpolating the two. For a vector $\mathbf{m}=(m_1,\dots,m_K)\in \mathbb{N}^K$, an instance of $\mathbf{m}$-MAB indicates that the arms are partitioned into $K$ groups and the $i$-th group contains $m_i$ arms. Once an arm is pulled, the losses of all arms in the same group are observed. We prove tight minimax regret bounds for $\mathbf{m}$-MAB and design an optimal PAC algorithm for its pure exploration version, $\mathbf{m}$-BAI, where the goal is to identify the arm with minimum loss with as few rounds as possible. We show that the minimax regret of $\mathbf{m}$-MAB is $\Theta\left(\sqrt{T\sum_{k=1}^K\log (m_k+1)}\right)$ and the minimum number of pulls for an $(\epsilon,0.05)$-PAC algorithm of $\mathbf{m}$-BAI is $\Theta\left(\frac{1}{\epsilon^2}\cdot \sum_{k=1}^K\log (m_k+1)\right)$. Both our upper bounds and lower bounds for $\mathbf{m}$-MAB can be extended to a more general setting, namely the bandit with graph feedback, in terms of the clique cover and related graph parameters. As consequences, we obtained tight minimax regret bounds for several families of feedback graphs.

Abstract: The selection of model's parameters plays an important role in the application of support vector classification (SVC). The commonly used method of selecting model's parameters is the k-fold cross validation with grid search (CV). It is extremely time-consuming because it needs to train a large number of SVC models. In this paper, a new method is proposed to train SVC with the selection of model's parameters. Firstly, training SVC with the selection of model's parameters is modeled as a minimax optimization problem (MaxMin-L2-SVC-NCH), in which the minimization problem is an optimization problem of finding the closest points between two normal convex hulls (L2-SVC-NCH) while the maximization problem is an optimization problem of finding the optimal model's parameters. A lower time complexity can be expected in MaxMin-L2-SVC-NCH because CV is abandoned. A gradient-based algorithm is then proposed to solve MaxMin-L2-SVC-NCH, in which L2-SVC-NCH is solved by a projected gradient algorithm (PGA) while the maximization problem is solved by a gradient ascent algorithm with dynamic learning rate. To demonstrate the advantages of the PGA in solving L2-SVC-NCH, we carry out a comparison of the PGA and the famous sequential minimal optimization (SMO) algorithm after a SMO algorithm and some KKT conditions for L2-SVC-NCH are provided. It is revealed that the SMO algorithm is a special case of the PGA. Thus, the PGA can provide more flexibility. The comparative experiments between MaxMin-L2-SVC-NCH and the classical parameter selection models on public datasets show that MaxMin-L2-SVC-NCH greatly reduces the number of models to be trained and the test accuracy is not lost to the classical models. It indicates that MaxMin-L2-SVC-NCH performs better than the other models. We strongly recommend MaxMin-L2-SVC-NCH as a preferred model for SVC task.

Abstract: This paper proposes a novel approach to integrating partial differential equation (PDE)-based evolution models into neural networks through a new type of regularization. Specifically, we propose inverse evolution layers (IELs) based on evolution equations. These layers can achieve specific regularization objectives and endow neural networks' outputs with corresponding properties of the evolution models. Moreover, IELs are straightforward to construct and implement, and can be easily designed for various physical evolutions and neural networks. Additionally, the design process for these layers can provide neural networks with intuitive and mathematical interpretability, thus enhancing the transparency and explainability of the approach. To demonstrate the effectiveness, efficiency, and simplicity of our approach, we present an example of endowing semantic segmentation models with the smoothness property based on the heat diffusion model. To achieve this goal, we design heat-diffusion IELs and apply them to address the challenge of semantic segmentation with noisy labels. The experimental results demonstrate that the heat-diffusion IELs can effectively mitigate the overfitting problem caused by noisy labels.

Abstract: The objective of clusterability evaluation is to check whether a clustering structure exists within the data set. As a crucial yet often-overlooked issue in cluster analysis, it is essential to conduct such a test before applying any clustering algorithm. If a data set is unclusterable, any subsequent clustering analysis would not yield valid results. Despite its importance, the majority of existing studies focus on numerical data, leaving the clusterability evaluation issue for categorical data as an open problem. Here we present TestCat, a testing-based approach to assess the clusterability of categorical data in terms of an analytical $p$-value. The key idea underlying TestCat is that clusterable categorical data possess many strongly correlated attribute pairs and hence the sum of chi-squared statistics of all attribute pairs is employed as the test statistic for $p$-value calculation. We apply our method to a set of benchmark categorical data sets, showing that TestCat outperforms those solutions based on existing clusterability evaluation methods for numeric data. To the best of our knowledge, our work provides the first way to effectively recognize the clusterability of categorical data in a statistically sound manner.

Abstract: The sparsity of Deep Neural Networks is well investigated to maximize the performance and reduce the size of overparameterized networks as possible. Existing methods focus on pruning parameters in the training process by using thresholds and metrics. Meanwhile, feature similarity between different layers has not been discussed sufficiently before, which could be rigorously proved to be highly correlated to the network sparsity in this paper. Inspired by interlayer feature similarity in overparameterized models, we investigate the intrinsic link between network sparsity and interlayer feature similarity. Specifically, we prove that reducing interlayer feature similarity based on Centered Kernel Alignment (CKA) improves the sparsity of the network by using information bottleneck theory. Applying such theory, we propose a plug-and-play CKA-based Sparsity Regularization for sparse network training, dubbed CKA-SR, which utilizes CKA to reduce feature similarity between layers and increase network sparsity. In other words, layers of our sparse network tend to have their own identity compared to each other. Experimentally, we plug the proposed CKA-SR into the training process of sparse network training methods and find that CKA-SR consistently improves the performance of several State-Of-The-Art sparse training methods, especially at extremely high sparsity. Code is included in the supplementary materials.

Abstract: Finding (bi-)clusters in bipartite graphs is a popular data analysis approach. Analysts typically want to visualize the clusters, which is simple as long as the clusters are disjoint. However, many modern algorithms find overlapping clusters, making visualization more complicated. In this paper, we study the problem of visualizing \emph{a given clustering} of overlapping clusters in bipartite graphs and the related problem of visualizing Boolean Matrix Factorizations. We conceptualize three different objectives that any good visualization should satisfy: (1) proximity of cluster elements, (2) large consecutive areas of elements from the same cluster, and (3) large uninterrupted areas in the visualization, regardless of the cluster membership. We provide objective functions that capture these goals and algorithms that optimize these objective functions. Interestingly, in experiments on real-world datasets, we find that the best trade-off between these competing goals is achieved by a novel heuristic, which locally aims to place rows and columns with similar cluster membership next to each other.

Abstract: Despite widespread adoption in practice, guarantees for the LASSO and Group LASSO are strikingly lacking in settings beyond statistical problems, and these algorithms are usually considered to be a heuristic in the context of sparse convex optimization on deterministic inputs. We give the first recovery guarantees for the Group LASSO for sparse convex optimization with vector-valued features. We show that if a sufficiently large Group LASSO regularization is applied when minimizing a strictly convex function $l$, then the minimizer is a sparse vector supported on vector-valued features with the largest $\ell_2$ norm of the gradient. Thus, repeating this procedure selects the same set of features as the Orthogonal Matching Pursuit algorithm, which admits recovery guarantees for any function $l$ with restricted strong convexity and smoothness via weak submodularity arguments. This answers open questions of Tibshirani et al. and Yasuda et al. Our result is the first to theoretically explain the empirical success of the Group LASSO for convex functions under general input instances assuming only restricted strong convexity and smoothness. Our result also generalizes provable guarantees for the Sequential Attention algorithm, which is a feature selection algorithm inspired by the attention mechanism proposed by Yasuda et al. As an application of our result, we give new results for the column subset selection problem, which is well-studied when the loss is the Frobenius norm or other entrywise matrix losses. We give the first result for general loss functions for this problem that requires only restricted strong convexity and smoothness.

Abstract: We propose an improved convergence analysis technique that characterizes the distributed learning paradigm of federated learning (FL) with imperfect/noisy uplink and downlink communications. Such imperfect communication scenarios arise in the practical deployment of FL in emerging communication systems and protocols. The analysis developed in this paper demonstrates, for the first time, that there is an asymmetry in the detrimental effects of uplink and downlink communications in FL. In particular, the adverse effect of the downlink noise is more severe on the convergence of FL algorithms. Using this insight, we propose improved Signal-to-Noise (SNR) control strategies that, discarding the negligible higher-order terms, lead to a similar convergence rate for FL as in the case of a perfect, noise-free communication channel while incurring significantly less power resources compared to existing solutions. In particular, we establish that to maintain the $O(\frac{1}{\sqrt{K}})$ rate of convergence like in the case of noise-free FL, we need to scale down the uplink and downlink noise by $\Omega({\sqrt{k}})$ and $\Omega({k})$ respectively, where $k$ denotes the communication round, $k=1,\dots, K$. Our theoretical result is further characterized by two major benefits: firstly, it does not assume the somewhat unrealistic assumption of bounded client dissimilarity, and secondly, it only requires smooth non-convex loss functions, a function class better suited for modern machine learning and deep learning models. We also perform extensive empirical analysis to verify the validity of our theoretical findings.

Abstract: We introduce the problem of curriculum discovery and describe a curriculum learning framework capable of discovering effective curricula in a curriculum space based on prior knowledge about sample difficulty. Using annotation entropy and loss as measures of difficulty, we show that (i): the top-performing discovered curricula for a given model and dataset are often non-monotonic as opposed to monotonic curricula in existing literature, (ii): the prevailing easy-to-hard or hard-to-easy transition curricula are often at the risk of underperforming, and (iii): the curricula discovered for smaller datasets and models perform well on larger datasets and models respectively. The proposed framework encompasses some of the existing curriculum learning approaches and can discover curricula that outperform them across several NLP tasks.

Abstract: This paper studies a new problem, \emph{active learning with partial labels} (ALPL). In this setting, an oracle annotates the query samples with partial labels, relaxing the oracle from the demanding accurate labeling process. To address ALPL, we first build an intuitive baseline that can be seamlessly incorporated into existing AL frameworks. Though effective, this baseline is still susceptible to the \emph{overfitting}, and falls short of the representative partial-label-based samples during the query process. Drawing inspiration from human inference in cognitive science, where accurate inferences can be explicitly derived from \emph{counter-examples} (CEs), our objective is to leverage this human-like learning pattern to tackle the \emph{overfitting} while enhancing the process of selecting representative samples in ALPL. Specifically, we construct CEs by reversing the partial labels for each instance, and then we propose a simple but effective WorseNet to directly learn from this complementary pattern. By leveraging the distribution gap between WorseNet and the predictor, this adversarial evaluation manner could enhance both the performance of the predictor itself and the sample selection process, allowing the predictor to capture more accurate patterns in the data. Experimental results on five real-world datasets and four benchmark datasets show that our proposed method achieves comprehensive improvements over ten representative AL frameworks, highlighting the superiority of WorseNet. The source code will be available at \url{https://github.com/Ferenas/APLL}.

Abstract: Molecular property prediction has gained significant attention due to its transformative potential in multiple scientific disciplines. Conventionally, a molecule graph can be represented either as a graph-structured data or a SMILES text. Recently, the rapid development of Large Language Models (LLMs) has revolutionized the field of NLP. Although it is natural to utilize LLMs to assist in understanding molecules represented by SMILES, the exploration of how LLMs will impact molecular property prediction is still in its early stage. In this work, we advance towards this objective through two perspectives: zero/few-shot molecular classification, and using the new explanations generated by LLMs as representations of molecules. To be specific, we first prompt LLMs to do in-context molecular classification and evaluate their performance. After that, we employ LLMs to generate semantically enriched explanations for the original SMILES and then leverage that to fine-tune a small-scale LM model for multiple downstream tasks. The experimental results highlight the superiority of text explanations as molecular representations across multiple benchmark datasets, and confirm the immense potential of LLMs in molecular property prediction tasks. Codes are available at \url{https://github.com/ChnQ/LLM4Mol}.

Abstract: In recent years, the integration of Machine Learning (ML) models with Operation Research (OR) tools has gained popularity across diverse applications, including cancer treatment, algorithmic configuration, and chemical process optimization. In this domain, the combination of ML and OR often relies on representing the ML model output using Mixed Integer Programming (MIP) formulations. Numerous studies in the literature have developed such formulations for many ML predictors, with a particular emphasis on Artificial Neural Networks (ANNs) due to their significant interest in many applications. However, ANNs frequently contain a large number of parameters, resulting in MIP formulations that are impractical to solve, thereby impeding scalability. In fact, the ML community has already introduced several techniques to reduce the parameter count of ANNs without compromising their performance, since the substantial size of modern ANNs presents challenges for ML applications as it significantly impacts computational efforts during training and necessitates significant memory resources for storage. In this paper, we showcase the effectiveness of pruning, one of these techniques, when applied to ANNs prior to their integration into MIPs. By pruning the ANN, we achieve significant improvements in the speed of the solution process. We discuss why pruning is more suitable in this context compared to other ML compression techniques, and we identify the most appropriate pruning strategies. To highlight the potential of this approach, we conduct experiments using feed-forward neural networks with multiple layers to construct adversarial examples. Our results demonstrate that pruning offers remarkable reductions in solution times without hindering the quality of the final decision, enabling the resolution of previously unsolvable instances.

Abstract: Federated learning (FL) combined with differential privacy (DP) offers machine learning (ML) training with distributed devices and with a formal privacy guarantee. With a large population of devices, FL with DP produces a performant model in a timely manner. However, for applications with a smaller population, not only does the model utility degrade as the DP noise is inversely proportional to population, but also the training latency increases since waiting for enough clients to become available from a smaller pool is slower. In this work, we thus propose expanding the population based on domain adaptation techniques to speed up the training and improves the final model quality when training with small populations. We empirically demonstrate that our techniques can improve the utility by 13% to 30% on real-world language modeling datasets.

Abstract: Unsupervised domain adaptation (UDA) has witnessed remarkable advancements in improving the accuracy of models for unlabeled target domains. However, the calibration of predictive uncertainty in the target domain, a crucial aspect of the safe deployment of UDA models, has received limited attention. The conventional in-domain calibration method, \textit{temperature scaling} (TempScal), encounters challenges due to domain distribution shifts and the absence of labeled target domain data. Recent approaches have employed importance-weighting techniques to estimate the target-optimal temperature based on re-weighted labeled source data. Nonetheless, these methods require source data and suffer from unreliable density estimates under severe domain shifts, rendering them unsuitable for source-free UDA settings. To overcome these limitations, we propose PseudoCal, a source-free calibration method that exclusively relies on unlabeled target data. Unlike previous approaches that treat UDA calibration as a \textit{covariate shift} problem, we consider it as an unsupervised calibration problem specific to the target domain. Motivated by the factorization of the negative log-likelihood (NLL) objective in TempScal, we generate a labeled pseudo-target set that captures the structure of the real target. By doing so, we transform the unsupervised calibration problem into a supervised one, enabling us to effectively address it using widely-used in-domain methods like TempScal. Finally, we thoroughly evaluate the calibration performance of PseudoCal by conducting extensive experiments on 10 UDA methods, considering both traditional UDA settings and recent source-free UDA scenarios. The experimental results consistently demonstrate the superior performance of PseudoCal, exhibiting significantly reduced calibration error compared to existing calibration methods.

Abstract: Models derived from other models are extremely common in machine learning (ML) today. For example, transfer learning is used to create task-specific models from "pre-trained" models through finetuning. This has led to an ecosystem where models are related to each other, sharing structure and often even parameter values. However, it is hard to manage these model derivatives: the storage overhead of storing all derived models quickly becomes onerous, prompting users to get rid of intermediate models that might be useful for further analysis. Additionally, undesired behaviors in models are hard to track down (e.g., is a bug inherited from an upstream model?). In this paper, we propose a model versioning and management system called MGit that makes it easier to store, test, update, and collaborate on model derivatives. MGit introduces a lineage graph that records provenance and versioning information between models, optimizations to efficiently store model parameters, as well as abstractions over this lineage graph that facilitate relevant testing, updating and collaboration functionality. MGit is able to reduce the lineage graph's storage footprint by up to 7x and automatically update downstream models in response to updates to upstream models.

Abstract: The monotonic dependence of the outputs of a neural network on some of its inputs is a crucial inductive bias in many scenarios where domain knowledge dictates such behavior. This is especially important for interpretability and fairness considerations. In a broader context, scenarios in which monotonicity is important can be found in finance, medicine, physics, and other disciplines. It is thus desirable to build neural network architectures that implement this inductive bias provably. In this work, we propose a weight-constrained architecture with a single residual connection to achieve exact monotonic dependence in any subset of the inputs. The weight constraint scheme directly controls the Lipschitz constant of the neural network and thus provides the additional benefit of robustness. Compared to currently existing techniques used for monotonicity, our method is simpler in implementation and in theory foundations, has negligible computational overhead, is guaranteed to produce monotonic dependence, and is highly expressive. We show how the algorithm is used to train powerful, robust, and interpretable discriminators that achieve competitive performance compared to current state-of-the-art methods across various benchmarks, from social applications to the classification of the decays of subatomic particles produced at the CERN Large Hadron Collider.