Machine Learning (cs.LG)

Abstract: Time-Series Mining (TSM) is an important research area since it shows great potential in practical applications. Deep learning models that rely on massive labeled data have been utilized for TSM successfully. However, constructing a large-scale well-labeled dataset is difficult due to data annotation costs. Recently, Pre-Trained Models have gradually attracted attention in the time series domain due to their remarkable performance in computer vision and natural language processing. In this survey, we provide a comprehensive review of Time-Series Pre-Trained Models (TS-PTMs), aiming to guide the understanding, applying, and studying TS-PTMs. Specifically, we first briefly introduce the typical deep learning models employed in TSM. Then, we give an overview of TS-PTMs according to the pre-training techniques. The main categories we explore include supervised, unsupervised, and self-supervised TS-PTMs. Further, extensive experiments are conducted to analyze the advantages and disadvantages of transfer learning strategies, Transformer-based models, and representative TS-PTMs. Finally, we point out some potential directions of TS-PTMs for future work.

Abstract: Non-stationary multi-armed bandit (NS-MAB) problems have recently received significant attention. NS-MAB are typically modelled in two scenarios: abruptly changing, where reward distributions remain constant for a certain period and change at unknown time steps, and smoothly changing, where reward distributions evolve smoothly based on unknown dynamics. In this paper, we propose Discounted Thompson Sampling (DS-TS) with Gaussian priors to address both non-stationary settings. Our algorithm passively adapts to changes by incorporating a discounted factor into Thompson Sampling. DS-TS method has been experimentally validated, but analysis of the regret upper bound is currently lacking. Under mild assumptions, we show that DS-TS with Gaussian priors can achieve nearly optimal regret bound on the order of $\tilde{O}(\sqrt{TB_T})$ for abruptly changing and $\tilde{O}(T^{\beta})$ for smoothly changing, where $T$ is the number of time steps, $B_T$ is the number of breakpoints, $\beta$ is associated with the smoothly changing environment and $\tilde{O}$ hides the parameters independent of $T$ as well as logarithmic terms. Furthermore, empirical comparisons between DS-TS and other non-stationary bandit algorithms demonstrate its competitive performance. Specifically, when prior knowledge of the maximum expected reward is available, DS-TS has the potential to outperform state-of-the-art algorithms.

Abstract: Long-term time series forecasting has gained significant attention in recent years. While there are various specialized designs for capturing temporal dependency, previous studies have demonstrated that a single linear layer can achieve competitive forecasting performance compared to other complex architectures. In this paper, we thoroughly investigate the intrinsic effectiveness of recent approaches and make three key observations: 1) linear mapping is critical to prior long-term time series forecasting efforts; 2) RevIN (reversible normalization) and CI (Channel Independent) play a vital role in improving overall forecasting performance; and 3) linear mapping can effectively capture periodic features in time series and has robustness for different periods across channels when increasing the input horizon. We provide theoretical and experimental explanations to support our findings and also discuss the limitations and future works. Our framework's code is available at \url{https://github.com/plumprc/RTSF}.

Abstract: Although Federated Learning (FL) enables global model training across clients without compromising their raw data, existing Federated Averaging (FedAvg)-based methods suffer from the problem of low inference performance, especially for unevenly distributed data among clients. This is mainly because i) FedAvg initializes client models with the same global models, which makes the local training hard to escape from the local search for optimal solutions; and ii) by averaging model parameters in a coarse manner, FedAvg eclipses the individual characteristics of local models. To address such issues that strongly limit the inference capability of FL, we propose a novel and effective FL paradigm named FedMR (Federated Model Recombination). Unlike conventional FedAvg-based methods, the cloud server of FedMR shuffles each layer of collected local models and recombines them to achieve new models for local training on clients. Due to the diversified initialization models for clients coupled with fine-grained model recombination, FedMR can converge to a well-generalized global model for all the clients, leading to a superior inference performance. Experimental results show that, compared with state-of-the-art FL methods, FedMR can significantly improve inference accuracy in a quicker manner without exposing client privacy.

Abstract: Deep graph clustering has recently received significant attention due to its ability to enhance the representation learning capabilities of models in unsupervised scenarios. Nevertheless, deep clustering for temporal graphs, which could capture crucial dynamic interaction information, has not been fully explored. It means that in many clustering-oriented real-world scenarios, temporal graphs can only be processed as static graphs. This not only causes the loss of dynamic information but also triggers huge computational consumption. To solve the problem, we propose a general framework for deep Temporal Graph Clustering called TGC, which adjusts deep clustering techniques (clustering assignment distribution and adjacency matrix reconstruction) to suit the interaction sequence-based batch-processing pattern of temporal graphs. In addition, we discuss differences between temporal graph clustering and existing static graph clustering from several levels. To verify the superiority of the proposed framework TGC, we conduct extensive experiments. The experimental results show that temporal graph clustering enables more flexibility in finding a balance between time and space requirements, and our framework can effectively improve the performance of existing temporal graph learning methods. Our code and supplementary material will be released after publication.

Abstract: Animal-borne sensors ('bio-loggers') can record a suite of kinematic and environmental data, which can elucidate animal ecophysiology and improve conservation efforts. Machine learning techniques are useful for interpreting the large amounts of data recorded by bio-loggers, but there exists no standard for comparing the different machine learning techniques in this domain. To address this, we present the Bio-logger Ethogram Benchmark (BEBE), a collection of datasets with behavioral annotations, standardized modeling tasks, and evaluation metrics. BEBE is to date the largest, most taxonomically diverse, publicly available benchmark of this type, and includes 1654 hours of data collected from 149 individuals across nine taxa. We evaluate the performance of ten different machine learning methods on BEBE, and identify key challenges to be addressed in future work. Datasets, models, and evaluation code are made publicly available at https://github.com/earthspecies/BEBE, to enable community use of BEBE as a point of comparison in methods development.

Abstract: Prior beliefs about the latent function to shape inductive biases can be incorporated into a Gaussian Process (GP) via the kernel. However, beyond kernel choices, the decision-making process of GP models remains poorly understood. In this work, we contribute an analysis of the loss landscape for GP models using methods from physics. We demonstrate $\nu$-continuity for Matern kernels and outline aspects of catastrophe theory at critical points in the loss landscape. By directly including $\nu$ in the hyperparameter optimisation for Matern kernels, we find that typical values of $\nu$ are far from optimal in terms of performance, yet prevail in the literature due to the increased computational speed. We also provide an a priori method for evaluating the effect of GP ensembles and discuss various voting approaches based on physical properties of the loss landscape. The utility of these approaches is demonstrated for various synthetic and real datasets. Our findings provide an enhanced understanding of the decision-making process behind GPs and offer practical guidance for improving their performance and interpretability in a range of applications.

Abstract: Recent years have witnessed the great success of Graph Neural Networks (GNNs) in handling graph-related tasks. However, MLPs remain the primary workhorse for practical industrial applications due to their desirable inference efficiency and scalability. To reduce their gaps, one can directly distill knowledge from a well-designed teacher GNN to a student MLP, which is termed as GNN-to-MLP distillation. However, the process of distillation usually entails a loss of information, and ``which knowledge patterns of GNNs are more likely to be left and distilled into MLPs?" becomes an important question. In this paper, we first factorize the knowledge learned by GNNs into low- and high-frequency components in the spectral domain and then derive their correspondence in the spatial domain. Furthermore, we identified a potential information drowning problem for existing GNN-to-MLP distillation, i.e., the high-frequency knowledge of the pre-trained GNNs may be overwhelmed by the low-frequency knowledge during distillation; we have described in detail what it represents, how it arises, what impact it has, and how to deal with it. In this paper, we propose an efficient Full-Frequency GNN-to-MLP (FF-G2M) distillation framework, which extracts both low-frequency and high-frequency knowledge from GNNs and injects it into MLPs. Extensive experiments show that FF-G2M improves over the vanilla MLPs by 12.6% and outperforms its corresponding teacher GNNs by 2.6% averaged over six graph datasets and three common GNN architectures.

Abstract: The layout design of pipelines is a critical task in the construction industry. Currently, pipeline layout is designed manually by engineers, which is time-consuming and laborious. Automating and streamlining this process can reduce the burden on engineers and save time. In this paper, we propose a method for generating three-dimensional layout of pipelines based on deep reinforcement learning (DRL). Firstly, we abstract the geometric features of space to establish a training environment and define reward functions based on three constraints: pipeline length, elbow, and installation distance. Next, we collect data through interactions between the agent and the environment and train the DRL model. Finally, we use the well-trained DRL model to automatically design a single pipeline. Our results demonstrate that DRL models can complete the pipeline layout task in space in a much shorter time than traditional algorithms while ensuring high-quality layout outcomes.

Abstract: Diffusion Probability Models (DPMs) have made impressive advancements in various machine learning domains. However, achieving high-quality synthetic samples typically involves performing a large number of sampling steps, which impedes the possibility of real-time sample synthesis. Traditional accelerated sampling algorithms via knowledge distillation rely on pre-trained model weights and discrete time step scenarios, necessitating additional training sessions to achieve their goals. To address these issues, we propose the Catch-Up Distillation (CUD), which encourages the current moment output of the velocity estimation model ``catch up'' with its previous moment output. Specifically, CUD adjusts the original Ordinary Differential Equation (ODE) training objective to align the current moment output with both the ground truth label and the previous moment output, utilizing Runge-Kutta-based multi-step alignment distillation for precise ODE estimation while preventing asynchronous updates. Furthermore, we investigate the design space for CUDs under continuous time-step scenarios and analyze how to determine the suitable strategies. To demonstrate CUD's effectiveness, we conduct thorough ablation and comparison experiments on CIFAR-10, MNIST, and ImageNet-64. On CIFAR-10, we obtain a FID of 2.80 by sampling in 15 steps under one-session training and the new state-of-the-art FID of 3.37 by sampling in one step with additional training. This latter result necessitated only 62w iterations with a batch size of 128, in contrast to Consistency Distillation, which demanded 210w iterations with a larger batch size of 256.

Abstract: Recent years have witnessed the rapid development of heterogeneous graph neural networks (HGNNs) in information retrieval (IR) applications. Many existing HGNNs design a variety of tailor-made graph convolutions to capture structural and semantic information in heterogeneous graphs. However, existing HGNNs usually represent each node as a single vector in the multi-layer graph convolution calculation, which makes the high-level graph convolution layer fail to distinguish information from different relations and different orders, resulting in the information loss in the message passing. %insufficient mining of information. To this end, we propose a novel heterogeneous graph neural network with sequential node representation, namely Seq-HGNN. To avoid the information loss caused by the single vector node representation, we first design a sequential node representation learning mechanism to represent each node as a sequence of meta-path representations during the node message passing. Then we propose a heterogeneous representation fusion module, empowering Seq-HGNN to identify important meta-paths and aggregate their representations into a compact one. We conduct extensive experiments on four widely used datasets from Heterogeneous Graph Benchmark (HGB) and Open Graph Benchmark (OGB). Experimental results show that our proposed method outperforms state-of-the-art baselines in both accuracy and efficiency. The source code is available at https://github.com/nobrowning/SEQ_HGNN.

Abstract: Despite the potential benefits of Diffusion Models for NLP applications, publicly available implementations, trained models, or reproducible training procedures currently need to be publicly available. We present the Democratized Diffusion Language Model (DDLM), based on the Continuous Diffusion for Categorical Data (CDCD) framework, to address these challenges. We propose a simplified training procedure for DDLM using the C4 dataset and perform an in-depth analysis of the trained model's behavior. Furthermore, we introduce a novel early-exiting strategy for faster sampling with models trained with score interpolation. Since no previous works aimed at solving downstream tasks with pre-trained Diffusion LM (e.g., classification tasks), we experimented with GLUE Benchmark to study the ability of DDLM to transfer knowledge. With this paper, we propose available training and evaluation pipelines to other researchers and pre-trained DDLM models, which could be used in future research with Diffusion LMs.

Abstract: In this paper, we propose Ahead-of-Time (AoT) P-Tuning, a novel parameter-efficient fine-tuning method for pre-trained Language Models (LMs) that adds input-dependent bias before each Transformer layer. We evaluate AoT P-Tuning on GLUE and SuperGLUE benchmarking datasets using RoBERTa and DeBERTa models, showing that it outperforms BitFit and is comparable or better than other baseline methods for efficient fine-tuning. Additionally, we assess the inference overhead of AoT P-Tuning and demonstrate that it introduces negligible overhead compared to established baseline methods. Our method enables multi-task inference with a single backbone LM, making it a practical solution for real-world applications.

Abstract: Recent years have witnessed the growing popularity of domain-specific accelerators (DSAs), such as Google's TPUs, for accelerating various applications such as deep learning, search, autonomous driving, etc. To facilitate DSA designs, high-level synthesis (HLS) is used, which allows a developer to compile a high-level description in the form of software code in C and C++ into a design in low-level hardware description languages (such as VHDL or Verilog) and eventually synthesized into a DSA on an ASIC (application-specific integrated circuit) or FPGA (field-programmable gate arrays). However, existing HLS tools still require microarchitecture decisions, expressed in terms of pragmas (such as directives for parallelization and pipelining). To enable more people to design DSAs, it is desirable to automate such decisions with the help of deep learning for predicting the quality of HLS designs. This requires us a deeper understanding of the program, which is a combination of original code and pragmas. Naturally, these programs can be considered as sequence data, for which large language models (LLM) can help. In addition, these programs can be compiled and converted into a control data flow graph (CDFG), and the compiler also provides fine-grained alignment between the code tokens and the CDFG nodes. However, existing works either fail to leverage both modalities or combine the two in shallow or coarse ways. We propose ProgSG allowing the source code sequence modality and the graph modalities to interact with each other in a deep and fine-grained way. To alleviate the scarcity of labeled designs, a pre-training method is proposed based on a suite of compiler's data flow analysis tasks. Experimental results on two benchmark datasets show the superiority of ProgSG over baseline methods that either only consider one modality or combine the two without utilizing the alignment information.

Abstract: Uncertainty-quantification methods are applied to estimate the confidence of deep-neural-networks classifiers over their predictions. However, most widely used methods are known to be overconfident. We address this problem by developing an algorithm that exploits the latent-space representation of data points fed into the network, to assess the accuracy of their prediction. Using the latent-space representation generated by the fraction of training set that the network classifies correctly, we build a statistical model that is able to capture the likelihood of a given prediction. We show on a synthetic dataset that commonly used methods are mostly overconfident. Overconfidence occurs also for predictions made on data points that are outside the distribution that generated the training data. In contrast, our method can detect such out-of-distribution data points as inaccurately predicted, thus aiding in the automatic detection of outliers.

Abstract: The difficulty of appropriately assigning credit is particularly heightened in cooperative MARL with sparse reward, due to the concurrent time and structural scales involved. Automatic subgoal generation (ASG) has recently emerged as a viable MARL approach inspired by utilizing subgoals in intrinsically motivated reinforcement learning. However, end-to-end learning of complex task planning from sparse rewards without prior knowledge, undoubtedly requires massive training samples. Moreover, the diversity-promoting nature of existing ASG methods can lead to the "over-representation" of subgoals, generating numerous spurious subgoals of limited relevance to the actual task reward and thus decreasing the sample efficiency of the algorithm. To address this problem and inspired by the disentangled representation learning, we propose a novel "disentangled" decision-making method, Semantically Aligned task decomposition in MARL (SAMA), that prompts pretrained language models with chain-of-thought that can suggest potential goals, provide suitable goal decomposition and subgoal allocation as well as self-reflection-based replanning. Additionally, SAMA incorporates language-grounded RL to train each agent's subgoal-conditioned policy. SAMA demonstrates considerable advantages in sample efficiency compared to state-of-the-art ASG methods, as evidenced by its performance on two challenging sparse-reward tasks, Overcooked and MiniRTS.

Abstract: Learning on graphs is becoming prevalent in a wide range of applications including social networks, robotics, communication, medicine, etc. These datasets belonging to entities often contain critical private information. The utilization of data for graph learning applications is hampered by the growing privacy concerns from users on data sharing. Existing privacy-preserving methods pre-process the data to extract user-side features, and only these features are used for subsequent learning. Unfortunately, these methods are vulnerable to adversarial attacks to infer private attributes. We present a novel privacy-respecting framework for distributed graph learning and graph-based machine learning. In order to perform graph learning and other downstream tasks on the server side, this framework aims to learn features as well as distances without requiring actual features while preserving the original structural properties of the raw data. The proposed framework is quite generic and highly adaptable. We demonstrate the utility of the Euclidean space, but it can be applied with any existing method of distance approximation and graph learning for the relevant spaces. Through extensive experimentation on both synthetic and real datasets, we demonstrate the efficacy of the framework in terms of comparing the results obtained without data sharing to those obtained with data sharing as a benchmark. This is, to our knowledge, the first privacy-preserving distributed graph learning framework.

Abstract: Recalibrating probabilistic classifiers is vital for enhancing the reliability and accuracy of predictive models. Despite the development of numerous recalibration algorithms, there is still a lack of a comprehensive theory that integrates calibration and sharpness (which is essential for maintaining predictive power). In this paper, we introduce the concept of minimum-risk recalibration within the framework of mean-squared-error (MSE) decomposition, offering a principled approach for evaluating and recalibrating probabilistic classifiers. Using this framework, we analyze the uniform-mass binning (UMB) recalibration method and establish a finite-sample risk upper bound of order $\tilde{O}(B/n + 1/B^2)$ where $B$ is the number of bins and $n$ is the sample size. By balancing calibration and sharpness, we further determine that the optimal number of bins for UMB scales with $n^{1/3}$, resulting in a risk bound of approximately $O(n^{-2/3})$. Additionally, we tackle the challenge of label shift by proposing a two-stage approach that adjusts the recalibration function using limited labeled data from the target domain. Our results show that transferring a calibrated classifier requires significantly fewer target samples compared to recalibrating from scratch. We validate our theoretical findings through numerical simulations, which confirm the tightness of the proposed bounds, the optimal number of bins, and the effectiveness of label shift adaptation.

Abstract: Moment restrictions and their conditional counterparts emerge in many areas of machine learning and statistics ranging from causal inference to reinforcement learning. Estimators for these tasks, generally called methods of moments, include the prominent generalized method of moments (GMM) which has recently gained attention in causal inference. GMM is a special case of the broader family of empirical likelihood estimators which are based on approximating a population distribution by means of minimizing a $\varphi$-divergence to an empirical distribution. However, the use of $\varphi$-divergences effectively limits the candidate distributions to reweightings of the data samples. We lift this long-standing limitation and provide a method of moments that goes beyond data reweighting. This is achieved by defining an empirical likelihood estimator based on maximum mean discrepancy which we term the kernel method of moments (KMM). We provide a variant of our estimator for conditional moment restrictions and show that it is asymptotically first-order optimal for such problems. Finally, we show that our method achieves competitive performance on several conditional moment restriction tasks.

Abstract: The trustworthiness of DNNs is often challenged by their vulnerability to minor adversarial perturbations, which may not only undermine prediction accuracy (robustness) but also cause biased predictions for similar inputs (individual fairness). Accurate fairness has been recently proposed to enforce a harmonic balance between accuracy and individual fairness. It induces the notion of fairness confusion matrix to categorize predictions as true fair, true biased, false fair, and false biased. This paper proposes a harmonic evaluation approach, RobustFair, for the accurate fairness of DNNs, using adversarial perturbations crafted through fairness confusion directed gradient search. By using Taylor expansions to approximate the ground truths of adversarial instances, RobustFair can particularly identify the robustness defects entangled for spurious fairness, which are often elusive in robustness evaluation, and missing in individual fairness evaluation. RobustFair can boost robustness and individual fairness evaluations by identifying robustness or fairness defects simultaneously. Empirical case studies on fairness benchmark datasets show that, compared with the state-of-the-art white-box robustness and individual fairness testing approaches, RobustFair detects significantly 1.77-11.87 times adversarial perturbations, yielding 1.83-13.12 times biased and 1.53-8.22 times false instances. The adversarial instances can then be effectively exploited to improve the accurate fairness (and hence accuracy and individual fairness) of the original deep neural network through retraining. The empirical case studies further show that the adversarial instances identified by RobustFair outperform those identified by the other testing approaches, in promoting 21% accurate fairness and 19% individual fairness on multiple sensitive attributes, without losing accuracy at all or even promoting it by up to 4%.

Abstract: Generative modeling has recently undergone remarkable advancements, primarily propelled by the transformative implications of Diffusion Probabilistic Models (DPMs). The impressive capability of these models, however, often entails significant computational overhead during both training and inference. To tackle this challenge, we present Diff-Pruning, an efficient compression method tailored for learning lightweight diffusion models from pre-existing ones, without the need for extensive re-training. The essence of Diff-Pruning is encapsulated in a Taylor expansion over pruned timesteps, a process that disregards non-contributory diffusion steps and ensembles informative gradients to identify important weights. Our empirical assessment, undertaken across four diverse datasets highlights two primary benefits of our proposed method: 1) Efficiency: it enables approximately a 50% reduction in FLOPs at a mere 10% to 20% of the original training expenditure; 2) Consistency: the pruned diffusion models inherently preserve generative behavior congruent with their pre-trained progenitors. Code is available at \url{https://github.com/VainF/Diff-Pruning}.

Abstract: Reducing the number of bits needed to encode the weights and activations of neural networks is highly desirable as it speeds up their training and inference time while reducing memory consumption. For these reasons, research in this area has attracted significant attention toward developing neural networks that leverage lower-precision computing, such as mixed-precision training. Interestingly, none of the existing approaches has investigated pure 16-bit floating-point settings. In this paper, we shed light on the overlooked efficiency of pure 16-bit floating-point neural networks. As such, we provide a comprehensive theoretical analysis to investigate the factors contributing to the differences observed between 16-bit and 32-bit models. We formalize the concepts of floating-point error and tolerance, enabling us to quantitatively explain the conditions under which a 16-bit model can closely approximate the results of its 32-bit counterpart. This theoretical exploration offers perspective that is distinct from the literature which attributes the success of low-precision neural networks to its regularization effect. This in-depth analysis is supported by an extensive series of experiments. Our findings demonstrate that pure 16-bit floating-point neural networks can achieve similar or even better performance than their mixed-precision and 32-bit counterparts. We believe the results presented in this paper will have significant implications for machine learning practitioners, offering an opportunity to reconsider using pure 16-bit networks in various applications.

Abstract: This paper proposes an actor-critic algorithm for controlling the temperature of a battery pack using a cooling fluid. This is modeled by a coupled 1D partial differential equation (PDE) with a controlled advection term that determines the speed of the cooling fluid. The Hamilton-Jacobi-Bellman (HJB) equation is a PDE that evaluates the optimality of the value function and determines an optimal controller. We propose an algorithm that treats the value network as a Physics-Informed Neural Network (PINN) to solve for the continuous-time HJB equation rather than a discrete-time Bellman optimality equation, and we derive an optimal controller for the environment that we exploit to achieve optimal control. Our experiments show that a hybrid-policy method that updates the value network using the HJB equation and updates the policy network identically to PPO achieves the best results in the control of this PDE system.

Abstract: Sparse Neural Networks (SNNs) can potentially demonstrate similar performance to their dense counterparts while saving significant energy and memory at inference. However, the accuracy drop incurred by SNNs, especially at high pruning ratios, can be an issue in critical deployment conditions. While recent works mitigate this issue through sophisticated pruning techniques, we shift our focus to an overlooked factor: hyperparameters and activation functions. Our analyses have shown that the accuracy drop can additionally be attributed to (i) Using ReLU as the default choice for activation functions unanimously, and (ii) Fine-tuning SNNs with the same hyperparameters as dense counterparts. Thus, we focus on learning a novel way to tune activation functions for sparse networks and combining these with a separate hyperparameter optimization (HPO) regime for sparse networks. By conducting experiments on popular DNN models (LeNet-5, VGG-16, ResNet-18, and EfficientNet-B0) trained on MNIST, CIFAR-10, and ImageNet-16 datasets, we show that the novel combination of these two approaches, dubbed Sparse Activation Function Search, short: SAFS, results in up to 15.53%, 8.88%, and 6.33% absolute improvement in the accuracy for LeNet-5, VGG-16, and ResNet-18 over the default training protocols, especially at high pruning ratios. Our code can be found at https://github.com/automl/SAFS

Abstract: The development of Policy Iteration (PI) has inspired many recent algorithms for Reinforcement Learning (RL), including several policy gradient methods, that gained both theoretical soundness and empirical success on a variety of tasks. The theory of PI is rich in the context of centralized learning, but its study is still in the infant stage under the federated setting. This paper explores the federated version of Approximate PI (API) and derives its error bound, taking into account the approximation error introduced by environment heterogeneity. We theoretically prove that a proper client selection scheme can reduce this error bound. Based on the theoretical result, we propose a client selection algorithm to alleviate the additional approximation error caused by environment heterogeneity. Experiment results show that the proposed algorithm outperforms other biased and unbiased client selection methods on the federated mountain car problem by effectively selecting clients with a lower level of heterogeneity from the population distribution.

Abstract: Generative models trained with Differential Privacy (DP) are increasingly used to produce synthetic data while reducing privacy risks. Navigating their specific privacy-utility tradeoffs makes it challenging to determine which models would work best for specific settings/tasks. In this paper, we fill this gap in the context of tabular data by analyzing how DP generative models distribute privacy budgets across rows and columns, arguably the main source of utility degradation. We examine the main factors contributing to how privacy budgets are spent, including underlying modeling techniques, DP mechanisms, and data dimensionality. Our extensive evaluation of both graphical and deep generative models sheds light on the distinctive features that render them suitable for different settings and tasks. We show that graphical models distribute the privacy budget horizontally and thus cannot handle relatively wide datasets while the performance on the task they were optimized for monotonically increases with more data. Deep generative models spend their budget per iteration, so their behavior is less predictable with varying dataset dimensions but could perform better if trained on more features. Also, low levels of privacy ($\epsilon\geq100$) could help some models generalize, achieving better results than without applying DP.

Abstract: Lifelong learning agents aim to learn multiple tasks sequentially over a lifetime. This involves the ability to exploit previous knowledge when learning new tasks and to avoid forgetting. Modulating masks, a specific type of parameter isolation approach, have recently shown promise in both supervised and reinforcement learning. While lifelong learning algorithms have been investigated mainly within a single-agent approach, a question remains on how multiple agents can share lifelong learning knowledge with each other. We show that the parameter isolation mechanism used by modulating masks is particularly suitable for exchanging knowledge among agents in a distributed and decentralized system of lifelong learners. The key idea is that the isolation of specific task knowledge to specific masks allows agents to transfer only specific knowledge on-demand, resulting in robust and effective distributed lifelong learning. We assume fully distributed and asynchronous scenarios with dynamic agent numbers and connectivity. An on-demand communication protocol ensures agents query their peers for specific masks to be transferred and integrated into their policies when facing each task. Experiments indicate that on-demand mask communication is an effective way to implement distributed lifelong reinforcement learning and provides a lifelong learning benefit with respect to distributed RL baselines such as DD-PPO, IMPALA, and PPO+EWC. The system is particularly robust to connection drops and demonstrates rapid learning due to knowledge exchange.

Abstract: Policy gradient algorithms are an important family of deep reinforcement learning techniques. Many past research endeavors focused on using the first-order policy gradient information to train policy networks. Different from these works, we conduct research in this paper driven by the believe that properly utilizing and controlling Hessian information associated with the policy gradient can noticeably improve the performance of policy gradient algorithms. One key Hessian information that attracted our attention is the Hessian trace, which gives the divergence of the policy gradient vector field in the Euclidean policy parametric space. We set the goal to generalize this Euclidean policy parametric space into a general Riemmanian manifold by introducing a metric tensor field $g_ab$ in the parametric space. This is achieved through newly developed mathematical tools, deep learning algorithms, and metric tensor deep neural networks (DNNs). Armed with these technical developments, we propose a new policy gradient algorithm that learns to minimize the absolute divergence in the Riemannian manifold as an important regularization mechanism, allowing the Riemannian manifold to smoothen its policy gradient vector field. The newly developed algorithm is experimentally studied on several benchmark reinforcement learning problems. Our experiments clearly show that the new metric tensor regularized algorithm can significantly outperform its counterpart that does not use our regularization technique. Additional experimental analysis further suggests that the trained metric tensor DNN and the corresponding metric tensor $g_{ab}$ can effectively reduce the absolute divergence towards zero in the Riemannian manifold.

Abstract: Reweighted wake-sleep (RWS) is a machine learning method for performing Bayesian inference in a very general class of models. RWS draws $K$ samples from an underlying approximate posterior, then uses importance weighting to provide a better estimate of the true posterior. RWS then updates its approximate posterior towards the importance-weighted estimate of the true posterior. However, recent work [Chattergee and Diaconis, 2018] indicates that the number of samples required for effective importance weighting is exponential in the number of latent variables. Attaining such a large number of importance samples is intractable in all but the smallest models. Here, we develop massively parallel RWS, which circumvents this issue by drawing $K$ samples of all $n$ latent variables, and individually reasoning about all $K^n$ possible combinations of samples. While reasoning about $K^n$ combinations might seem intractable, the required computations can be performed in polynomial time by exploiting conditional independencies in the generative model. We show considerable improvements over standard "global" RWS, which draws $K$ samples from the full joint.

Abstract: While policy optimization algorithms have played an important role in recent empirical success of Reinforcement Learning (RL), the existing theoretical understanding of policy optimization remains rather limited -- they are either restricted to tabular MDPs or suffer from highly suboptimal sample complexity, especial in online RL where exploration is necessary. This paper proposes a simple efficient policy optimization framework -- Optimistic NPG for online RL. Optimistic NPG can be viewed as simply combining of the classic natural policy gradient (NPG) algorithm [Kakade, 2001] with optimistic policy evaluation subroutines to encourage exploration. For $d$-dimensional linear MDPs, Optimistic NPG is computationally efficient, and learns an $\varepsilon$-optimal policy within $\tilde{O}(d^2/\varepsilon^3)$ samples, which is the first computationally efficient algorithm whose sample complexity has the optimal dimension dependence $\tilde{\Theta}(d^2)$. It also improves over state-of-the-art results of policy optimization algorithms [Zanette et al., 2021] by a factor of $d$. For general function approximation that subsumes linear MDPs, Optimistic NPG, to our best knowledge, is also the first policy optimization algorithm that achieves the polynomial sample complexity for learning near-optimal policies.

Abstract: We address the problem of denoising data from a Gaussian mixture using a two-layer non-linear autoencoder with tied weights and a skip connection. We consider the high-dimensional limit where the number of training samples and the input dimension jointly tend to infinity while the number of hidden units remains bounded. We provide closed-form expressions for the denoising mean-squared test error. Building on this result, we quantitatively characterize the advantage of the considered architecture over the autoencoder without the skip connection that relates closely to principal component analysis. We further show that our results accurately capture the learning curves on a range of real data sets.

Abstract: This paper presents a general methodology for deriving information-theoretic generalization bounds for learning algorithms. The main technical tool is a probabilistic decorrelation lemma based on a change of measure and a relaxation of Young's inequality in $L_{\psi_p}$ Orlicz spaces. Using the decorrelation lemma in combination with other techniques, such as symmetrization, couplings, and chaining in the space of probability measures, we obtain new upper bounds on the generalization error, both in expectation and in high probability, and recover as special cases many of the existing generalization bounds, including the ones based on mutual information, conditional mutual information, stochastic chaining, and PAC-Bayes inequalities. In addition, the Fernique-Talagrand upper bound on the expected supremum of a subgaussian process emerges as a special case.

Abstract: Minimizing the difference of two submodular (DS) functions is a problem that naturally occurs in various machine learning problems. Although it is well known that a DS problem can be equivalently formulated as the minimization of the difference of two convex (DC) functions, existing algorithms do not fully exploit this connection. A classical algorithm for DC problems is called the DC algorithm (DCA). We introduce variants of DCA and its complete form (CDCA) that we apply to the DC program corresponding to DS minimization. We extend existing convergence properties of DCA, and connect them to convergence properties on the DS problem. Our results on DCA match the theoretical guarantees satisfied by existing DS algorithms, while providing a more complete characterization of convergence properties. In the case of CDCA, we obtain a stronger local minimality guarantee. Our numerical results show that our proposed algorithms outperform existing baselines on two applications: speech corpus selection and feature selection.

Abstract: Typical generative diffusion models rely on a Gaussian diffusion process for training the backward transformations, which can then be used to generate samples from Gaussian noise. However, real world data often takes place in discrete-state spaces, including many scientific applications. Here, we develop a theoretical formulation for arbitrary discrete-state Markov processes in the forward diffusion process using exact (as opposed to variational) analysis. We relate the theory to the existing continuous-state Gaussian diffusion as well as other approaches to discrete diffusion, and identify the corresponding reverse-time stochastic process and score function in the continuous-time setting, and the reverse-time mapping in the discrete-time setting. As an example of this framework, we introduce ``Blackout Diffusion'', which learns to produce samples from an empty image instead of from noise. Numerical experiments on the CIFAR-10, Binarized MNIST, and CelebA datasets confirm the feasibility of our approach. Generalizing from specific (Gaussian) forward processes to discrete-state processes without a variational approximation sheds light on how to interpret diffusion models, which we discuss.

Abstract: Foundation models (e.g., CLIP or DINOv2) have shown their impressive learning and transferring capabilities on a wide range of visual tasks, by training on a large corpus of data and adapting to specific downstream tasks. It is, however, interesting that foundation models have not been fully explored for universal domain adaptation (UniDA), which is to learn models using labeled data in a source domain and unlabeled data in a target one, such that the learned models can successfully adapt to the target data. In this paper, we make comprehensive empirical studies of state-of-the-art UniDA methods using foundation models. We first demonstrate that, while foundation models greatly improve the performance of the baseline methods that train the models on the source data alone, existing UniDA methods generally fail to improve over the baseline. This suggests that new research efforts are very necessary for UniDA using foundation models. To this end, we propose a very simple method of target data distillation on the CLIP model, and achieves consistent improvement over the baseline across all the UniDA benchmarks. Our studies are under a newly proposed evaluation metric of universal classification rate (UCR), which is threshold- and ratio-free and addresses the threshold-sensitive issue encountered when using the existing H-score metric.

Abstract: We introduce Clifford Group Equivariant Neural Networks: a novel approach for constructing $\mathrm{E}(n)$-equivariant networks. We identify and study the $\textit{Clifford group}$, a subgroup inside the Clifford algebra, whose definition we slightly adjust to achieve several favorable properties. Primarily, the group's action forms an orthogonal automorphism that extends beyond the typical vector space to the entire Clifford algebra while respecting the multivector grading. This leads to several non-equivalent subrepresentations corresponding to the multivector decomposition. Furthermore, we prove that the action respects not just the vector space structure of the Clifford algebra but also its multiplicative structure, i.e., the geometric product. These findings imply that every polynomial in multivectors, including their grade projections, constitutes an equivariant map with respect to the Clifford group, allowing us to parameterize equivariant neural network layers. Notable advantages are that these layers operate directly on a vector basis and elegantly generalize to any dimension. We demonstrate, notably from a single core implementation, state-of-the-art performance on several distinct tasks, including a three-dimensional $n$-body experiment, a four-dimensional Lorentz-equivariant high-energy physics experiment, and a five-dimensional convex hull experiment.

Abstract: The rise of machine learning (ML) systems has exacerbated their carbon footprint due to increased capabilities and model sizes. However, there is scarce knowledge on how the carbon footprint of ML models is actually measured, reported, and evaluated. In light of this, the paper aims to analyze the measurement of the carbon footprint of 1,417 ML models and associated datasets on Hugging Face, which is the most popular repository for pretrained ML models. The goal is to provide insights and recommendations on how to report and optimize the carbon efficiency of ML models. The study includes the first repository mining study on the Hugging Face Hub API on carbon emissions. This study seeks to answer two research questions: (1) how do ML model creators measure and report carbon emissions on Hugging Face Hub?, and (2) what aspects impact the carbon emissions of training ML models? The study yielded several key findings. These include a decreasing proportion of carbon emissions-reporting models, a slight decrease in reported carbon footprint on Hugging Face over the past 2 years, and a continued dominance of NLP as the main application domain. Furthermore, the study uncovers correlations between carbon emissions and various attributes such as model size, dataset size, and ML application domains. These results highlight the need for software measurements to improve energy reporting practices and promote carbon-efficient model development within the Hugging Face community. In response to this issue, two classifications are proposed: one for categorizing models based on their carbon emission reporting practices and another for their carbon efficiency. The aim of these classification proposals is to foster transparency and sustainable model development within the ML community.

Abstract: We derive upper bounds for random design linear regression with dependent ($\beta$-mixing) data absent any realizability assumptions. In contrast to the strictly realizable martingale noise regime, no sharp instance-optimal non-asymptotics are available in the literature. Up to constant factors, our analysis correctly recovers the variance term predicted by the Central Limit Theorem -- the noise level of the problem -- and thus exhibits graceful degradation as we introduce misspecification. Past a burn-in, our result is sharp in the moderate deviations regime, and in particular does not inflate the leading order term by mixing time factors.

Abstract: We present evidence that language models can learn meaning despite being trained only to perform next token prediction on text, specifically a corpus of programs. Each program is preceded by a specification in the form of (textual) input-output examples. Working with programs enables us to precisely define concepts relevant to meaning in language (e.g., correctness and semantics), making program synthesis well-suited as an intermediate testbed for characterizing the presence (or absence) of meaning in language models. We first train a Transformer model on the corpus of programs, then probe the trained model's hidden states as it completes a program given a specification. Despite providing no inductive bias toward learning the semantics of the language, we find that a linear probe is able to extract abstractions of both current and future program states from the model states. Moreover, there is a strong, statistically significant correlation between the accuracy of the probe and the model's ability to generate a program that implements the specification. To evaluate whether the semantics are represented in the model states rather than learned by the probe, we design a novel experimental procedure that intervenes on the semantics of the language while preserving the lexicon and syntax. We also demonstrate that the model learns to generate correct programs that are, on average, shorter than those in the training set, which is evidence that language model outputs may differ from the training distribution in semantically meaningful ways. In summary, this paper does not propose any new techniques for training language models, but develops an experimental framework for and provides insights into the acquisition and representation of (formal) meaning in language models.