Machine Learning (cs.LG)

Abstract: This study presents a benchmark for evaluating action-constrained reinforcement learning (RL) algorithms. In action-constrained RL, each action taken by the learning system must comply with certain constraints. These constraints are crucial for ensuring the feasibility and safety of actions in real-world systems. We evaluate existing algorithms and their novel variants across multiple robotics control environments, encompassing multiple action constraint types. Our evaluation provides the first in-depth perspective of the field, revealing surprising insights, including the effectiveness of a straightforward baseline approach. The benchmark problems and associated code utilized in our experiments are made available online at github.com/omron-sinicx/action-constrained-RL-benchmark for further research and development.

Abstract: Weather forecasting is a long-standing computational challenge with direct societal and economic impacts. This task involves a large amount of continuous data collection and exhibits rich spatiotemporal dependencies over long periods, making it highly suitable for deep learning models. In this paper, we apply pre-training techniques to weather forecasting and propose W-MAE, a Weather model with Masked AutoEncoder pre-training for multi-variable weather forecasting. W-MAE is pre-trained in a self-supervised manner to reconstruct spatial correlations within meteorological variables. On the temporal scale, we fine-tune the pre-trained W-MAE to predict the future states of meteorological variables, thereby modeling the temporal dependencies present in weather data. We pre-train W-MAE using the fifth-generation ECMWF Reanalysis (ERA5) data, with samples selected every six hours and using only two years of data. Under the same training data conditions, we compare W-MAE with FourCastNet, and W-MAE outperforms FourCastNet in precipitation forecasting. In the setting where the training data is far less than that of FourCastNet, our model still performs much better in precipitation prediction (0.80 vs. 0.98). Additionally, experiments show that our model has a stable and significant advantage in short-to-medium-range forecasting (i.e., forecasting time ranges from 6 hours to one week), and the longer the prediction time, the more evident the performance advantage of W-MAE, further proving its robustness.

Abstract: With Reinforcement Learning (RL) for inventory management (IM) being a nascent field of research, approaches tend to be limited to simple, linear environments with implementations that are minor modifications of off-the-shelf RL algorithms. Scaling these simplistic environments to a real-world supply chain comes with a few challenges such as: minimizing the computational requirements of the environment, specifying agent configurations that are representative of dynamics at real world stores and warehouses, and specifying a reward framework that encourages desirable behavior across the whole supply chain. In this work, we present a system with a custom GPU-parallelized environment that consists of one warehouse and multiple stores, a novel architecture for agent-environment dynamics incorporating enhanced state and action spaces, and a shared reward specification that seeks to optimize for a large retailer's supply chain needs. Each vertex in the supply chain graph is an independent agent that, based on its own inventory, able to place replenishment orders to the vertex upstream. The warehouse agent, aside from placing orders from the supplier, has the special property of also being able to constrain replenishment to stores downstream, which results in it learning an additional allocation sub-policy. We achieve a system that outperforms standard inventory control policies such as a base-stock policy and other RL-based specifications for 1 product, and lay out a future direction of work for multiple products.

Abstract: Large-scale Dynamic Networks (LDNs) are becoming increasingly important in the Internet age, yet the dynamic nature of these networks captures the evolution of the network structure and how edge weights change over time, posing unique challenges for data analysis and modeling. A Latent Factorization of Tensors (LFT) model facilitates efficient representation learning for a LDN. But the existing LFT models are almost based on Canonical Polyadic Factorization (CPF). Therefore, this work proposes a model based on Tensor Ring (TR) decomposition for efficient representation learning for a LDN. Specifically, we incorporate the principle of single latent factor-dependent, non-negative, and multiplicative update (SLF-NMU) into the TR decomposition model, and analyze the particular bias form of TR decomposition. Experimental studies on two real LDNs demonstrate that the propose method achieves higher accuracy than existing models.

Abstract: Source localization is the inverse problem of graph information dissemination and has broad practical applications. However, the inherent intricacy and uncertainty in information dissemination pose significant challenges, and the ill-posed nature of the source localization problem further exacerbates these challenges. Recently, deep generative models, particularly diffusion models inspired by classical non-equilibrium thermodynamics, have made significant progress. While diffusion models have proven to be powerful in solving inverse problems and producing high-quality reconstructions, applying them directly to the source localization is infeasible for two reasons. Firstly, it is impossible to calculate the posterior disseminated results on a large-scale network for iterative denoising sampling, which would incur enormous computational costs. Secondly, in the existing methods for this field, the training data itself are ill-posed (many-to-one); thus simply transferring the diffusion model would only lead to local optima. To address these challenges, we propose a two-stage optimization framework, the source localization denoising diffusion model (SL-Diff). In the coarse stage, we devise the source proximity degrees as the supervised signals to generate coarse-grained source predictions. This aims to efficiently initialize the next stage, significantly reducing its convergence time and calibrating the convergence process. Furthermore, the introduction of cascade temporal information in this training method transforms the many-to-one mapping relationship into a one-to-one relationship, perfectly addressing the ill-posed problem. In the fine stage, we design a diffusion model for the graph inverse problem that can quantify the uncertainty in the dissemination. The proposed SL-Diff yields excellent prediction results within a reasonable sampling time at extensive experiments.

Abstract: Safe reinforcement learning (RL) aims to solve an optimal control problem under safety constraints. Existing $\textit{direct}$ safe RL methods use the original constraint throughout the learning process. They either lack theoretical guarantees of the policy during iteration or suffer from infeasibility problems. To address this issue, we propose an $\textit{indirect}$ safe RL method called feasible policy iteration (FPI) that iteratively uses the feasible region of the last policy to constrain the current policy. The feasible region is represented by a feasibility function called constraint decay function (CDF). The core of FPI is a region-wise policy update rule called feasible policy improvement, which maximizes the return under the constraint of the CDF inside the feasible region and minimizes the CDF outside the feasible region. This update rule is always feasible and ensures that the feasible region monotonically expands and the state-value function monotonically increases inside the feasible region. Using the feasible Bellman equation, we prove that FPI converges to the maximum feasible region and the optimal state-value function. Experiments on classic control tasks and Safety Gym show that our algorithms achieve lower constraint violations and comparable or higher performance than the baselines.

Abstract: Federated learning (FL) enables multiple parties to collaboratively train a machine learning model without sharing their data; rather, they train their own model locally and send updates to a central server for aggregation. Depending on how the data is distributed among the participants, FL can be classified into Horizontal (HFL) and Vertical (VFL). In VFL, the participants share the same set of training instances but only host a different and non-overlapping subset of the whole feature space. Whereas in HFL, each participant shares the same set of features while the training set is split into locally owned training data subsets. VFL is increasingly used in applications like financial fraud detection; nonetheless, very little work has analyzed its security. In this paper, we focus on robustness in VFL, in particular, on backdoor attacks, whereby an adversary attempts to manipulate the aggregate model during the training process to trigger misclassifications. Performing backdoor attacks in VFL is more challenging than in HFL, as the adversary i) does not have access to the labels during training and ii) cannot change the labels as she only has access to the feature embeddings. We present a first-of-its-kind clean-label backdoor attack in VFL, which consists of two phases: a label inference and a backdoor phase. We demonstrate the effectiveness of the attack on three different datasets, investigate the factors involved in its success, and discuss countermeasures to mitigate its impact.

Abstract: Most machine learning methods assume that the input data distribution is the same in the training and testing phases. However, in practice, this stationarity is usually not met and the distribution of inputs differs, leading to unexpected performance of the learned model in deployment. The issue in which the training and test data inputs follow different probability distributions while the input-output relationship remains unchanged is referred to as covariate shift. In this paper, the performance of conventional machine learning models was experimentally evaluated in the presence of covariate shift. Furthermore, a region-based evaluation was performed by decomposing the domain of probability density function of the input data to assess the classifier's performance per domain region. Distributional changes were simulated in a two-dimensional classification problem. Subsequently, a higher four-dimensional experiments were conducted. Based on the experimental analysis, the Random Forests algorithm is the most robust classifier in the two-dimensional case, showing the lowest degradation rate for accuracy and F1-score metrics, with a range between 0.1% and 2.08%. Moreover, the results reveal that in higher-dimensional experiments, the performance of the models is predominantly influenced by the complexity of the classification function, leading to degradation rates exceeding 25% in most cases. It is also concluded that the models exhibit high bias towards the region with high density in the input space domain of the training samples.

Abstract: In this paper, we extends original Neural Collapse Phenomenon by proving Generalized Neural Collapse hypothesis. We obtain Grassmannian Frame structure from the optimization and generalization of classification. This structure maximally separates features of every two classes on a sphere and does not require a larger feature dimension than the number of classes. Out of curiosity about the symmetry of Grassmannian Frame, we conduct experiments to explore if models with different Grassmannian Frames have different performance. As a result, we discover the Symmetric Generalization phenomenon. We provide a theorem to explain Symmetric Generalization of permutation. However, the question of why different directions of features can lead to such different generalization is still open for future investigation.

Abstract: In this paper, we primarily focus on understanding the data preprocessing pipeline for DNN Training in the public cloud. First, we run experiments to test the performance implications of the two major data preprocessing methods using either raw data or record files. The preliminary results show that data preprocessing is a clear bottleneck, even with the most efficient software and hardware configuration enabled by NVIDIA DALI, a high-optimized data preprocessing library. Second, we identify the potential causes, exercise a variety of optimization methods, and present their pros and cons. We hope this work will shed light on the new co-design of ``data storage, loading pipeline'' and ``training framework'' and flexible resource configurations between them so that the resources can be fully exploited and performance can be maximized.

Abstract: Graph Neural Networks (GNNs) have become a popular tool for learning on graphs, but their widespread use raises privacy concerns as graph data can contain personal or sensitive information. Differentially private GNN models have been recently proposed to preserve privacy while still allowing for effective learning over graph-structured datasets. However, achieving an ideal balance between accuracy and privacy in GNNs remains challenging due to the intrinsic structural connectivity of graphs. In this paper, we propose a new differentially private GNN called ProGAP that uses a progressive training scheme to improve such accuracy-privacy trade-offs. Combined with the aggregation perturbation technique to ensure differential privacy, ProGAP splits a GNN into a sequence of overlapping submodels that are trained progressively, expanding from the first submodel to the complete model. Specifically, each submodel is trained over the privately aggregated node embeddings learned and cached by the previous submodels, leading to an increased expressive power compared to previous approaches while limiting the incurred privacy costs. We formally prove that ProGAP ensures edge-level and node-level privacy guarantees for both training and inference stages, and evaluate its performance on benchmark graph datasets. Experimental results demonstrate that ProGAP can achieve up to 5%-10% higher accuracy than existing state-of-the-art differentially private GNNs.

Abstract: An appropriate reward function is of paramount importance in specifying a task in reinforcement learning (RL). Yet, it is known to be extremely challenging in practice to design a correct reward function for even simple tasks. Human-in-the-loop (HiL) RL allows humans to communicate complex goals to the RL agent by providing various types of feedback. However, despite achieving great empirical successes, HiL RL usually requires too much feedback from a human teacher and also suffers from insufficient theoretical understanding. In this paper, we focus on addressing this issue from a theoretical perspective, aiming to provide provably feedback-efficient algorithmic frameworks that take human-in-the-loop to specify rewards of given tasks. We provide an active-learning-based RL algorithm that first explores the environment without specifying a reward function and then asks a human teacher for only a few queries about the rewards of a task at some state-action pairs. After that, the algorithm guarantees to provide a nearly optimal policy for the task with high probability. We show that, even with the presence of random noise in the feedback, the algorithm only takes $\widetilde{O}(H{{\dim_{R}^2}})$ queries on the reward function to provide an $\epsilon$-optimal policy for any $\epsilon > 0$. Here $H$ is the horizon of the RL environment, and $\dim_{R}$ specifies the complexity of the function class representing the reward function. In contrast, standard RL algorithms require to query the reward function for at least $\Omega(\operatorname{poly}(d, 1/\epsilon))$ state-action pairs where $d$ depends on the complexity of the environmental transition.

Abstract: Federated learning (FL) is a promising paradigm that enables distributed clients to collaboratively train a shared global model while keeping the training data locally. However, the performance of FL is often limited by poor communication links and slow convergence when FL is deployed over wireless networks. Besides, due to the limited radio resources, it is crucial to select clients and control resource allocation accurately for improved FL performance. Motivated by these challenges, a joint optimization problem of client selection and resource allocation is formulated in this paper, aiming to minimize the total time consumption of each round in FL over non-orthogonal multiple access (NOMA) enabled wireless network. Specifically, based on a metric termed the age of update (AoU), we first propose a novel client selection scheme by accounting for the staleness of the received local FL models. After that, the closed-form solutions of resource allocation are obtained by monotonicity analysis and dual decomposition method. Moreover, to further improve the performance of FL, the deployment of artificial neural network (ANN) at the server is proposed to predict the local FL models of the unselected clients at each round. Finally, extensive simulation results demonstrate the superior performance of the proposed schemes.

Abstract: Learning disentangled representations is important in representation learning, aiming to learn a low dimensional representation of data where each dimension corresponds to one underlying generative factor. Due to the possibility of causal relationships between generative factors, causal disentangled representation learning has received widespread attention. In this paper, we first propose a new flows that can incorporate causal structure information into the model, called causal flows. Based on the variational autoencoders(VAE) commonly used in disentangled representation learning, we design a new model, CF-VAE, which enhances the disentanglement ability of the VAE encoder by utilizing the causal flows. By further introducing the supervision of ground-truth factors, we demonstrate the disentanglement identifiability of our model. Experimental results on both synthetic and real datasets show that CF-VAE can achieve causal disentanglement and perform intervention experiments. Moreover, CF-VAE exhibits outstanding performance on downstream tasks and has the potential to learn causal structure among factors.

Abstract: Lumped parameter methods aim to simplify the evolution of spatially-extended or continuous physical systems to that of a "lumped" element representative of the physical scales of the modeled system. For systems where the definition of a lumped element or its associated physics may be unknown, modeling tasks may be restricted to full-fidelity simulations of the physics of a system. In this work, we consider data-driven modeling tasks with limited point-wise measurements of otherwise continuous systems. We build upon the notion of the Universal Differential Equation (UDE) to construct data-driven models for reducing dynamics to that of a lumped parameter and inferring its properties. The flexibility of UDEs allow for composing various known physical priors suitable for application-specific modeling tasks, including lumped parameter methods. The motivating example for this work is the plunge and dwell stages for friction-stir welding; specifically, (i) mapping power input into the tool to a point-measurement of temperature and (ii) using this learned mapping for process control.

Abstract: A lot of recent progress has been made in ultra low-bit quantization, promising significant improvements in latency, memory footprint and energy consumption on edge devices. Quantization methods such as Learned Step Size Quantization can achieve model accuracy that is comparable to full-precision floating-point baselines even with sub-byte quantization. However, it is extremely challenging to deploy these ultra low-bit quantized models on mainstream CPU devices because commodity SIMD (Single Instruction, Multiple Data) hardware typically supports no less than 8-bit precision. To overcome this limitation, we propose DeepGEMM, a lookup table based approach for the execution of ultra low-precision convolutional neural networks on SIMD hardware. The proposed method precomputes all possible products of weights and activations, stores them in a lookup table, and efficiently accesses them at inference time to avoid costly multiply-accumulate operations. Our 2-bit implementation outperforms corresponding 8-bit integer kernels in the QNNPACK framework by up to 1.74x on x86 platforms.

Abstract: Stochastic gradient Langevin dynamics (SGLD) are a useful methodology for sampling from probability distributions. This paper provides a finite sample analysis of a passive stochastic gradient Langevin dynamics algorithm (PSGLD) designed to achieve inverse reinforcement learning. By "passive", we mean that the noisy gradients available to the PSGLD algorithm (inverse learning process) are evaluated at randomly chosen points by an external stochastic gradient algorithm (forward learner). The PSGLD algorithm thus acts as a randomized sampler which recovers the cost function being optimized by this external process. Previous work has analyzed the asymptotic performance of this passive algorithm using stochastic approximation techniques; in this work we analyze the non-asymptotic performance. Specifically, we provide finite-time bounds on the 2-Wasserstein distance between the passive algorithm and its stationary measure, from which the reconstructed cost function is obtained.

Abstract: In statistical learning theory, determining the sample complexity of realizable binary classification for VC classes was a long-standing open problem. The results of Simon and Hanneke established sharp upper bounds in this setting. However, the reliance of their argument on the uniform convergence principle limits its applicability to more general learning settings such as multiclass classification. In this paper, we address this issue by providing optimal high probability risk bounds through a framework that surpasses the limitations of uniform convergence arguments. Our framework converts the leave-one-out error of permutation invariant predictors into high probability risk bounds. As an application, by adapting the one-inclusion graph algorithm of Haussler, Littlestone, and Warmuth, we propose an algorithm that achieves an optimal PAC bound for binary classification. Specifically, our result shows that certain aggregations of one-inclusion graph algorithms are optimal, addressing a variant of a classic question posed by Warmuth. We further instantiate our framework in three settings where uniform convergence is provably suboptimal. For multiclass classification, we prove an optimal risk bound that scales with the one-inclusion hypergraph density of the class, addressing the suboptimality of the analysis of Daniely and Shalev-Shwartz. For partial hypothesis classification, we determine the optimal sample complexity bound, resolving a question posed by Alon, Hanneke, Holzman, and Moran. For realizable bounded regression with absolute loss, we derive an optimal risk bound that relies on a modified version of the scale-sensitive dimension, refining the results of Bartlett and Long. Our rates surpass standard uniform convergence-based results due to the smaller complexity measure in our risk bound.