ICML 2024 论文列表

We investigate the out-of-domain generalization of random feature (RF) models and Transformers. We first prove that in the `generalization on the unseen (GOTU)' setting, where training data is fully seen in some part of the domain but testing is made on another part, and for RF models in the small feature regime, the convergence takes place to interpolators of minimal degree as in the Boolean case (Abbe et al., 2023). We then consider the sparse target regime and explain how this regime relates to the small feature regime, but with a different regularization term that can alter the picture in the non-Boolean case. We show two different outcomes for the sparse regime with q-ary data tokens: (1) if the data is embedded with roots of unities, then a min-degree interpolator is learned like in the Boolean case for RF models, (2) if the data is not embedded as such, e.g., simply as integers, then RF models and Transformers may not learn minimal degree interpolators. This shows that the Boolean setting and its roots of unities generalization are special cases where the minimal degree interpolator offers a rare characterization of how learning takes place. For more general integer and real-valued settings, a more nuanced picture remains to be fully characterized.

Federated learning (FL) enables collaborative machine learning across distributed data owners, but data heterogeneity poses a challenge for model calibration. While prior work focused on improving accuracy for non-iid data, calibration remains under-explored. This study reveals existing FL aggregation approaches lead to sub-optimal calibration, and theoretical analysis shows despite constraining variance in clients’ label distributions, global calibration error is still asymptotically lower bounded. To address this, we propose a novel Federated Calibration (FedCal) approach, emphasizing both local and global calibration. It leverages client-specific scalers for local calibration to effectively correct output misalignment without sacrificing prediction accuracy. These scalers are then aggregated via weight averaging to generate a global scaler, minimizing the global calibration error. Extensive experiments demonstrate that FedCal significantly outperforms the best-performing baseline, reducing global calibration error by 47.66% on average.

Novel reinforcement learning algorithms, or improvements on existing ones, are commonly justified by evaluating their performance on benchmark environments and are compared to an ever-changing set of standard algorithms. However, despite numerous calls for improvements, experimental practices continue to produce misleading or unsupported claims. One reason for the ongoing substandard practices is that conducting rigorous benchmarking experiments requires substantial computational time. This work investigates the sources of increased computation costs in rigorous experiment designs. We show that conducting rigorous performance benchmarks will likely have computational costs that are often prohibitive. As a result, we argue for using an additional experimentation paradigm to overcome the limitations of benchmarking.

Shuffling gradient methods are widely used in modern machine learning tasks and include three popular implementations: Random Reshuffle (RR), Shuffle Once (SO), and Incremental Gradient (IG). Compared to the empirical success, the theoretical guarantee of shuffling gradient methods was not well-understood for a long time. Until recently, the convergence rates had just been established for the average iterate for convex functions and the last iterate for strongly convex problems (using squared distance as the metric). However, when using the function value gap as the convergence criterion, existing theories cannot interpret the good performance of the last iterate in different settings (e.g., constrained optimization). To bridge this gap between practice and theory, we prove the first last-iterate convergence rates for shuffling gradient methods with respect to the objective value even without strong convexity. Our new results either (nearly) match the existing last-iterate lower bounds or are as fast as the previous best upper bounds for the average iterate.

Bilevel optimization has been recently applied to many machine learning tasks. However, their applications have been restricted to the supervised learning setting, where static objective functions with benign structures are considered. But bilevel problems such as incentive design, inverse reinforcement learning (RL), and RL from human feedback (RLHF) are often modeled as dynamic objective functions that go beyond the simple static objective structures, which pose significant challenges of using existing bilevel solutions. To tackle this new class of bilevel problems, we introduce the first principled algorithmic framework for solving bilevel RL problems through the lens of penalty formulation. We provide theoretical studies of the problem landscape and its penalty-based (policy) gradient algorithms. We demonstrate the effectiveness of our algorithms via simulations in the Stackelberg game and RLHF.

We examine both the manifold hypothesis (Bengio et al., 2013) and the union of manifold hypothesis (Brown et al., 2023), and argue that, in contrast to these hypotheses, the local intrinsic dimension varies from point to point even in the same connected component. We propose an alternative CW complex hypothesis that image data is distributed in ``manifolds with skeletons". We support the hypothesis through visualization of distributions of image data of random geometric objects, as well as by introducing and testing a criterion on natural image datasets. One motivation of our work is to explain why diffusion models have difficulty generating accurate higher dimensional details such as human hands. Under the CW complex hypothesis and with both theoretical and empirical evidences, we provide an interpretation that the mixture of higher and lower dimensional components in data obstructs diffusion models from efficient learning.

Many machine learning tasks can be formulated as a stochastic compositional optimization (SCO) problem such as reinforcement learning, AUC maximization and meta-learning, where the objective function involves a nested composition associated with an expectation. Although many studies have been devoted to studying the convergence behavior of SCO algorithms, there is little work on understanding their generalization, that is, how these learning algorithms built from training data would behave on future test examples. In this paper, we provide the stability and generalization analysis of stochastic compositional gradient descent algorithms in the framework of statistical learning theory. Firstly, we introduce a stability concept called *compositional uniform stability* and establish its quantitative relation with generalization for SCO problems. Then, we establish the compositional uniform stability results for two notable stochastic compositional gradient descent algorithms, namely SCGD and SCSC. Finally, we derive *dimension-independent* excess risk bounds for SCGD and SCSC by balancing stability results and optimization errors. To the best of our knowledge, these are the first-ever known results on stability and generalization analysis of stochastic compositional gradient descent algorithms.

We introduce Magicoder, a series of fully open-source (code, weights, and data) Large Language Models (LLMs) for code that significantly closes the gap with top code models while having no more than 7B parameters. Magicoder models are trained on 75K synthetic instruction data using **OSS-Instruct**, a novel approach to enlightening LLMs with open-source code snippets to generate diverse instruction data for code. Our main motivation is to mitigate the inherent bias of the synthetic data generated by LLMs through the wealth of open-source references for the production of more realistic and controllable data. The orthogonality of OSS-Instruct and other data generation methods like Evol-Instruct further enables us to build an enhanced MagicoderS. Both Magicoder and MagicoderS substantially outperform state-of-the-art code models with similar or even larger sizes on a wide range of coding benchmarks. Notably, MagicoderS-CL-7B based on CodeLlama even surpasses the prominent ChatGPT on HumanEval+ (66.5 vs. 65.9 in pass@1 ). Overall, OSS-Instruct opens a new direction for crafting diverse synthetic instruction data for code using abundant open-source references.

Modern advancements in large-scale machine learning would be impossible without the paradigm of data-parallel distributed computing. Since distributed computing with large-scale models imparts excessive pressure on communication channels, significant recent research has been directed toward co-designing communication compression strategies and training algorithms with the goal of reducing communication costs. While pure data parallelism allows better data scaling, it suffers from poor model scaling properties. Indeed, compute nodes are severely limited by memory constraints, preventing further increases in model size. For this reason, the latest achievements in training giant neural network models also rely on some form of model parallelism. In this work, we take a closer theoretical look at Independent Subnetwork Training (IST), which is a recently proposed and highly effective technique for solving the aforementioned problems. We identify fundamental differences between IST and alternative approaches, such as distributed methods with compressed communication, and provide a precise analysis of its optimization performance on a quadratic model.

Low-rank adaptation (LoRA) has recently gained much interest in fine-tuning foundation models. It effectively reduces the number of trainable parameters by incorporating low-rank matrices $A$ and $B$ to represent the weight change, i.e., $\Delta W=BA$. Despite LoRA's progress, it faces storage challenges when handling extensive customization adaptations or larger base models. In this work, we aim to further compress trainable parameters by enjoying the powerful expressiveness of the Fourier transform. Specifically, we introduce FourierFT, which treats $\Delta W$ as a matrix in the spatial domain and learns only a small fraction of its spectral coefficients. With the trained spectral coefficients, we implement the inverse discrete Fourier transform to recover $\Delta W$. Empirically, our FourierFT method shows comparable or better performance with fewer parameters than LoRA on various tasks, including natural language understanding, natural language generation, instruction tuning, and image classification. For example, when performing instruction tuning on the LLaMA2-7B model, FourierFT surpasses LoRA with only 0.064M trainable parameters, compared to LoRA's 33.5M. Our code is released at [this link](https://github.com/Chaos96/fourierft).

A significant amount of protein function requires binding small molecules, including enzymatic catalysis. As such, designing binding pockets for small molecules has several impactful applications ranging from drug synthesis to energy storage. Towards this goal, we first develop HarmonicFlow, an improved generative process over 3D protein-ligand binding structures based on our self-conditioned flow matching objective. FlowSite extends this flow model to jointly generate a protein pocket's discrete residue types and the molecule's binding 3D structure. We show that HarmonicFlow improves upon state-of-the-art generative processes for docking in simplicity, generality, and average sample quality in pocket-level docking. Enabled by this structure modeling, FlowSite designs binding sites substantially better than baseline approaches.

We present Clifford-Steerable Convolutional Neural Networks (CS-CNNs), a novel class of ${\operatorname{E}}(p, q)$-equivariant CNNs. CS-CNNs process multivector fields on pseudo-Euclidean spaces $\mathbb{R}^{p,q}$. They specialize, for instance, to ${\operatorname{E}}(3)$-equivariance on $\mathbb{R}^3$ and Poincaré-equivariance on Minkowski spacetime $\mathbb{R}^{1,3}$. Our approach is based on an implicit parametrization of ${\operatorname{O}}(p,q)$-steerable kernels via Clifford group equivariant neural networks. We significantly and consistently outperform baseline methods on fluid dynamics as well as relativistic electrodynamics forecasting tasks.

In modern machine learning, models can often fit training data in numerous ways, some of which perform well on unseen (test) data, while others do not. Remarkably, in such cases gradient descent frequently exhibits an implicit bias that leads to excellent performance on unseen data. This implicit bias was extensively studied in supervised learning, but is far less understood in optimal control (reinforcement learning). There, learning a controller applied to a system via gradient descent is known as policy gradient, and a question of prime importance is the extent to which a learned controller extrapolates to unseen initial states. This paper theoretically studies the implicit bias of policy gradient in terms of extrapolation to unseen initial states. Focusing on the fundamental Linear Quadratic Regulator (LQR) problem, we establish that the extent of extrapolation depends on the degree of exploration induced by the system when commencing from initial states included in training. Experiments corroborate our theory, and demonstrate its conclusions on problems beyond LQR, where systems are non-linear and controllers are neural networks. We hypothesize that real-world optimal control may be greatly improved by developing methods for informed selection of initial states to train on.

In many predictive contexts (e.g., credit lending), true outcomes are only observed for samples that were positively classified in the past. These past observations, in turn, form training datasets for classifiers that make future predictions. However, such training datasets lack information about the outcomes of samples that were (incorrectly) negatively classified in the past and can lead to erroneous classifiers. We present an approach that trains a classifier using available data and comes with a family of exploration strategies to collect outcome data about subpopulations that otherwise would have been ignored. For any exploration strategy, the approach comes with guarantees that (1) all sub-populations are explored, (2) the fraction of false positives is bounded, and (3) the trained classifier converges to a "desired" classifier. The right exploration strategy is context-dependent; it can be chosen to improve learning guarantees and encode context-specific group fairness properties. Evaluation on real-world datasets shows that this approach consistently boosts the quality of collected outcome data and improves the fraction of true positives for all groups, with only a small reduction in predictive utility.

Conventional community detection methods often categorize all nodes into clusters. However, the presumed community structure of interest may only be valid for a subset of nodes (named as `tight nodes'), while the rest of the network may consist of noninformative ``scattered nodes''. For example, a protein-protein network often contains proteins that do not belong to specific biological functional modules but are involved in more general processes, or act as bridges between different functional modules. Forcing each of these proteins into a single cluster introduces unwanted biases and obscures the underlying biological implication. To address this issue, we propose a tight community detection (TCD) method to identify tight communities excluding scattered nodes. The algorithm enjoys a strong theoretical guarantee of tight node identification accuracy and is scalable for large networks. The superiority of the proposed method is demonstrated by various synthetic and real experiments.

The performance of learning models often deteriorates when deployed in out-of-sample environments. To ensure reliable deployment, we propose a stability evaluation criterion based on distributional perturbations. Conceptually, our stability evaluation criterion is defined as the minimal perturbation required on our observed dataset to induce a prescribed deterioration in risk evaluation. In this paper, we utilize the optimal transport (OT) discrepancy with moment constraints on the (sample, density) space to quantify this perturbation. Therefore, our stability evaluation criterion can address both data corruptions and sub-population shifts—the two most common types of distribution shifts in real-world scenarios. To further realize practical benefits, we present a series of tractable convex formulations and computational methods tailored to different classes of loss functions. The key technical tool to achieve this is the strong duality theorem provided in this paper. Empirically, we validate the practical utility of our stability evaluation criterion across a host of real-world applications. These empirical studies showcase the criterion's ability not only to compare the stability of different learning models and features but also to provide valuable guidelines and strategies to further improve models.

Mechanistic Interpretability (MI) proposes a path toward fully understanding how neural networks make their predictions. Prior work demonstrates that even when trained to perform simple arithmetic, models can implement a variety of algorithms (sometimes concurrently) depending on initialization and hyperparameters. Does this mean neuron-level interpretability techniques have limited applicability? Here, we argue that high-dimensional neural networks can learn *useful* low-dimensional representations of the data they were trained on, going beyond simply making good predictions: Such representations can be understood with the MI lens and provide insights that are surprisingly faithful to human-derived domain knowledge. This indicates that such approaches to interpretability can be useful for deriving a new understanding of a problem from models trained to solve it. As a case study, we extract nuclear physics concepts by studying models trained to reproduce nuclear data.

In contrast to close-set scenarios that restore images from a predefined set of degradations, open-set image restoration aims to handle the unknown degradations that were unforeseen during the pretraining phase, which is less-touched as far as we know. This work study this challenging problem and reveal its essence as unidentified distribution shifts between the test and training data. Recently, test-time adaptation has emerged as a fundamental method to address this inherent disparities. Inspired by it, we propose a test-time degradation adaptation framework for open-set image restoration, which consists of three components, *i.e.*, i) a pre-trained and degradation-agnostic diffusion model for generating clean images, ii) a test-time degradation adapter adapts the unknown degradations based on the input image during the testing phase, and iii) the adapter-guided image restoration guides the model through the adapter to produce the corresponding clean image. Through experiments on multiple degradations, we show that our method achieves comparable even better performance than those task-specific methods. The code is available at https://github.com/XLearning-SCU/2024-ICML-TAO.

The growing use of machine learning (ML) has raised concerns that an ML model may reveal private information about an individual who has contributed to the training dataset. To prevent leakage of sensitive data, we consider using differentially- private (DP), synthetic training data instead of real training data to train an ML model. A key desirable property of synthetic data is its ability to preserve the low-order marginals of the original distribution. Our main contribution comprises novel upper and lower bounds on the excess empirical risk of linear models trained on such synthetic data, for continuous and Lipschitz loss functions. We perform extensive experimentation alongside our theoretical results.