ICML 2023 论文列表

Modern deep learning models with great expressive power can be trained to overfit the training data but still generalize well. This phenomenon is referred to as benign overfitting. Recently, a few studies have attempted to theoretically understand benign overfitting in neural networks. However, these works are either limited to neural networks with smooth activation functions or to the neural tangent kernel regime. How and when benign overfitting can occur in ReLU neural networks remains an open problem. In this work, we seek to answer this question by establishing algorithm-dependent risk bounds for learning two-layer ReLU convolutional neural networks with label-flipping noise. We show that, under mild conditions, the neural network trained by gradient descent can achieve near-zero training loss and Bayes optimal test risk. Our result also reveals a sharp transition between benign and harmful overfitting under different conditions on data distribution in terms of test risk. Experiments on synthetic data back up our theory.

We consider the problem of performing Bayesian inference in probabilistic models where observations are accompanied by uncertainty, referred to as "uncertain evidence.'' We explore how to interpret uncertain evidence, and by extension the importance of proper interpretation as it pertains to inference about latent variables. We consider a recently-proposed method "distributional evidence'' as well as revisit two older methods: Jeffrey's rule and virtual evidence. We devise guidelines on how to account for uncertain evidence and we provide new insights, particularly regarding consistency. To showcase the impact of different interpretations of the same uncertain evidence, we carry out experiments in which one interpretation is defined as "correct.'' We then compare inference results from each different interpretation illustrating the importance of careful consideration of uncertain evidence.

Forward Gradients - the idea of using directional derivatives in forward differentiation mode - have recently been shown to be utilizable for neural network training while avoiding problems generally associated with backpropagation gradient computation, such as locking and memorization requirements. The cost is the requirement to guess the step direction, which is hard in high dimensions. While current solutions rely on weighted averages over isotropic guess vector distributions, we propose to strongly bias our gradient guesses in directions that are much more promising, such as feedback obtained from small, local auxiliary networks. For a standard computer vision neural network, we conduct a rigorous study systematically covering a variety of combinations of gradient targets and gradient guesses, including those previously presented in the literature. We find that using gradients obtained from a local loss as a candidate direction drastically improves on random noise in Forward Gradient methods.

Transformer-based models have delivered impressive results on many tasks, particularly vision and language tasks. In many model training situations, conventional configurations are often adopted. For example, we usually set the base model with hidden size (i.e. model width) to be 768 and the number of transformer layers (i.e. model depth) to be 12. In this paper, we revisit these conventional configurations by studying the the relationship between transformer configuration and training objective. We show that the optimal transformer configuration is closely related to the training objective. Specifically, compared with the simple classification objective, the masked autoencoder is effective in alleviating the over-smoothing issue in deep transformer training. Based on this finding, we propose ``Bamboo'', a notion of using deeper and narrower transformer configurations, for masked autoencoder training. On ImageNet, with such a simple change in configuration, the re-designed Base-level transformer achieves 84.2% top-1 accuracy and outperforms SoTA models like MAE by $0.9\%$. On language tasks, re-designed model outperforms BERT with the default setting by 1.1 points on average, on GLUE benchmark with 8 datasets.

Machine learning (ML) holds great potential for accurately forecasting treatment outcomes over time, which could ultimately enable the adoption of more individualized treatment strategies in many practical applications. However, a significant challenge that has been largely overlooked by the ML literature on this topic is the presence of informative sampling in observational data. When instances are observed irregularly over time, sampling times are typically not random, but rather informative–depending on the instance's characteristics, past outcomes, and administered treatments. In this work, we formalize informative sampling as a covariate shift problem and show that it can prohibit accurate estimation of treatment outcomes if not properly accounted for. To overcome this challenge, we present a general framework for learning treatment outcomes in the presence of informative sampling using inverse intensity-weighting, and propose a novel method, TESAR-CDE, that instantiates this framework using Neural CDEs. Using a simulation environment based on a clinical use case, we demonstrate the effectiveness of our approach in learning under informative sampling.

Two-point zeroth order methods are important in many applications of zeroth-order optimization arising in robotics, wind farms, power systems, online optimization, and adversarial robustness to black-box attacks in deep neural networks, where the problem can be high-dimensional and/or time-varying. Furthermore, such problems may be nonconvex and contain saddle points. While existing works have shown that zeroth-order methods utilizing $\Omega(d)$ function valuations per iteration (with $d$ denoting the problem dimension) can escape saddle points efficiently, it remains an open question if zeroth-order methods based on two-point estimators can escape saddle points. In this paper, we show that by adding an appropriate isotropic perturbation at each iteration, a zeroth-order algorithm based on $2m$ (for any $1 \leq m \leq d$) function evaluations per iteration can not only find $\epsilon$-second order stationary points polynomially fast, but do so using only $\tilde{O}(\frac{d}{m\epsilon^{2}\bar{\psi}})$ function evaluations, where $\bar{\psi} \geq \tilde{\Omega}(\sqrt{\epsilon})$ is a parameter capturing the extent to which the function of interest exhibits the strict saddle property.

There is a growing concern that learned conditional generative models may output samples that are substantially similar to some copyrighted data $C$ that was in their training set. We give a formal definition of near access-freeness (NAF) and prove bounds on the probability that a model satisfying this definition outputs a sample similar to $C$, even if $C$ is included in its training set. Roughly speaking, a generative model $p$ is $k$-NAF if for every potentially copyrighted data $C$, the output of $p$ diverges by at most $k$-bits from the output of a model $q$ that did not access $C$ at all. We also give generative model learning algorithms, which efficiently modify the original generative model learning algorithm in a black box manner, that output generative models with strong bounds on the probability of sampling protected content. Furthermore, we provide promising experiments for both language (transformers) and image (diffusion) generative models, showing minimal degradation in output quality while ensuring strong protections against sampling protected content.

In the context of adversarial robustness, we make three strongly related contributions. First, we prove that while attacking ReLU classifiers is $\mathit{NP}$-hard, ensuring their robustness at training time is $\Sigma^2_P$-hard (even on a single example). This asymmetry provides a rationale for the fact that robust classifications approaches are frequently fooled in the literature. Second, we show that inference-time robustness certificates are not affected by this asymmetry, by introducing a proof-of-concept approach named Counter-Attack (CA). Indeed, CA displays a reversed asymmetry: running the defense is $\mathit{NP}$-hard, while attacking it is $\Sigma_2^P$-hard. Finally, motivated by our previous result, we argue that adversarial attacks can be used in the context of robustness certification, and provide an empirical evaluation of their effectiveness. As a byproduct of this process, we also release UG100, a benchmark dataset for adversarial attacks.

Many deep generative models are defined as a push-forward of a Gaussian measure by a continuous generator, such as Generative Adversarial Networks (GANs) or Variational Auto-Encoders (VAEs). This work explores the latent space of such deep generative models. A key issue with these models is their tendency to output samples outside of the support of the target distribution when learning disconnected distributions. We investigate the relationship between the performance of these models and the geometry of their latent space. Building on recent developments in geometric measure theory, we prove a sufficient condition for optimality in the case where the dimension of the latent space is larger than the number of modes. Through experiments on GANs, we demonstrate the validity of our theoretical results and gain new insights into the latent space geometry of these models. Additionally, we propose a truncation method that enforces a simplicial cluster structure in the latent space and improves the performance of GANs.

We study geodesically convex (g-convex) problems that can be written as a difference of Euclidean convex functions. This structure arises in key applications such as matrix scaling, M- estimators of scatter matrices, and Brascamp-Lieb inequalities. In particular, we exploit this structure to make use of the Convex-Concave Procedure (CCCP), which helps us bypass potentially expensive Riemannian operations and leads to very competitive solvers. Importantly, unlike existing theory for CCCP that ensures convergence to stationary points, we exploit the overall g-convexity structure and provide iteration complexity results for global optimality. We illustrate our results by specializing them to a few concrete optimization problems that have been previously studied in the machine learning literature. We hope our work spurs the study of mixed Euclidean-Riemannian optimization algorithms.

Harmonic functions are abundant in nature, appearing in limiting cases of Maxwell's, Navier-Stokes equations, the heat and the wave equation. Consequently, there are many applications of harmonic functions from industrial process optimisation to robotic path planning and the calculation of first exit times of random walks. Despite their ubiquity and relevance, there have been few attempts to incorporate inductive biases towards harmonic functions in machine learning contexts. In this work, we demonstrate effective means of representing harmonic functions in neural networks and extend such results also to quantum neural networks to demonstrate the generality of our approach. We benchmark our approaches against (quantum) physics-informed neural networks, where we show favourable performance.

We present a representation-driven framework for reinforcement learning. By representing policies as estimates of their expected values, we leverage techniques from contextual bandits to guide exploration and exploitation. Particularly, embedding a policy network into a linear feature space allows us to reframe the exploration-exploitation problem as a representation-exploitation problem, where good policy representations enable optimal exploration. We demonstrate the effectiveness of this framework through its application to evolutionary and policy gradient-based approaches, leading to significantly improved performance compared to traditional methods. Our framework provides a new perspective on reinforcement learning, highlighting the importance of policy representation in determining optimal exploration-exploitation strategies.

Simple regret is a natural and parameter-free performance criterion for pure exploration in multi-armed bandits yet is less popular than the probability of missing the best arm or an $\epsilon$-good arm, perhaps due to lack of easy ways to characterize it. In this paper, we make a significant progress on minimizing simple regret in both data-rich ($T\ge n$) and data-poor regime ($T \le n$) where $n$ is the number of arms and $T$ is the number of samples. At its heart is our improved instance-dependent analysis of the well-known Sequential Halving (SH) algorithm where we bound the probability of returning an arm whose mean reward is not within $\epsilon$ from the best (i.e., not $\epsilon$-good) for *any* choice of $\epsilon>0$, although $\epsilon$ is not an input to SH. Our bound not only leads to an optimal worst-case simple regret bound of $\sqrt{n/T}$ up to logarithmic factors but also essentially matches the instance-dependent lower bound for returning an $\epsilon$-good arm reported by Katz-Samuels and Jamieson (2020). For the more challenging data-poor regime, we propose Bracketing SH (BSH) that enjoys the same improvement even without sampling each arm at least once. Our empirical study shows that BSH outperforms existing methods on real-world tasks.

This paper focuses on continual meta-learning, where few-shot tasks are heterogeneous and sequentially available. Recent works use a mixture model for meta-knowledge to deal with the heterogeneity. However, these methods suffer from parameter inefficiency caused by two reasons: (1) the underlying assumption of mutual exclusiveness among mixture components hinders sharing meta-knowledge across heterogeneous tasks. (2) they only allow increasing mixture components and cannot adaptively filter out redundant components. In this paper, we propose an Adaptive Compositional Continual Meta-Learning (ACML) algorithm, which employs a compositional premise to associate a task with a subset of mixture components, allowing meta-knowledge sharing among heterogeneous tasks. Moreover, to adaptively adjust the number of mixture components, we propose a component sparsification method based on evidential theory to filter out redundant components. Experimental results show ACML outperforms strong baselines, showing the effectiveness of our compositional meta-knowledge, and confirming that ACML can adaptively learn meta-knowledge.

In cooperative multi-agent reinforcement learning, centralized training with decentralized execution (CTDE) shows great promise for a trade-off between independent Q-learning and joint action learning. However, vanilla CTDE methods assumed a fixed number of agents could hardly adapt to real-world scenarios where dynamic team compositions typically suffer from dramatically variant partial observability. Specifically, agents with extensive sight ranges are prone to be affected by trivial environmental substrates, dubbed the "distracted attention" issue; ones with limited observation can hardly sense their teammates, degrading the cooperation quality. In this paper, we propose Complementary Attention for Multi-Agent reinforcement learning (CAMA), which applies a divide-and-conquer strategy on input entities accompanied with the complementary attention of enhancement and replenishment. Concretely, to tackle the distracted attention issue, highly contributed entities' attention is enhanced by the execution-related representation extracted via action prediction with an inverse model. For better out-of-sight-range cooperation, the lowly contributed ones are compressed to brief messages with a conditional mutual information estimator. Our CAMA facilitates stable and sustainable teamwork, which is justified by the impressive results reported on the challenging StarCraftII, MPE, and Traffic Junction benchmarks.

A critical element of learning with noisy labels is noisy label detection. Notably, numerous previous works assume that no source of labels can be clean in a noisy label detection context. In this work, we relax this assumption and assume that a small subset of the training data is clean, which enables substantial noisy label detection performance gains. Specifically, we propose a novel framework that leverages clean data by framing the problem of noisy label detection with clean data as a multiple hypothesis testing problem. Moreover, we propose BHN, a simple yet effective approach for noisy label detection that integrates the Benjamini-Hochberg (BH) procedure into deep neural networks. BHN achieves $\textit{state-of-the-art}$ performance and outperforms baselines by $\textbf{28.48}$% in terms of false discovery rate (FDR) and by $\textbf{18.99}$% in terms of F1 on CIFAR-10. Extensive ablation studies further demonstrate the superiority of BHN. Our code is available at https://github.com/ChenglinYu/BHN.

A *treatment policy* defines when and what treatments are applied to affect some outcome of interest. Data-driven decision-making requires the ability to predict *what happens if a policy is changed*. Existing methods that predict how the outcome evolves under different scenarios assume that the tentative sequences of future treatments are fixed in advance, while in practice the treatments are determined stochastically by a policy and may depend, for example, on the efficiency of previous treatments. Therefore, the current methods are not applicable if the treatment policy is unknown or a counterfactual analysis is needed. To handle these limitations, we model the treatments and outcomes jointly in continuous time, by combining Gaussian processes and point processes. Our model enables the estimation of a treatment policy from observational sequences of treatments and outcomes, and it can predict the interventional and counterfactual progression of the outcome *after an intervention on the treatment policy* (in contrast with the causal effect of a single treatment). We show with real-world and semi-synthetic data on blood glucose progression that our method can answer causal queries more accurately than existing alternatives.

In Federated Learning (FL), adaptive optimization is an effective approach to addressing the statistical heterogeneity issue but cannot adapt quickly to concept drifts. In this work, we propose a novel adaptive optimizer called Flash that simultaneously addresses both statistical heterogeneity and the concept drift issues. The fundamental insight is that a concept drift can be detected based on the magnitude of parameter updates that are required to fit the global model to each participating client's local data distribution. Flash uses a two-pronged approach that synergizes client-side early-stopping training to facilitate detection of concept drifts and the server-side drift-aware adaptive optimization to effectively adjust effective learning rate. We theoretically prove that Flash matches the convergence rate of state-of-the-art adaptive optimizers and further empirically evaluate the efficacy of Flash on a variety of FL benchmarks using different concept drift settings.

We present Topological Point Cloud Clustering (TPCC), a new method to cluster points in an arbitrary point cloud based on their contribution to global topological features. TPCC synthesizes desirable features from spectral clustering and topological data analysis and is based on considering the spectral properties of a simplicial complex associated to the considered point cloud. As it is based on considering sparse eigenvector computations, TPCC is similarly easy to interpret and implement as spectral clustering. However, by focusing not just on a single matrix associated to a graph created from the point cloud data, but on a whole set of Hodge-Laplacians associated to an appropriately constructed simplicial complex, we can leverage a far richer set of topological features to characterize the data points within the point cloud and benefit from the relative robustness of topological techniques against noise. We test the performance of TPCC on both synthetic and real-world data and compare it with classical spectral clustering.