ICML 2023 论文列表

Large transformer models powered by diverse data and model scale have dominated natural language modeling and computer vision and pushed the frontier of multiple AI areas. In reinforcement learning (RL), despite many efforts into transformer-based policies, a key limitation, however, is that current transformer-based policies cannot learn by directly combining information from multiple sub-optimal trials. In this work, we address this issue using recently proposed chain of hindsight to relabel experience, where we train a transformer on a sequence of trajectory experience ascending sorted according to their total rewards. Our method consists of relabelling target return of each trajectory to the maximum total reward among in sequence of trajectories and training an autoregressive model to predict actions conditioning on past states, actions, rewards, target returns, and task completion tokens, the resulting model, Agentic Transformer (AT), can learn to improve upon itself both at training and test time. As we show on D4RL and ExoRL benchmarks, to the best our knowledge, this is the first time that a simple transformer-based model performs competitively with both temporal-difference and imitation-learning-based approaches, even from sub-optimal data. Our Agentic Transformer also shows a promising scaling trend that bigger models consistently improve results.

It has become standard to solve NLP tasks by fine-tuning pre-trained language models (LMs), especially in low-data settings. There is minimal theoretical understanding of empirical success, e.g., why fine-tuning a model with $10^8$ or more parameters on a couple dozen training points does not result in overfitting. We investigate whether the Neural Tangent Kernel (NTK)---which originated as a model to study the gradient descent dynamics of infinitely wide networks with suitable random initialization---describes fine-tuning of pre-trained LMs. This study was inspired by the decent performance of NTK for computer vision tasks (Wei et al., 2022). We extend the NTK formalism to Adam and use Tensor Programs (Yang, 2020) to characterize conditions under which the NTK lens may describe fine-tuning updates to pre-trained language models. Extensive experiments on 14 NLP tasks validate our theory and show that formulating the downstream task as a masked word prediction problem through prompting often induces kernel-based dynamics during fine-tuning. Finally, we use this kernel view to propose an explanation for the success of parameter-efficient subspace-based fine-tuning methods.

Despite their importance for assessing reliability of predictions, uncertainty quantification (UQ) measures in machine learning models have only recently begun to be rigorously characterized. One prominent issue is the *curse of dimensionality*: it is commonly believed that the marginal likelihood should be reminiscent of cross-validation metrics and both should deteriorate with larger input dimensions. However, we prove that by tuning hyperparameters to maximize marginal likelihood (the empirical Bayes procedure), performance, as measured by the marginal likelihood, *improves monotonically* with the input dimension. On the other hand, cross-validation metrics exhibit qualitatively different behavior that is characteristic of double descent. Cold posteriors, which have recently attracted interest due to their improved performance in certain settings, appear to exacerbate these phenomena. We verify empirically that our results hold for real data, beyond our considered assumptions, and we explore consequences involving synthetic covariates.

Sequential decision making algorithms often struggle to leverage different sources of unstructured offline interaction data. Imitation learning (IL) methods based on supervised learning are robust, but require optimal demonstrations, which are hard to collect. Offline goal-conditioned reinforcement learning (RL) algorithms promise to learn from sub-optimal data, but face optimization challenges especially with high-dimensional data. To bridge the gap between IL and RL, we introduce Distance Weighted Supervised Learning or DWSL, a supervised method for learning goal-conditioned policies from offline data. DWSL models the entire distribution of time-steps between states in offline data with only supervised learning, and uses this distribution to approximate shortest path distances. To extract a policy, we weight actions by their reduction in distance estimates. Theoretically, DWSL converges to an optimal policy constrained to the data distribution, an attractive property for offline learning, without any bootstrapping. Across all datasets we test, DWSL empirically maintains behavior cloning as a lower bound while still exhibiting policy improvement. In high-dimensional image domains, DWSL surpasses the performance of both prior goal-conditioned IL and RL algorithms. Visualizations and code can be found at https://sites.google.com/view/dwsl/home.

We consider random walks on discrete state spaces, such as general undirected graphs, where the random walkers are designed to approximate a target quantity over the network topology via sampling and neighborhood exploration in the form of Markov chain Monte Carlo (MCMC) procedures. Given any Markov chain corresponding to a target probability distribution, we design a *self-repellent random walk* (SRRW) which is less likely to transition to nodes that were highly visited in the past, and more likely to transition to seldom visited nodes. For a class of SRRWs parameterized by a positive real $\alpha$, we prove that the empirical distribution of the process converges almost surely to the the target (stationary) distribution of the underlying Markov chain kernel. We then provide a central limit theorem and derive the exact form of the arising asymptotic co-variance matrix, which allows us to show that the SRRW with a stronger repellence (larger $\alpha$) always achieves a smaller asymptotic covariance, in the sense of Loewner ordering of co-variance matrices. Especially for SRRW-driven MCMC algorithms, we show that the decrease in the asymptotic sampling variance is of the order $O(1/\alpha)$, eventually going down to zero. Finally, we provide numerical simulations complimentary to our theoretical results, also empirically testing a version of SRRW with $\alpha$ increasing in time to combine the benefits of smaller asymptotic variance due to large $\alpha$, with empirically observed faster mixing properties of SRRW with smaller $\alpha$.

This paper presents a novel generative model for Computer Aided Design (CAD) that 1) represents high-level design concepts of a CAD model as a three-level hierarchical tree of neural codes, from global part arrangement down to local curve geometry; and 2) controls the generation or completion of CAD models by specifying the target design using a code tree. Concretely, a novel variant of a vector quantized VAE with "masked skip connection" extracts design variations as neural codebooks at three levels. Two-stage cascaded auto-regressive transformers learn to generate code trees from incomplete CAD models and then complete CAD models following the intended design. Extensive experiments demonstrate superior performance on conventional tasks such as unconditional generation while enabling novel interaction capabilities on conditional generation tasks. The code is available at https://github.com/samxuxiang/hnc-cad.

Adversarial attacks in reinforcement learning (RL) often assume highly-privileged access to the victim’s parameters, environment, or data. Instead, this paper proposes a novel adversarial setting called a Cheap Talk MDP in which an Adversary can merely append deterministic messages to the Victim’s observation, resulting in a minimal range of influence. The Adversary cannot occlude ground truth, influence underlying environment dynamics or reward signals, introduce non-stationarity, add stochasticity, see the Victim’s actions, or access their parameters. Additionally, we present a simple meta-learning algorithm called Adversarial Cheap Talk (ACT) to train Adversaries in this setting. We demonstrate that an Adversary trained with ACT can still significantly influence the Victim’s training and testing performance, despite the highly constrained setting. Affecting train-time performance reveals a new attack vector and provides insight into the success and failure modes of existing RL algorithms. More specifically, we show that an ACT Adversary is capable of harming performance by interfering with the learner’s function approximation, or instead helping the Victim’s performance by outputting useful features. Finally, we show that an ACT Adversary can manipulate messages during train-time to directly and arbitrarily control the Victim at test-time.

Unsupervised anomaly detection (UAD) of multivariate time series (MTS) aims to learn robust representations of normal multivariate temporal patterns. Existing UAD methods try to learn a fixed set of mappings for each MTS, entailing expensive computation and limited model adaptation. To address this pivotal issue, we propose a prototype-oriented UAD (PUAD) method under a probabilistic framework. Specifically, instead of learning the mappings for each MTS, the proposed PUAD views multiple MTSs as the distribution over a group of prototypes, which are extracted to represent a diverse set of normal patterns. To learn and regulate the prototypes, PUAD introduces a reconstruction-based unsupervised anomaly detection approach, which incorporates a prototype-oriented optimal transport method into a Transformer-powered probabilistic dynamical generative framework. Leveraging meta-learned transferable prototypes, PUAD can achieve high model adaptation capacity for new MTSs. Experiments on five public MTS datasets all verify the effectiveness of the proposed UAD method.

In large scale machine learning, *random sampling* is a popular way to approximate datasets by a small representative subset of examples. In particular, *sensitivity sampling* is an intensely studied technique which provides provable guarantees on the quality of approximation, while reducing the number of examples to the product of the *VC dimension* $d$ and the *total sensitivity* $\mathfrak{S}$ in remarkably general settings. However, guarantees going beyond this general bound of $\mathfrak{S} d$ are known in perhaps only one setting, for *$\ell_2$ subspace embeddings*, despite intense study of sensitivity sampling in prior work. In this work, we show the first bounds for sensitivity sampling for $\ell_p$ subspace embeddings for $p\neq 2$ that improve over the general $\mathfrak{S} d$ bound, achieving a bound of roughly $\mathfrak{S}^{2/p}$ for $1\leq p<2$ and $\mathfrak{S}^{2-2/p}$ for $2<p<\infty$. For $1\leq p<2$, we show that this bound is tight, in the sense that there exist matrices for which $\mathfrak{S}^{2/p}$ samples is necessary. Furthermore, our techniques yield further new results in the study of sampling algorithms, showing that the *root leverage score sampling* algorithm achieves a bound of roughly $d$ for $1\leq p<2$, and that a combination of leverage score and sensitivity sampling achieves an improved bound of roughly $d^{2/p}\mathfrak{S}^{2-4/p}$ for $2<p<\infty$. Our sensitivity sampling results yield the best known sample complexity for a wide class of structured matrices that have small $\ell_p$ sensitivity.

Robustness to adversarial attacks is typically evaluated with adversarial accuracy. While essential, this metric does not capture all aspects of robustness and in particular leaves out the question of how many perturbations can be found for each point. In this work, we introduce an alternative approach, adversarial sparsity, which quantifies how difficult it is to find a successful perturbation given both an input point and a constraint on the direction of the perturbation. We show that sparsity provides valuable insight into neural networks in multiple ways: for instance, it illustrates important differences between current state-of-the-art robust models them that accuracy analysis does not, and suggests approaches for improving their robustness. When applying broken defenses effective against weak attacks but not strong ones, sparsity can discriminate between the totally ineffective and the partially effective defenses. Finally, with sparsity we can measure increases in robustness that do not affect accuracy: we show for example that data augmentation can by itself increase adversarial robustness, without using adversarial training.

Graph Neural Networks (GNNs) are powerful models for non-Euclidean data, but their training is often accentuated by massive unnecessary computation: on the one hand, training on non-Euclidean data has relatively high computational cost due to its irregular density properties; on the other hand, the class imbalance property often associated with non-Euclidean data cannot be alleviated by the massiveness of the data, thus hindering the generalisation of the models. To address the above issues, theoretically, we start with a hypothesis about the effectiveness of using a subset of training data for GNNs, which is guaranteed by the gradient distance between the subset and the full set. Empirically, we also observe that a subset of the data can provide informative gradients for model optimization and which changes over time dynamically. We name this phenomenon dynamic data sparsity. Additionally, we find that pruned sparse contrastive models may miss valuable information, leading to a large loss value on the informative subset. Motivated by the above findings, we develop a unified data model dynamic sparsity framework called Data Decantation (DataDec) to address the above challenges. The key idea of DataDec is to identify the informative subset dynamically during the training process by applying sparse graph contrastive learning. The effectiveness of DataDec is comprehensively evaluated on graph benchmark datasets and we also verify its generalizability on image data.

Contrastive Learning has recently achieved state-of-the-art performance in a wide range of unimodal and multimodal tasks. Many contrastive learning approaches use mined hard negatives to make batches more informative during training but these approaches are inefficient as they increase epoch length proportional to the number of mined negatives and require frequent updates of nearest neighbor indices or mining from recent batches. In this work, we provide an alternative to hard negative mining, Global Contrastive Batch Sampling (GCBS), an efficient approximation to the batch assignment problem that upper bounds the gap between the global and training losses, $\mathcal{L}^{Global} - \mathcal{L}^{Train}$, in contrastive learning settings. Through experimentation we find GCBS improves state-of-the-art performance in sentence embedding and code-search tasks. Additionally, GCBS is easy to implement as it requires only a few additional lines of code, does not maintain external data structures such as nearest neighbor indices, is more computationally efficient than the most minimal hard negative mining approaches, and makes no changes to the model being trained. Code is available at https://github.com/vinayak1/GCBS.

Accurate probabilistic predictions are essential for optimal decision making. While neural network miscalibration has been studied primarily in classification, we investigate this in the less-explored domain of regression. We conduct the largest empirical study to date to assess the probabilistic calibration of neural networks. We also analyze the performance of recalibration, conformal, and regularization methods to enhance probabilistic calibration. Additionally, we introduce novel differentiable recalibration and regularization methods, uncovering new insights into their effectiveness. Our findings reveal that regularization methods offer a favorable tradeoff between calibration and sharpness. Post-hoc methods exhibit superior probabilistic calibration, which we attribute to the finite-sample coverage guarantee of conformal prediction. Furthermore, we demonstrate that quantile recalibration can be considered as a specific case of conformal prediction. Our study is fully reproducible and implemented in a common code base for fair comparisons.

This paper considers the learning of logical (Boolean) functions with focus on the generalization on the unseen (GOTU) setting, a strong case of out-of-distribution generalization. This is motivated by the fact that the rich combinatorial nature of data in certain reasoning tasks (e.g., arithmetic/logic) makes representative data sampling challenging, and learning successfully under GOTU gives a first vignette of an 'extrapolating' or 'reasoning' learner. We then study how different network architectures trained by (S)GD perform under GOTU and provide both theoretical and experimental evidence that for a class of network models including instances of Transformers, random features models, and diagonal linear networks, a min-degree-interpolator is learned on the unseen. We also provide evidence that other instances with larger learning rates or mean-field networks reach leaky min-degree solutions. These findings lead to two implications: (1) we provide an explanation to the length generalization problem (e.g., Anil et al. 2022); (2) we introduce a curriculum learning algorithm called Degree-Curriculum that learns monomials more efficiently by incrementing supports.

We study convergence lower bounds of without-replacement stochastic gradient descent (SGD) for solving smooth (strongly-)convex finite-sum minimization problems. Unlike most existing results focusing on final iterate lower bounds in terms of the number of components $n$ and the number of epochs $K$, we seek bounds for arbitrary weighted average iterates that are tight in all factors including the condition number $\kappa$. For SGD with Random Reshuffling, we present lower bounds that have tighter $\kappa$ dependencies than existing bounds. Our results are the first to perfectly close the gap between lower and upper bounds for weighted average iterates in both strongly-convex and convex cases. We also prove weighted average iterate lower bounds for arbitrary permutation-based SGD, which apply to all variants that carefully choose the best permutation. Our bounds improve the existing bounds in factors of $n$ and $\kappa$ and thereby match the upper bounds shown for a recently proposed algorithm called GraB.

While it has long been empirically observed that adversarial robustness may be at odds with standard accuracy and may have further disparate impacts on different classes, it remains an open question to what extent such observations hold and how the class imbalance plays a role within. In this paper, we attempt to understand this question of accuracy disparity by taking a closer look at linear classifiers under a Gaussian mixture model. We decompose the impact of adversarial robustness into two parts: an inherent effect that will degrade the standard accuracy on all classes due to the robustness constraint, and the other caused by the class imbalance ratio, which will increase the accuracy disparity compared to standard training. Furthermore, we also show that such effects extend beyond the Gaussian mixture model, by generalizing our data model to the general family of stable distributions. More specifically, we demonstrate that while the constraint of adversarial robustness consistently degrades the standard accuracy in the balanced class setting, the class imbalance ratio plays a fundamentally different role in accuracy disparity compared to the Gaussian case, due to the heavy tail of the stable distribution. We additionally perform experiments on both synthetic and real-world datasets to corroborate our theoretical findings. Our empirical results also suggest that the implications may extend to nonlinear models over real-world datasets. Our code is publicly available on GitHub at https://github.com/Accuracy-Disparity/AT-on-AD.

Efficient numerical solvers for partial differential equations empower science and engineering. One commonly employed numerical solver is the preconditioned conjugate gradient (PCG) algorithm, whose performance is largely affected by the preconditioner quality. However, designing high-performing preconditioner with traditional numerical methods is highly non-trivial, often requiring problem-specific knowledge and meticulous matrix operations. We present a new method that leverages learning-based approach to obtain an approximate matrix factorization to the system matrix to be used as a preconditioner in the context of PCG solvers. Our high-level intuition comes from the shared property between preconditioners and network-based PDE solvers that excels at obtaining approximate solutions at a low computational cost. Such observation motivates us to represent preconditioners as graph neural networks (GNNs). In addition, we propose a new loss function that rewrites traditional preconditioner metrics to incorporate inductive bias from PDE data distributions, enabling effective training of high-performing preconditioners. We conduct extensive experiments to demonstrate the efficacy and generalizability of our proposed approach on solving various 2D and 3D linear second-order PDEs.

Empirical neural tangent kernels (eNTKs) can provide a good understanding of a given network's representation: they are often far less expensive to compute and applicable more broadly than infinite-width NTKs. For networks with $O$ output units (e.g. an $O$-class classifier), however, the eNTK on $N$ inputs is of size $NO \times NO$, taking $\mathcal O\big( (N O)^2\big)$ memory and up to $\mathcal O\big( (N O)^3 \big)$ computation to use. Most existing applications have therefore used one of a handful of approximations yielding $N \times N$ kernel matrices, saving orders of magnitude of computation, but with limited to no justification. We prove that one such approximation, which we call "sum of logits," converges to the true eNTK at initialization. Our experiments demonstrate the quality of this approximation for various uses across a range of settings.

Mixture models are universal approximators of smooth densities but are difficult to utilize in complicated datasets due to restrictions on typically available modes and challenges with initialiations. We show that by continuously parameterizing a mixture of factor analyzers using a learned ordinary differential equation, we can improve the fit of mixture models over direct methods. Once trained, the mixture components can be extracted and the neural ODE can be discarded, leaving us with an effective, but low-resource model. We additionally explore the use of a training curriculum from an easy-to-model latent space extracted from a normalizing flow to the more complex input space and show that the smooth curriculum helps to stabilize and improve results with and without the continuous parameterization. Finally, we introduce a hierarchical version of the model to enable more flexible, robust classification and clustering, and show substantial improvements against traditional parameterizations of GMMs.