ICML 2024 Papers

This work studies sparse adversarial perturbations bounded by $l_0$ norm. We propose a white-box PGD-like attack method named sparse-PGD to effectively and efficiently generate such perturbations. Furthermore, we combine sparse-PGD with a black-box attack to comprehensively and more reliably evaluate the models' robustness against $l_0$ bounded adversarial perturbations. Moreover, the efficiency of sparse-PGD enables us to conduct adversarial training to build robust models against sparse perturbations. Extensive experiments demonstrate that our proposed attack algorithm exhibits strong performance in different scenarios. More importantly, compared with other robust models, our adversarially trained model demonstrates state-of-the-art robustness against various sparse attacks. Codes are available at https://github.com/CityU-MLO/sPGD.

The partially observable generalized linear model (POGLM) is a powerful tool for understanding neural connectivities under the assumption of existing hidden neurons. With spike trains only recorded from visible neurons, existing works use variational inference to learn POGLM meanwhile presenting the difficulty of learning this latent variable model. There are two main issues: (1) the sampled Poisson hidden spike count hinders the use of the pathwise gradient estimator in VI; and (2) the existing design of the variational model is neither expressive nor time-efficient, which further affects the performance. For (1), we propose a new differentiable POGLM, which enables the pathwise gradient estimator, better than the score function gradient estimator used in existing works. For (2), we propose the forward-backward message-passing sampling scheme for the variational model. Comprehensive experiments show that our differentiable POGLMs with our forward-backward message passing produce a better performance on one synthetic and two real-world datasets. Furthermore, our new method yields more interpretable parameters, underscoring its significance in neuroscience.

*Inverse reinforcement learning* (IRL) aims to recover the reward function of an *expert* agent from demonstrations of behavior. It is well-known that the IRL problem is fundamentally ill-posed, i.e., many reward functions can explain the demonstrations. For this reason, IRL has been recently reframed in terms of estimating the *feasible reward set* (Metelli et al., 2021), thus, postponing the selection of a single reward. However, so far, the available formulations and algorithmic solutions have been proposed and analyzed mainly for the *online* setting, where the learner can interact with the environment and query the expert at will. This is clearly unrealistic in most practical applications, where the availability of an *offline* dataset is a much more common scenario. In this paper, we introduce a novel notion of feasible reward set capturing the opportunities and limitations of the offline setting and we analyze the complexity of its estimation. This requires the introduction an original learning framework that copes with the intrinsic difficulty of the setting, for which data coverage is not under control. Then, we propose two computationally and statistically efficient algorithms, IRLO and PIRLO, for addressing the problem. In particular, the latter adopts a specific form of *pessimism* to enforce the novel, desirable property of *inclusion monotonicity* of the delivered feasible set. With this work, we aim to provide a panorama of the challenges of the offline IRL problem and how they can be fruitfully addressed.

The advancement toward deeper graph neural networks is currently obscured by two inherent issues in message passing, *oversmoothing* and *oversquashing*. We identify the root cause of these issues as information loss due to *heterophily mixing* in aggregation, where messages of diverse category semantics are mixed. We propose a novel multi-track graph convolutional network to address oversmoothing and oversquashing effectively. Our basic idea is intuitive: if messages are separated and independently propagated according to their category semantics, heterophilic mixing can be prevented. Consequently, we present a novel multi-track message passing scheme capable of preventing heterophilic mixing, enhancing long-distance information flow, and improving separation condition. Empirical validations show that our model achieved state-of-the-art performance on several graph datasets and effectively tackled oversmoothing and oversquashing, setting a new benchmark of $86.4$% accuracy on Cora.

Bayesian flow networks (BFNs) iteratively refine the parameters, instead of the samples in diffusion models (DMs), of distributions at various noise levels through Bayesian inference. Owing to its differentiable nature, BFNs are promising in modeling both continuous and discrete data, while simultaneously maintaining fast sampling capabilities. This paper aims to understand and enhance BFNs by connecting them with DMs through stochastic differential equations (SDEs). We identify the linear SDEs corresponding to the noise-addition processes in BFNs, demonstrate that BFN's regression losses are aligned with denoise score matching, and validate the sampler in BFN as a first-order solver for the respective reverse-time SDE. Based on these findings and existing recipes of fast sampling in DMs, we propose specialized solvers for BFNs that markedly surpass the original BFN sampler in terms of sample quality with a limited number of function evaluations (e.g., 10) on both image and text datasets. Notably, our best sampler achieves an increase in speed of $5\sim20$ times for free.

Relieving the reliance of neural network training on a global back-propagation (BP) has emerged as a notable research topic due to the biological implausibility and huge memory consumption caused by BP. Among the existing solutions, local learning optimizes gradient-isolated modules of a neural network with local errors and has been proved to be effective even on large-scale datasets. However, the reconciliation among local errors has never been investigated. In this paper, we first theoretically study non-greedy layer-wise training and show that the convergence cannot be assured when the local gradient in a module w.r.t. its input is not reconciled with the local gradient in the previous module w.r.t. its output. Inspired by the theoretical result, we further propose a local training strategy that successively regularizes the gradient reconciliation between neighboring modules without breaking gradient isolation or introducing any learnable parameters. Our method can be integrated into both local-BP and BP-free settings. In experiments, we achieve significant performance improvements compared to previous methods. Particularly, our method for CNN and Transformer architectures on ImageNet is able to attain a competitive performance with global BP, saving more than 40% memory consumption.

Autoencoders are a prominent model in many empirical branches of machine learning and lossy data compression. However, basic theoretical questions remain unanswered even in a shallow two-layer setting. In particular, to what degree does a shallow autoencoder capture the structure of the underlying data distribution? For the prototypical case of the 1-bit compression of *sparse* Gaussian data, we prove that gradient descent converges to a solution that completely disregards the sparse structure of the input. Namely, the performance of the algorithm is the same as if it was compressing a Gaussian source -- with no sparsity. For general data distributions, we give evidence of a phase transition phenomenon in the shape of the gradient descent minimizer, as a function of the data sparsity: below the critical sparsity level, the minimizer is a rotation taken uniformly at random (just like in the compression of non-sparse data); above the critical sparsity, the minimizer is the identity (up to a permutation). Finally, by exploiting a connection with approximate message passing algorithms, we show how to improve upon Gaussian performance for the compression of sparse data: adding a denoising function to a shallow architecture already reduces the loss provably, and a suitable multi-layer decoder leads to a further improvement. We validate our findings on image datasets, such as CIFAR-10 and MNIST.

A central aspect of machine learning research is experimentation, the process of designing and running experiments, analyzing the results, and iterating towards some positive outcome (e.g., improving accuracy). Could agents driven by powerful language models perform machine learning experimentation effectively? To answer this question, we introduce MLAgentBench, a suite of 13 tasks ranging from improving model performance on CIFAR-10 to recent research problems like BabyLM. For each task, an agent can perform actions like reading/writing files, executing code, and inspecting outputs. We then construct an agent that can perform ML experimentation based on ReAct framework. We benchmark agents based on Claude v1.0, Claude v2.1, Claude v3 Opus, GPT-4, GPT-4-turbo, Gemini-Pro, and Mixtral and find that a Claude v3 Opus agent is the best in terms of success rate. It can build compelling ML models over many tasks in MLAgentBench with 37.5% average success rate. Our agents also display highly interpretable plans and actions. However, the success rates vary considerably; they span from 100% on well-established older datasets to as low as 0% on recent Kaggle challenges created potentially after the underlying LM was trained. Finally, we identify several key challenges for LM-based agents such as long-term planning and reducing hallucination.

Online linear programming plays an important role in both revenue management and resource allocation, and recent research has focused on developing efficient first-order online learning algorithms. Despite the empirical success of first-order methods, they typically achieve regret no better than $\mathcal{O}(\sqrt{T})$, which is suboptimal compared to the $\mathcal{O}(\log T)$ result guaranteed by the state-of-the-art linear programming (LP)-based online algorithms. This paper establishes several important facts about online linear programming, which unveils the challenge for first-order online algorithms to achieve beyond $\mathcal{O}(\sqrt{T})$ regret. To address this challenge, we introduce a new algorithmic framework which decouples learning from decision-making. For the first time, we show that first-order methods can achieve regret $\mathcal{O}(T^{1/3})$ with this new framework.

Layer normalization (LN) is a ubiquitous technique in deep learning but our theoretical understanding to it remains elusive. This paper investigates a new theoretical direction for LN, regarding to its nonlinearity and representation capacity. We investigate the representation capacity of a network with layerwise composition of linear and LN transformations, referred to as LN-Net. We theoretically show that, given $m$ samples with any label assignment, an LN-Net with only 3 neurons in each layer and $O(m)$ LN layers can correctly classify them. We further show the lower bound of the VC dimension of an LN-Net. The nonlinearity of LN can be amplified by group partition, which is also theoretically demonstrated with mild assumption and empirically supported by our experiments. Based on our analyses, we consider to design neural architecture by exploiting and amplifying the nonlinearity of LN, and the effectiveness is supported by our experiments.

Existing research often posits spurious features as **easier** to learn than core features in neural network optimization, but the impact of their relative simplicity remains under-explored. Moreover, studies mainly focus on end performance rather than the learning dynamics of feature learning. In this paper, we propose a theoretical framework and an associated synthetic dataset grounded in boolean function analysis. This setup allows for fine-grained control over the relative complexity (compared to core features) and correlation strength (with respect to the label) of spurious features to study the dynamics of feature learning under spurious correlations. Our findings uncover several interesting phenomena: (1) stronger spurious correlations or simpler spurious features slow down the learning rate of the core features, (2) two distinct subnetworks are formed to learn core and spurious features separately, (3) learning phases of spurious and core features are not always separable, (4) spurious features are not forgotten even after core features are fully learned. We demonstrate that our findings justify the success of retraining the last layer to remove spurious correlation and also identifies limitations of popular debiasing algorithms that exploit early learning of spurious features. We support our empirical findings with theoretical analyses for the case of learning XOR features with a one-hidden-layer ReLU network.

Variational families with full-rank covariance approximations are known not to work well in black-box variational inference (BBVI), both empirically and theoretically. In fact, recent computational complexity results for BBVI have established that full-rank variational families scale poorly with the dimensionality of the problem compared to *e.g.* mean-field families. This is particularly critical to hierarchical Bayesian models with local variables; their dimensionality increases with the size of the datasets. Consequently, one gets an iteration complexity with an explicit $\mathcal{O}(N^2)$ dependence on the dataset size $N$. In this paper, we explore a theoretical middle ground *between* mean-field variational families and full-rank families: *structured* variational families. We rigorously prove that certain scale matrix structures can achieve a better iteration complexity of $\mathcal{O}\left(N\right)$, implying better scaling with respect to $N$. We empirically verify our theoretical results on large-scale hierarchical models.

Entropy minimization (EM) is frequently used to increase the accuracy of classification models when they're faced with new data at test time. EM is a self-supervised learning method that optimizes classifiers to assign even higher probabilities to their top predicted classes. In this paper, we analyze why EM works when adapting a model for a few steps and why it eventually fails after adapting for many steps. We show that, at first, EM causes the model to embed test images close to training images, thereby increasing model accuracy. After many steps of optimization, EM makes the model embed test images far away from the embeddings of training images, which results in a degradation of accuracy. Building upon our insights, we present a method for solving a practical problem: estimating a model's accuracy on a given arbitrary dataset without having access to its labels. Our method estimates accuracy by looking at how the embeddings of input images change as the model is optimized to minimize entropy. Experiments on 23 challenging datasets show that our method sets the SoTA with a mean absolute error of 5.75%, an improvement of 29.62% over the previous SoTA on this task. Our code is available at: https://github.com/oripress/EntropyEnigma

Given the increasing number of parameter-efficient adapters of large language models (LLMs), how can we reuse them to improve LLM performance on new tasks? We study how to best build a *library* of adapters given multi-task data and devise techniques for both *zero-shot* and *supervised* task generalization through *routing* in such library. We benchmark existing approaches to build this library and introduce model-based clustering, $\texttt{MBC}$, a method that groups tasks based on the similarity of their adapter parameters, indirectly optimizing for transfer across the multi-task dataset. In order to reuse the library, we present a novel zero-shot routing mechanism, $\texttt{Arrow}$, which enables dynamic selection of the most relevant adapters for new inputs without the need for retraining. We experiment with several LLMs, such as Phi-2 and Mistral, on a wide array of held-out tasks, verifying that MBC-based adapters and Arrow routing lead to superior generalization to new tasks. Thus, we make steps towards creating modular, adaptable LLMs that can match or outperform traditional joint training.

Dataset Distillation seeks to summarize a large dataset by generating a reduced set of synthetic samples. While there has been much success at distilling small datasets such as CIFAR-10 on smaller neural architectures, Dataset Distillation methods fail to scale to larger high-resolution datasets and architectures. In this work, we introduce **D**ataset **D**istillation with **D**omain **S**hift (**D3S**), a scalable distillation algorithm, made by reframing the dataset distillation problem as a *domain shift* one. In doing so, we derive a universal bound on the distillation loss, and provide a method for efficiently approximately optimizing it. We achieve state-of-the-art results on Tiny-ImageNet, ImageNet-1k, and ImageNet-21K over a variety of recently proposed baselines, including high cross-architecture generalization. Additionally, our ablation studies provide lessons on the importance of validation-time hyperparameters on distillation performance, motivating the need for standardization.

We propose a new Q-learning variant, called 2RA Q-learning, that addresses some weaknesses of existing Q-learning methods in a principled manner. One such weakness is an underlying estimation bias which cannot be controlled and often results in poor performance. We propose a distributionally robust estimator for the maximum expected value term, which allows us to precisely control the level of estimation bias introduced. The distributionally robust estimator admits a closed-form solution such that the proposed algorithm has a computational cost per iteration comparable to Watkins' Q-learning. For the tabular case, we show that 2RA Q-learning converges to the optimal policy and analyze its asymptotic mean-squared error. Lastly, we conduct numerical experiments for various settings, which corroborate our theoretical findings and indicate that 2RA Q-learning often performs better than existing methods.

In practice, training using federated learning can be orders of magnitude slower than standard centralized training. This severely limits the amount of experimentation and tuning that can be done, making it challenging to obtain good performance on a given task. Server-side proxy data can be used to run training simulations, for instance for hyperparameter tuning. This can greatly speed up the training pipeline by reducing the number of tuning runs to be performed overall on the true clients. However, it is challenging to ensure that these simulations accurately reflect the dynamics of the real federated training. In particular, the proxy data used for simulations often comes as a single centralized dataset without a partition into distinct clients, and partitioning this data in a naive way can lead to simulations that poorly reflect real federated training. In this paper we address the challenge of how to partition centralized data in a way that reflects the statistical heterogeneity of the true federated clients. We propose a fully federated, theoretically justified, algorithm that efficiently learns the distribution of the true clients and observe improved server-side simulations when using the inferred distribution to create simulated clients from the centralized data.