ICML 2024 Papers

Existing works in federated learning (FL) often assume either full client or uniformly distributed client participation. However, in reality, some clients may never participate in FL training (aka incomplete client participation) due to various system heterogeneity factors. A popular solution is the server-assisted federated learning (SA-FL) framework, where the server uses an auxiliary dataset. Despite empirical evidence of SA-FL's effectiveness in addressing incomplete client participation, theoretical understanding of SA-FL is lacking. Furthermore, the effects of incomplete client participation in conventional FL are poorly understood. This motivates us to rigorously investigate SA-FL. Toward this end, we first show that conventional FL is *not* PAC-learnable under incomplete client participation in the worst case. Then, we show that the PAC-learnability of FL with incomplete client participation can indeed be revived by SA-FL, which theoretically justifies the use of SA-FL for the first time. Lastly, to provide practical guidance for SA-FL training under *incomplete client participation*, we propose the SAFARI (server-assisted federated averaging) algorithm that enjoys the same linear convergence speedup guarantees as classic FL with ideal client participation assumptions, offering the first SA-FL algorithm with convergence guarantee. Extensive experiments on different datasets show SAFARI significantly improves the performance under incomplete client participation.

While Transformers have been the main architecture behind deep learning's success in language modeling, state-space models (SSMs) such as Mamba have recently been shown to match or outperform Transformers at small to medium scale. We show that these families of models are actually quite closely related, and develop a rich framework of theoretical connections between SSMs and variants of attention, connected through various decompositions of a well-studied class of structured *semiseparable matrices*. Our state space duality (SSD) framework allows us to design a new architecture (**Mamba-2**) whose core layer is an a refinement of Mamba's selective SSM that is 2-8$\times$ faster, while continuing to be competitive with Transformers on language modeling.

Graph neural networks (GNNs) have exhibited prominent performance in learning graph-structured data. Considering node classification task, based on the i.i.d assumption among node labels, the traditional supervised learning simply sums up cross-entropy losses of the independent training nodes and applies the average loss to optimize GNNs' weights. But different from other data formats, the nodes are naturally connected. It is found that the independent distribution modeling of node labels restricts GNNs' capability to generalize over the entire graph and defend adversarial attacks. In this work, we propose a new framework, termed joint-cluster supervised learning, to model the joint distribution of each node with its corresponding cluster. We learn the joint distribution of node and cluster labels conditioned on their representations, and train GNNs with the obtained joint loss. In this way, the data-label reference signals extracted from the local cluster explicitly strengthen the discrimination ability on the target node. The extensive experiments demonstrate that our joint-cluster supervised learning can effectively bolster GNNs' node classification accuracy. Furthermore, being benefited from the reference signals which may be free from spiteful interference, our learning paradigm significantly protects the node classification from being affected by the adversarial attack.

Neural networks typically generalize well when fitting the data perfectly, even though they are heavily overparameterized. Many factors have been pointed out as the reason for this phenomenon, including an implicit bias of stochastic gradient descent (SGD) and a possible simplicity bias arising from the neural network architecture. The goal of this paper is to disentangle the factors that influence generalization stemming from optimization and architectural choices by studying *random* and *SGD-optimized* networks that achieve zero training error. We experimentally show, in the low sample regime, that overparameterization in terms of increasing width is beneficial for generalization, and this benefit is due to the bias of SGD and not due to an architectural bias. In contrast, for increasing depth, overparameterization is detrimental for generalization, but random and SGD-optimized networks behave similarly, so this can be attributed to an architectural bias.

Graph Contrastive Learning (GCL) has emerged as a highly effective self-supervised approach in graph representation learning. However, prevailing GCL methods confront two primary challenges: 1) They predominantly operate under homophily assumptions, focusing on low-frequency signals in node features while neglecting heterophilic edges that connect nodes with dissimilar features. 2) Their reliance on neighborhood aggregation for inference leads to scalability challenges and hinders deployment in real-time applications. In this paper, we introduce S3GCL, an innovative framework designed to tackle these challenges. Inspired by spectral GNNs, we initially demonstrate the correlation between frequency and homophily levels. Then, we propose a novel cosine-parameterized Chebyshev polynomial as low/high-pass filters to generate biased graph views. To resolve the inference dilemma, we incorporate an MLP encoder and enhance its awareness of graph context by introducing structurally and semantically neighboring nodes as positive pairs in the spatial domain. Finally, we formulate a cross-pass GCL objective between full-pass MLP and biased-pass GNN filtered features, eliminating the need for augmentation. Extensive experiments on real-world tasks validate S3GCL proficiency in generalization to diverse homophily levels and its superior inference efficiency.

*Counterfactuals*, or modified inputs that lead to a different outcome, are an important tool for understanding the logic used by machine learning classifiers and how to change an undesirable classification. Even if a counterfactual changes a classifier's decision, however, it may not affect the true underlying class probabilities, i.e. the counterfactual may act like an adversarial attack and ``fool'' the classifier. We propose a new framework for creating modified inputs that change the true underlying probabilities in a beneficial way which we call *Trustworthy Actionable Perturbations* (TAP). This includes a novel verification procedure to ensure that TAP change the true class probabilities instead of acting adversarially. Our framework also includes new cost, reward, and goal definitions that are better suited to effectuating change in the real world. We present PAC-learnability results for our verification procedure and theoretically analyze our new method for measuring reward. We also develop a methodology for creating TAP and compare our results to those achieved by previous counterfactual methods.

We study how to best spend a budget of noisy labels to compare the accuracy of two binary classifiers. It’s common practice to collect and aggregate multiple noisy labels for a given data point into a less noisy label via a majority vote. We prove a theorem that runs counter to conventional wisdom. If the goal is to identify the better of two classifiers, we show it’s best to spend the budget on collecting a single label for more samples. Our result follows from a non-trivial application of Cramér’s theorem, a staple in the theory of large deviations. We discuss the implications of our work for the design of machine learning benchmarks, where they overturn some time-honored recommendations. In addition, our results provide sample size bounds superior to what follows from Hoeffding’s bound.

Binarized image super-resolution (SR) has attracted much research attention due to its potential to drastically reduce parameters and operations. However, most binary SR works binarize network weights directly, which hinders high-frequency information extraction. Furthermore, as a pixel-wise reconstruction task, binarization often results in heavy representation content distortion. To address these issues, we propose a flexible residual binarization (FRB) method for image SR. We first propose a second-order residual binarization (SRB), to counter the information loss caused by binarization. In addition to the primary weight binarization, we also binarize the reconstruction error, which is added as a residual term in the prediction. Furthermore, to narrow the representation content gap between the binarized and full-precision networks, we propose Distillation-guided Binarization Training (DBT). We uniformly align the contents of different bit widths by constructing a normalized attention form. Finally, we generalize our method by applying our FRB to binarize convolution and Transformer-based SR networks, resulting in two binary baselines: FRBC and FRBT. We conduct extensive experiments and comparisons with recent leading binarization methods. Our proposed baselines, FRBC and FRBT, achieve superior performance both quantitatively and visually. The code and model will be released.

Large language models (LLMs) have demonstrated remarkable capabilities in language generation, understanding, and few-shot learning in recent years. An extensive body of work has explored how their performance may be further improved through the tools of prompting, ranging from verification, self-consistency, or intermediate scratchpads. In this paper, we present a complementary approach to improve language responses where multiple language model instances propose and debate their individual responses and reasoning processes over multiple rounds to arrive at a common final answer. Our findings indicate that this approach significantly enhances mathematical and strategic reasoning across a number of tasks. We also demonstrate that our approach improves the factual validity of generated content, reducing fallacious answers and hallucinations that contemporary models are prone to. Our approach may be directly applied to existing black-box models and uses identical procedure and prompts for all tasks we investigate. Overall, our findings suggest that such "society of minds" approach has the potential to significantly advance the capabilities of LLMs and pave the way for further breakthroughs in language generation and understanding.

Gaussian process factor analysis (GPFA) is a latent variable modeling technique commonly used to identify smooth, low-dimensional latent trajectories underlying high-dimensional neural recordings. Specifically, researchers model spiking rates as Gaussian observations, resulting in tractable inference. Recently, GPFA has been extended to model spike count data. However, due to the non-conjugacy of the likelihood, the inference becomes intractable. Prior works rely on either black-box inference techniques, numerical integration or polynomial approximations of the likelihood to handle intractability. To overcome this challenge, we propose a conditionally-conjugate Gaussian process factor analysis (ccGPFA) resulting in both analytically and computationally tractable inference for modeling neural activity from spike count data. In particular, we develop a novel data augmentation based method that renders the model conditionally conjugate. Consequently, our model enjoys the advantage of simple closed-form updates using a variational EM algorithm. Furthermore, due to its conditional conjugacy, we show our model can be readily scaled using sparse Gaussian Processes and accelerated inference via natural gradients. To validate our method, we empirically demonstrate its efficacy through experiments.

In this work, we give a new technique for analyzing individualized privacy accounting via the following simple observation: if an algorithm is one-sided add-DP, then its subsampled variant satisfies two-sided DP. From this, we obtain several improved algorithms for private combinatorial optimization problems, including decomposable submodular maximization and set cover. Our error guarantees are asymptotically tight and our algorithm satisfies pure-DP while previously known algorithms (Gupta et al., 2010; Chaturvedi et al., 2021) are approximate-DP. We also show an application of our technique beyond combinatorial optimization by giving a pure-DP algorithm for the shifting heavy hitter problem in a stream; previously, only an approximate-DP algorithm was known (Kaplan et al., 2021; Cohen & Lyu, 2023).

Deep heteroscedastic regression involves jointly optimizing the mean and covariance of the predicted distribution using the negative log-likelihood. However, recent works show that this may result in sub-optimal convergence due to the challenges associated with covariance estimation. While the literature addresses this by proposing alternate formulations to mitigate the impact of the predicted covariance, we focus on improving the predicted covariance itself. We study two questions: (1) Does the predicted covariance truly capture the randomness of the predicted mean? (2) In the absence of supervision, how can we quantify the accuracy of covariance estimation? We address (1) with a _Taylor Induced Covariance (TIC)_, which captures the randomness of the predicted mean by incorporating its gradient and curvature through the second order Taylor polynomial. Furthermore, we tackle (2) by introducing a _Task Agnostic Correlations (TAC)_ metric, which combines the notion of correlations and absolute error to evaluate the covariance. We evaluate TIC-TAC across multiple experiments spanning synthetic and real-world datasets. Our results show that not only does TIC accurately learn the covariance, it additionally facilitates an improved convergence of the negative log-likelihood. Our code is available at [https://github.com/vita-epfl/TIC-TAC](https://github.com/vita-epfl/TIC-TAC)

In this work, we study the issue of reward hacking on the response length, a challenge emerging in Reinforcement Learning from Human Feedback (RLHF) on LLMs. A well-formatted, verbose but less helpful response from the LLMs can often deceive LLMs or even human evaluators and achieve high scores. The same issue also holds for some reward models in RL. To address the challenges in both training and evaluation, we establish a more reliable evaluation protocol for comparing different training configurations, which inspects the trade-off between LLM evaluation score and response length obtained by varying training hyperparameters. Based on this evaluation, we conduct large-scale studies, where the results shed insights into the efficacy of hyperparameters and tricks used in RL on mitigating length bias. We further propose to improve the reward model by jointly training two linear heads to predict the preference, one trained to correlate with length and the other trained to decorrelate with length and therefore focusing more on the actual content. We then discard the length head in RL to ignore the spurious length reward. Experiments demonstrate that our approach eliminates the reward correlation with length, and improves the obtained policy by a significant margin.

Differentially private ML approaches seek to learn models which may be publicly released while guaranteeing that the input data is kept private. One issue with this construction is that further model releases based on the same training data (e.g. for a new task) incur a further privacy budget cost. Privacy-preserving synthetic data generation is one possible solution to this conundrum. However, models trained on synthetic private data struggle to approach the performance of private, ad-hoc models. In this paper, we present a novel method based on sum-product networks that is able to perform both privacy-preserving classification and privacy-preserving data generation with a single model. To the best of our knowledge, ours is the first approach that provides both discriminative and generative capabilities to differentially private ML. We show that our approach outperforms the state of the art in terms of stability (i.e. number of training runs required for convergence) and utility of the generated data.

Mixed integer linear programs (MILP) are flexible and powerful tools for modeling and solving many difficult real-world combinatorial optimization problems. In this paper, we propose a novel machine learning (ML)-based framework ConPaS that learns to predict solutions to MILPs with contrastive learning. For training, we collect high-quality solutions as positive samples. We also collect low-quality or infeasible solutions as negative samples using novel optimization-based or sampling approaches. We then learn to make discriminative predictions by contrasting the positive and negative samples. During testing, we predict and fix the assignments for a subset of integer variables and then solve the resulting reduced MILP to find high-quality solutions. Empirically, ConPaS achieves state-of-the-art results compared to other ML-based approaches in terms of the quality of and the speed at which solutions are found.

This paper contributes a new approach for distributional reinforcement learning which elucidates a clean separation of transition structure and reward in the learning process. Analogous to how the successor representation (SR) describes the expected consequences of behaving according to a given policy, our distributional successor measure (SM) describes the distributional consequences of this behaviour. We formulate the distributional SM as a distribution over distributions and provide theory connecting it with distributional and model-based reinforcement learning. Moreover, we propose an algorithm that learns the distributional SM from data by minimizing a two-level maximum mean discrepancy. Key to our method are a number of algorithmic techniques that are independently valuable for learning generative models of state. As an illustration of the usefulness of the distributional SM, we show that it enables zero-shot risk-sensitive policy evaluation in a way that was not previously possible.

We investigate online Linear Quadratic Regulator (LQR) with bandit feedback and semi-adversarial disturbances. Previous works assume costs with *homogeneous* curvatures (i.e., with a uniform strong convexity lower bound), which can be hard to satisfy in many real scenarios and prohibits adapting to true curvatures for better performance. In this paper, we initiate the study of bandit LQR control with *heterogeneous* cost curvatures, aiming to strengthen the algorithm's adaptivity. To achieve this, we reduce the problem to bandit convex optimization with memory via a ``with-history'' reduction to avoid hard-to-control truncation errors. Then we provide a novel analysis for an important *stability* term that appeared in both regret and memory, using *Newton decrement* developed in interior-point methods. The analysis enables us to guarantee memory-related terms introduced in the reduction and also provide a simplified analysis for handling heterogeneous curvatures in bandit convex optimization. Finally, we achieve interpolated guarantees that can not only recover existing bounds for convex and quadratic costs but also attain new implications for cases of corrupted and decaying quadraticity.

Though feature-alignment based Domain Adaptive Object Detection (DAOD) methods have achieved remarkable progress, they ignore the source bias issue, i.e., the detector tends to acquire more source-specific knowledge, impeding its generalization capabilities in the target domain. Furthermore, these methods face a more formidable challenge in achieving consistent classification and localization in the target domain compared to the source domain. To overcome these challenges, we propose a novel Distillation-based Source Debiasing (DSD) framework for DAOD, which can distill domain-agnostic knowledge from a pre-trained teacher model, improving the detector's performance on both domains. In addition, we design a Target-Relevant Object Localization Network (TROLN), which can mine target-related localization information from source and target-style mixed data. Accordingly, we present a Domain-aware Consistency Enhancing (DCE) strategy, in which these information are formulated into a new localization representation to further refine classification scores in the testing stage, achieving a harmonization between classification and localization. Extensive experiments have been conducted to manifest the effectiveness of this method, which consistently improves the strong baseline by large margins, outperforming existing alignment-based works.

A large body of work has investigated the properties of graph neural networks and identified several limitations, particularly pertaining to their expressive power. Their inability to count certain *patterns* (e.g., cycles) in a graph lies at the heart of such limitations, since many functions to be learned rely on the ability of counting such patterns. Two prominent paradigms aim to address this limitation by enriching the graph features with *subgraph* or *homomorphism* pattern counts. In this work, we show that both of these approaches are sub-optimal in a certain sense and argue for a more *fine-grained* approach, which incorporates the homomorphism counts of *all* structures in the ``basis'' of the target pattern. This yields strictly more expressive architectures without incurring any additional overhead in terms of computational complexity compared to existing approaches. We prove a series of theoretical results on node-level and graph-level *motif parameters* and empirically validate them on standard benchmark datasets.