ICML 2024 论文列表

As a branch of clustering, multi-view clustering has received much attention in recent years. In practical applications, a common phenomenon is that partial views of some samples may be missing in the collected multi-view data, which poses a severe challenge to design the multi-view learning model and explore complementary and consistent information. Currently, most of the incomplete multi-view clustering methods only focus on exploring the information of available views while few works study the missing view recovery for incomplete multi-view learning. To this end, we propose an innovative diffusion-based missing view generation (DMVG) network. Moreover, for the scenarios with high missing rates, we further propose an incomplete multi-view data augmentation strategy to enhance the recovery quality for the missing views. Extensive experimental results show that the proposed DMVG can not only accurately predict missing views, but also further enhance the subsequent clustering performance in comparison with several state-of-the-art incomplete multi-view clustering methods.

We propose HyperAgent, a reinforcement learning (RL) algorithm based on the hypermodel framework for exploration in RL. HyperAgent allows for the efficient incremental approximation of posteriors associated with an optimal action-value function ($Q^\star$) without the need for conjugacy and follows the greedy policies w.r.t. these approximate posterior samples. We demonstrate that HyperAgent offers robust performance in large-scale deep RL benchmarks. It can solve Deep Sea hard exploration problems with episodes that optimally scale with problem size and exhibits significant efficiency gains in the Atari suite. Implementing HyperAgent requires minimal code addition to well-established deep RL frameworks like DQN. We theoretically prove that, under tabular assumptions, HyperAgent achieves logarithmic per-step computational complexity while attaining sublinear regret, matching the best known randomized tabular RL algorithm.

Semiparametric statistics play a pivotal role in a wide range of domains, including but not limited to missing data, causal inference, and transfer learning, to name a few. In many settings, semiparametric theory leads to (nearly) statistically optimal procedures that yet involve numerically solving Fredholm integral equations of the second kind. Traditional numerical methods, such as polynomial or spline approximations, are difficult to scale to multi-dimensional problems. Alternatively, statisticians may choose to approximate the original integral equations by ones with closed-form solutions, resulting in computationally more efficient, but statistically suboptimal or even incorrect procedures. To bridge this gap, we propose a novel framework by formulating the semiparametric estimation problem as a bi-level optimization problem; and then we propose a scalable algorithm called **D**eep **N**eural-Nets **A**ssisted **S**emiparametric **E**stimation ($\mathsf{DNA\mbox{-}SE}$) by leveraging the universal approximation property of Deep Neural-Nets (DNN) to streamline semiparametric procedures. Through extensive numerical experiments and a real data analysis, we demonstrate the numerical and statistical advantages of $\mathsf{DNA\mbox{-}SE}$ over traditional methods. To the best of our knowledge, we are the first to bring DNN into semiparametric statistics as a numerical solver of integral equations in our proposed general framework.

Large Language Model or LLM inference has two phases, the prompt (or prefill) phase to output the first token and the extension (or decoding) phase to the generate subsequent tokens. In this work, we propose an efficient parallelization scheme, KV-Runahead to accelerate the prompt phase. The key observation is that the extension phase generates tokens faster than the prompt phase because of key-value cache (KV-cache). Hence, KV-Runahead parallelizes the prompt phase by orchestrating multiple processes to populate the KV-cache and minimizes the time-to-first-token (TTFT). Dual-purposing the KV-cache scheme has two main benefits. First, since KV-cache is designed to leverage the causal attention map, we minimize computation and computation automatically. Second, since it already exists for the extension phase, KV-Runahead is easy to implement. We further propose context-level load-balancing to handle uneven KV-cache generation (due to the causal attention) and to optimize TTFT. Compared with an existing parallelization scheme such as tensor or sequential parallelization where keys and values are locally generated and exchanged via all-gather collectives, our experimental results demonstrate that KV-Runahead can offer over 1.4× and 1.6× speedups for Llama 7B and Falcon 7B respectively.

In-context learning is a powerful emergent ability in transformer models. Prior work in mechanistic interpretability has identified a circuit element that may be critical for in-context learning – the induction head (IH), which performs a match-and-copy operation. During training of large transformers on natural language data, IHs emerge around the same time as a notable phase change in the loss. Despite the robust evidence for IHs and this interesting coincidence with the phase change, relatively little is known about the diversity and emergence dynamics of IHs. Why is there more than one IH, and how are they dependent on each other? Why do IHs appear all of a sudden, and what are the subcircuits that enable them to emerge? We answer these questions by studying IH emergence dynamics in a controlled setting by training on synthetic data. In doing so, we develop and share a novel optogenetics-inspired causal framework for modifying activations throughout training. Using this framework, we delineate the diverse and additive nature of IHs. By "clamping" subsets of activations throughout training, we then identify three underlying subcircuits that interact to drive IH formation, yielding the phase change. Furthermore, these subcircuits shed light on data-dependent properties of formation, such as phase change timing, already showing the promise of this more in-depth understanding of subcircuits that need to "go right" for an induction head.

In most real-world reinforcement learning applications, state information is only partially observable, which breaks the Markov decision process assumption and leads to inferior performance for algorithms that conflate observations with state. Partially Observable Markov Decision Processes (POMDPs), on the other hand, provide a general framework that allows for partial observability to be accounted for in *learning, exploration and planning*, but presents significant computational and statistical challenges. To address these difficulties, we develop a representation-based perspective that leads to a coherent framework and tractable algorithmic approach for practical reinforcement learning from partial observations. We provide a theoretical analysis for justifying the statistical efficiency of the proposed algorithm, and also empirically demonstrate the proposed algorithm can surpass state-of-the-art performance with partial observations across various benchmarks, advancing reliable reinforcement learning towards more practical applications.

In class-imbalanced learning, the scarcity of information about minority classes presents challenges in obtaining generalizable features for these classes. Leveraging large-scale pre-trained models with powerful generalization capabilities as teacher models can help fill this information gap. Traditional knowledge distillation transfers the label distribution $p(\boldsymbol{y}|\boldsymbol{x})$ predicted by the teacher model to the student model. However, this method falls short on imbalanced data as it fails to capture the class-conditional probability distribution $p(\boldsymbol{x}|\boldsymbol{y})$ from the teacher model, which is crucial for enhancing generalization. To overcome this, we propose Class-Conditional Knowledge Distillation (CCKD), a novel approach that enables learning of the teacher model’s class-conditional probability distribution during the distillation process. Additionally, we introduce Augmented CCKD (ACCKD), which involves distillation on a constructed class-balanced dataset (formed through data mixing) and feature imitation on the entire dataset to further facilitate the learning of features. Experimental results on various imbalanced datasets demonstrate an average accuracy improvement of 7.4% using our method.

Harnessing the power of human-annotated data through Supervised Fine-Tuning (SFT) is pivotal for advancing Large Language Models (LLMs). In this paper, we delve into the prospect of growing a strong LLM out of a weak one without the need for acquiring additional human-annotated data. We propose a new fine-tuning method called Self-Play fIne-tuNing (SPIN), which starts from a supervised fine-tuned model. At the heart of SPIN lies a self-play mechanism, where the LLM refines its capability by playing against instances of itself. More specifically, the LLM generates its own training data from its previous iterations, refining its policy by discerning these self-generated responses from those obtained from human-annotated data. Our method progressively elevates the LLM from a nascent model to a formidable one, unlocking the full potential of human-annotated demonstration data for SFT. Theoretically, we prove that the global optimum to the training objective function of our method is achieved only when the LLM policy aligns with the target data distribution. Empirically, we evaluate our method on several benchmark datasets including the HuggingFace Open LLM Leaderboard, MT-Bench, and datasets from Big-Bench. Our results show that SPIN can significantly improve the LLM's performance across a variety of benchmarks and even outperform models trained through direct preference optimization (DPO) supplemented with extra GPT-4 preference data. This sheds light on the promise of self-play, enabling the achievement of human-level performance in LLMs without the need for expert opponents.

Incomplete multi-view clustering (IMVC) methods typically encounter three drawbacks: (1) intense time and/or space overheads; (2) intractable hyper-parameters; (3) non-zero variance results. With these concerns in mind, we give a simple yet effective IMVC scheme, termed as ToRES. Concretely, instead of self-expression affinity, we manage to construct prototype-sample affinity for incomplete data so as to decrease the memory requirements. To eliminate hyper-parameters, besides mining complementary features among views by view-wise prototypes, we also attempt to devise cross-view prototypes to capture consensus features for jointly forming high-quality clustering representation. To avoid the variance, we successfully unify representation learning and clustering operation, and directly optimize the discrete cluster indicators from incomplete data. Then, for the resulting objective function, we provide two equivalent solutions from perspectives of feasible region partitioning and objective transformation. Many results suggest that ToRES exhibits advantages against 20 SOTA algorithms, even in scenarios with a higher ratio of incomplete data.

Graph self-supervised learning has sparked a research surge in training informative representations without accessing any labeled data. However, our understanding of graph self-supervised learning remains limited, and the inherent relationships between various self-supervised tasks are still unexplored. Our paper aims to provide a fresh understanding of graph self-supervised learning based on task correlations. Specifically, we evaluate the performance of the representations trained by one specific task on other tasks and define correlation values to quantify task correlations. Through this process, we unveil the task correlations between various self-supervised tasks and can measure their expressive capabilities, which are closely related to downstream performance. By analyzing the correlation values between tasks across various datasets, we reveal the complexity of task correlations and the limitations of existing multi-task learning methods. To obtain more capable representations, we propose Graph Task Correlation Modeling (GraphTCM) to illustrate the task correlations and utilize it to enhance graph self-supervised training. The experimental results indicate that our method significantly outperforms existing methods across various downstream tasks.

Class-imbalanced node classification tasks are prevalent in real-world scenarios. Due to the uneven distribution of nodes across different classes, learning high-quality node representations remains a challenging endeavor. The engineering of loss functions has shown promising potential in addressing this issue. It involves the meticulous design of loss functions, utilizing information about the quantities of nodes in different categories and the network's topology to learn unbiased node representations. However, the design of these loss functions heavily relies on human expert knowledge and exhibits limited adaptability to specific target tasks. In this paper, we introduce a high-performance, flexible, and generalizable automated loss function search framework to tackle this challenge. Across 15 combinations of graph neural networks and datasets, our framework achieves a significant improvement in performance compared to state-of-the-art methods. Additionally, we observe that homophily in graph-structured data significantly contributes to the transferability of the proposed framework.

Counterfactual estimation from observations represents a critical endeavor in numerous application fields, such as healthcare and finance, with the primary challenge being the mitigation of treatment bias. The balancing strategy aimed at reducing covariate disparities between different treatment groups serves as a universal solution. However, when it comes to the time series data, the effectiveness of balancing strategies remains an open question, with a thorough analysis of the robustness and applicability of balancing strategies still lacking. This paper revisits counterfactual estimation in the temporal setting and provides a brief overview of recent advancements in balancing strategies. More importantly, we conduct a critical empirical examination for the effectiveness of the balancing strategies within the realm of temporal counterfactual estimation in various settings on multiple datasets. Our findings could be of significant interest to researchers and practitioners and call for a reexamination of the balancing strategy in time series settings.

Byzantine-robust learning has emerged as a prominent fault-tolerant distributed machine learning framework. However, most techniques focus on the *static* setting, wherein the identity of Byzantine workers remains unchanged throughout the learning process. This assumption fails to capture real-world *dynamic* Byzantine behaviors, which may include intermittent malfunctions or targeted, time-limited attacks. Addressing this limitation, we propose DynaBRO -- a new method capable of withstanding any sub-linear number of identity changes across rounds. Specifically, when the number of such changes is $\mathcal{O}(\sqrt{T})$ (where $T$ is the total number of training rounds), DynaBRO nearly matches the state-of-the-art asymptotic convergence rate of the static setting. Our method utilizes a multi-level Monte Carlo (MLMC) gradient estimation technique applied at the server to robustly aggregated worker updates. By additionally leveraging an adaptive learning rate, we circumvent the need for prior knowledge of the fraction of Byzantine workers.

Differentially private (DP) machine learning is considered the gold-standard solution for training a model from sensitive data while still preserving privacy. However, a major barrier to achieving this ideal is its sub-optimal privacy-accuracy trade-off, which is particularly visible in DP representation learning. Specifically, it has been shown that under modest privacy budgets, most models learn representations that are not significantly better than hand-crafted features. In this work, we show that effective DP representation learning can be done via image captioning and scaling up to internet-scale multimodal datasets. Through a series of engineering tricks, we successfully train a DP image captioner (DP-Cap) on a 233M subset of LAION-2B from scratch using a reasonable amount of computation, and obtaining unprecedented high-quality image features that can be used in a variety of downstream vision and vision-language tasks. For example, under a privacy budget of $\varepsilon=8$ for the LAION dataset, a linear classifier trained on top of learned DP-Cap features attains $65.8\%$ accuracy on ImageNet-1K, considerably improving the previous SOTA of $56.5\%$. Our work challenges the prevailing sentiment that high-utility DP representation learning cannot be achieved by training from scratch.

The performance of optimization methods is often tied to the spectrum of the objective Hessian. Yet, conventional assumptions, such as smoothness, do often not enable us to make finely-grained convergence statements—particularly not for non-convex problems. Striving for a more intricate characterization of complexity, we introduce a unique concept termed graded non-convexity. This allows to partition the class of non-convex problems into a nested chain of subclasses. Interestingly, many traditional non-convex objectives, including partially convex problems, matrix factorizations, and neural networks, fall within these subclasses. As a second contribution, we propose gradient methods with spectral preconditioning, which employ inexact top eigenvectors of the Hessian to address the ill-conditioning of the problem, contingent on the grade. Our analysis reveals that these new methods provide provably superior convergence rates compared to basic gradient descent on applicable problem classes, particularly when large gaps exist between the top eigenvalues of the Hessian. Our theory is validated by numerical experiments executed on multiple practical machine learning problems.

To address the shortcomings of real-world datasets, robust learning algorithms have been designed to overcome arbitrary and indiscriminate data corruption. However, practical processes of gathering data may lead to patterns of data corruption that are localized to specific partitions of the training dataset. Motivated by critical applications where the learned model is deployed to make predictions about people from a rich collection of overlapping subpopulations, we initiate the study of *multigroup robust* algorithms whose robustness guarantees for each subpopulation only degrade with the amount of data corruption *inside* that subpopulation. When the data corruption is not distributed uniformly over subpopulations, our algorithms provide more meaningful robustness guarantees than standard guarantees that are oblivious to how the data corruption and the affected subpopulations are related. Our techniques establish a new connection between multigroup fairness and robustness.

Large language models (LLMs) have achieved impressive performance on various natural language generation tasks. Nonetheless, they suffer from generating negative and harmful contents that are biased against certain demographic groups (e.g., female), raising severe fairness concerns. As remedies, prior works intervened the generation by removing attitude or demographic information, inevitably degrading the generation quality and resulting in notable *fairness-fluency* trade-offs. However, it is still under-explored to what extent the fluency *has to* be affected in order to achieve a desired level of fairness. In this work, we conduct the first formal study from an information-theoretic perspective. We show that previous approaches are excessive for debiasing and propose LIDAO, a general framework to debias a (L)LM at a better fluency provably. We further robustify LIDAO in adversarial scenarios, where a carefully-crafted prompt may stimulate LLMs exhibiting instruction-following abilities to generate texts with fairness issue appears only when the prompt is also taken into account. Experiments on three LMs ranging from 0.7B to 7B parameters demonstrate the superiority of our method.

Previous efforts using frozen Large Language Models (LLMs) for visual understanding, via image captioning or image-text retrieval tasks, face challenges when dealing with complex multimodal scenarios. In order to enhance the capabilities of Multimodal Large Language Models (MLLM) in comprehending the context of vision and language, we introduce Multimodal Composition Learning (MCL) for the purpose of mapping or aligning the vision and language input. In particular, we introduce two tasks: Multimodal-Context Captioning (MC-Cap) and Multimodal-Context Retrieval (MC-Ret) to guide a frozen LLM in comprehending the vision and language context. These specialized tasks are crafted to improve the LLM’s capacity for efficient processing and utilization of multimodal inputs, thereby enhancing its proficiency in generating more accurate text or visual representations. Extensive experiments on both retrieval tasks (i.e., zero-shot composed image retrieval, visual storytelling image retrieval and visual dialog image retrieval) and text generation tasks (i.e., visual question answering) demonstrate the effectiveness of the proposed method. The code is available at: https://github.com/dhg-wei/MCL.

Temporal credit assignment in reinforcement learning is challenging due to delayed and stochastic outcomes. Monte Carlo targets can bridge long delays between action and consequence but lead to high-variance targets due to stochasticity. Temporal difference (TD) learning uses bootstrapping to overcome variance but introduces a bias that can only be corrected through many iterations. TD($\lambda$) provides a mechanism to navigate this bias-variance tradeoff smoothly. Appropriately selecting $\lambda$ can significantly improve performance. Here, we propose Chunked-TD, which uses predicted probabilities of transitions from a model for computing $\lambda$-return targets. Unlike other model-based solutions to credit assignment, Chunked-TD is less vulnerable to model inaccuracies. Our approach is motivated by the principle of history compression and ‘chunks’ trajectories for conventional TD learning. Chunking with learned world models compresses near-deterministic regions of the environment-policy interaction to speed up credit assignment while still bootstrapping when necessary. We propose algorithms that can be implemented online and show that they solve some problems much faster than conventional TD($\lambda$).