ICML 2024 Papers

We propose a model-based offline reinforcement learning (RL) algorithm for confounded partially observable Markov decision processes (POMDPs) under general function approximations and show it is provably efficient under some technical conditions such as the partial coverage imposed on the offline data distribution. Specifically, we first establish a novel model-based identification result for learning the effect of any action on the reward and future transitions in the confounded POMDP. Using this identification result, we then design a nonparametric two-stage estimation procedure to construct an estimator for off-policy evaluation (OPE), which permits general function approximations. Finally, we learn the optimal policy by performing a conservative policy optimization within the confidence regions based on the proposed estimation procedure for OPE. Under some mild conditions, we establish a finite-sample upper bound on the suboptimality of the learned policy in finding the optimal one, which depends on the sample size and the length of horizons polynomially.

Nowadays, Large Language Models (LLMs) mostly possess billions of parameters, bringing significant challenges to hardware platforms. Although quantization is an efficient approach to reduce computation and memory overhead for inference optimization, we stress the challenge that mainstream low-bit quantization approaches still suffer from either various data distribution outliers or a lack of hardware efficiency. We also find that low-bit data format has further potential expressiveness to cover the atypical language data distribution. In this paper, we propose a novel numerical representation, Bi-Exponent Block Floating Point (BiE), and a new quantization flow. BiE quantization shows accuracy superiority and hardware friendliness on various models and benchmarks.

Large language models have ushered in a new era of artificial intelligence research. However, their substantial training costs hinder further development and widespread adoption. In this paper, inspired by the redundancy in the parameters of large language models, we propose a novel training paradigm: Evolving Subnetwork Training (EST). EST samples subnetworks from the layers of the large language model and from commonly used modules within each layer, Multi-Head Attention (MHA) and Multi-Layer Perceptron (MLP). By gradually increasing the size of the subnetworks during the training process, EST can save the cost of training. We apply EST to train GPT2 model and TinyLlama model, resulting in 26.7% FLOPs saving for GPT2 and 25.0% for TinyLlama without an increase in loss on the pre-training dataset. Moreover, EST leads to performance improvements in downstream tasks, indicating that it benefits generalization. Additionally, we provide intuitive theoretical studies based on training dynamics and Dropout theory to ensure the feasibility of EST.

Amortized Bayesian inference trains neural networks to solve stochastic inference problems using model simulations, thereby making it possible to rapidly perform Bayesian inference for any newly observed data. However, current simulation-based amortized inference methods are simulation-hungry and inflexible: They require the specification of a fixed parametric prior, simulator, and inference tasks ahead of time. Here, we present a new amortized inference method---the Simformer---which overcomes these limitations. By training a probabilistic diffusion model with transformer architectures, the Simformer outperforms current state-of-the-art amortized inference approaches on benchmark tasks and is substantially more flexible: It can be applied to models with function-valued parameters, it can handle inference scenarios with missing or unstructured data, and it can sample arbitrary conditionals of the joint distribution of parameters and data, including both posterior and likelihood. We showcase the performance and flexibility of the Simformer on simulators from ecology, epidemiology, and neuroscience, and demonstrate that it opens up new possibilities and application domains for amortized Bayesian inference on simulation-based models.

One-shot imitation learning (OSIL) is to learn an imitator agent that can execute multiple tasks with only a single demonstration. In real-world scenario, the environment is dynamic, e.g., unexpected changes can occur after demonstration. Thus, achieving generalization of the imitator agent is crucial as agents would inevitably face situations unseen in the provided demonstrations. While traditional OSIL methods excel in relatively stationary settings, their adaptability to such unforeseen changes, which asking for a higher level of generalization ability for the imitator agents, is limited and rarely discussed. In this work, we present a new algorithm called Deep Demonstration Tracing (DDT). In DDT, we propose a demonstration transformer architecture to encourage agents to adaptively trace suitable states in demonstrations. Besides, it integrates OSIL into a meta-reinforcement-learning training paradigm, providing regularization for policies in unexpected situations. We evaluate DDT on a new navigation task suite and robotics tasks, demonstrating its superior performance over existing OSIL methods across all evaluated tasks in dynamic environments with unforeseen changes. The project page is in https://osil-ddt.github.io.

This work proposes an algorithm improving the dimensionality dependence for gradient-free optimisation over cross-polytopes, which has many applications such as adversarial attacks, explainable AI and sparse regression. For bandit convex optimisation with two-point feedback over cross-polytopes, the state-of-the-art algorithms have a dimensionality dependence of $\mathcal{O}(\sqrt{d\log d})$, while the known lower bound is of the form $\Omega(\sqrt{d(\log d)^{-1}})$. We propose a mirror descent algorithm equipped with a symmetric version of the negative $\frac{1}{2}$-Tsallis entropy. Combined with an $\ell_1$-ellipsoidal smoothing-based gradient estimator, the proposed algorithm guarantees a dimensionality dependence on $\mathcal{O}(\sqrt{d})$, which improves the state-of-the-art algorithms by a factor of $\sqrt{\log d}$. The idea can be further applied to optimising non-smooth and non-convex functions. We propose an algorithm with a convergence depending on $\mathcal{O}(d)$, which is the best-known dimensionality dependence.

The weighted squared loss is a common component in several Collaborative Filtering (CF) algorithms for item recommendation, including the representative implicit Alternating Least Squares (iALS). Despite its widespread use, this loss function lacks a clear connection to ranking objectives such as Discounted Cumulative Gain (DCG), posing a fundamental challenge in explaining the exceptional ranking performance observed in these algorithms. In this work, we make a breakthrough by establishing a connection between squared loss and ranking metrics through a Taylor expansion of the DCG-consistent surrogate loss—softmax loss. We also discover a new surrogate squared loss function, namely Ranking-Generalizable Squared (RG$^2$) loss, and conduct thorough theoretical analyses on the DCG-consistency of the proposed loss function. Later, we present an example of utilizing the RG$^2$ loss with Matrix Factorization (MF), coupled with a generalization upper bound and an ALS optimization algorithm that leverages closed-form solutions over all items. Experimental results over three public datasets demonstrate the effectiveness of the RG$^2$ loss, exhibiting ranking performance on par with, or even surpassing, the softmax loss while achieving faster convergence.

Neural collapse provides an elegant mathematical characterization of learned last layer representations (a.k.a. features) and classifier weights in deep classification models. Such results not only provide insights but also motivate new techniques for improving practical deep models. However, most of the existing empirical and theoretical studies in neural collapse focus on the case that the number of classes is small relative to the dimension of the feature space. This paper extends neural collapse to cases where the number of classes are much larger than the dimension of feature space, which broadly occur for language models, retrieval systems, and face recognition applications. We show that the features and classifier exhibit a generalized neural collapse phenomenon, where the minimum one-vs-rest margins is maximized. We provide empirical study to verify the occurrence of generalized neural collapse in practical deep neural networks. Moreover, we provide theoretical study to show that the generalized neural collapse provably occurs under unconstrained feature model with spherical constraint, under certain technical conditions on feature dimension and number of classes.

Vision Transformers (ViT), when paired with large-scale pretraining, have shown remarkable performance across various computer vision tasks, primarily due to their weak inductive bias. However, while such weak inductive bias aids in pretraining scalability, this may hinder the effective adaptation of ViTs for visuo-motor control tasks as a result of the absence of control-centric inductive biases. Such absent inductive biases include spatial locality and translation equivariance bias which convolutions naturally offer. To this end, we introduce Convolution Injector (CoIn), an add-on module that injects convolutions which are rich in locality and equivariance biases into a pretrained ViT for effective adaptation in visuo-motor control. We evaluate CoIn with three distinct types of pretrained ViTs (CLIP, MVP, VC-1) across 12 varied control tasks within three separate domains (Adroit, MetaWorld, DMC), and demonstrate that CoIn consistently enhances control task performance across all experimented environments and models, validating the effectiveness of providing pretrained ViTs with control-centric biases.

The objective in eXtreme Classification (XC) is to find relevant labels for a document from an exceptionally large label space. Most XC application scenarios have rich auxiliary data associated with the input documents, e.g., frequently clicked webpages for search queries in sponsored search. Unfortunately, most of the existing XC methods do not use any auxiliary data. In this paper, we propose a novel framework, Online Auxiliary Knowledge (OAK), which harnesses auxiliary information linked to the document to improve XC accuracy. OAK stores information learnt from the auxiliary data in a knowledge bank and during a forward pass, retrieves relevant auxiliary knowledge embeddings for a given document. An enriched embedding is obtained by fusing these auxiliary knowledge embeddings with the document's embedding, thereby enabling much more precise candidate label selection and final classification. OAK training involves three stages. (1) Training a linker module to link documents to relevant auxiliary data points. (2) Learning an embedding for documents enriched using linked auxiliary information. (3) Using the enriched document embeddings to learn the final classifiers. OAK outperforms current state-of-the-art XC methods by up to $\sim 5 \%$ on academic datasets, and by $\sim 3 \%$ on an auxiliary data-augmented variant of LF-ORCAS-800K dataset in Precision@1. OAK also demonstrates statistically significant improvements in sponsored search metrics when deployed on a large scale search engine.

In this work, we aimed at forecasting the number of SARS-CoV-2 hospitalized patients at 14 days to help anticipate the bed requirements of a large scale hospital using public data and electronic health records data. Previous attempts led to mitigated performance in this high-dimension setting; we introduce a novel approach to time series forecasting by providing an alternative to conventional methods to deal with high number of potential features of interest (409 predictors). We integrate Reservoir Computing (RC) with feature selection using a genetic algorithm (GA) to gather optimal non-linear combinations of inputs to improve prediction in sample-efficient context. We illustrate that the RC-GA combination exhibits excellent performance in forecasting SARS-CoV-2 hospitalizations. This approach outperformed the use of RC alone and other conventional methods: LSTM, Transformers, Elastic-Net, XGBoost. Notably, this work marks the pioneering use of RC (along with GA) in the realm of short and high-dimensional time series, positioning it as a competitive and innovative approach in comparison to standard methods.

As large language models (LLMs) become more capable, fine-tuning techniques for aligning with human intent are increasingly important. A key consideration for aligning these models is how to most effectively use human resources, or model resources in the case where LLMs themselves are used as oracles. Reinforcement learning from Human or AI preferences (RLHF/RLAIF) is the most prominent example of such a technique, but is complex and often unstable. Direct Preference Optimization (DPO) has recently been proposed as a simpler and more stable alternative. In this work, we develop an active learning strategy for DPO to make better use of preference labels. We propose a practical acquisition function for prompt/completion pairs based on the predictive entropy of the language model and a measure of certainty of the implicit preference model optimized by DPO. We demonstrate how our approach improves both the rate of learning and final performance of fine-tuning on pairwise preference data.

We propose Deep Longitudinal Targeted Minimum Loss-based Estimation (Deep LTMLE), a novel approach to estimate the counterfactual mean of outcome under dynamic treatment policies in longitudinal problem settings. Our approach utilizes a transformer architecture with heterogeneous type embedding trained using temporal-difference learning. After obtaining an initial estimate using the transformer, following the targeted minimum loss-based likelihood estimation (TMLE) framework, we statistically corrected for the bias commonly associated with machine learning algorithms. Furthermore, our method also facilitates statistical inference by enabling the provision of 95% confidence intervals grounded in asymptotic statistical theory. Simulation results demonstrate our method's superior performance over existing approaches, particularly in complex, long time-horizon scenarios. It remains effective in small-sample, short-duration contexts, matching the performance of asymptotically efficient estimators. To demonstrate our method in practice, we applied our method to estimate counterfactual mean outcomes for standard versus intensive blood pressure management strategies in a real-world cardiovascular epidemiology cohort study.

Time series data, characterized by its intrinsic long and short-range dependencies, poses a unique challenge across analytical applications. While Transformer-based models excel at capturing long-range dependencies, they face limitations in noise sensitivity, computational efficiency, and overfitting with smaller datasets. In response, we introduce a novel **T**ime **S**eries **L**ightweight **A**daptive **Net**work (**TSLANet**), as a universal convolutional model for diverse time series tasks. Specifically, we propose an Adaptive Spectral Block, harnessing Fourier analysis to enhance feature representation and to capture both long-term and short-term interactions while mitigating noise via adaptive thresholding. Additionally, we introduce an Interactive Convolution Block and leverage self-supervised learning to refine the capacity of TSLANet for decoding complex temporal patterns and improve its robustness on different datasets. Our comprehensive experiments demonstrate that TSLANet outperforms state-of-the-art models in various tasks spanning classification, forecasting, and anomaly detection, showcasing its resilience and adaptability across a spectrum of noise levels and data sizes. The code is available at https://github.com/emadeldeen24/TSLANet.

Visual Model-Based Reinforcement Learning (MBRL) promises to encapsulate agent's knowledge about the underlying dynamics of the environment, enabling learning a world model as a useful planner. However, top MBRL agents such as Dreamer often struggle with visual pixel-based inputs in the presence of exogenous or irrelevant noise in the observation space, due to failure to capture task-specific features while filtering out irrelevant spatio-temporal details. To tackle this problem, we apply a spatio-temporal masking strategy, a bisimulation principle, combined with latent reconstruction, to capture endogenous task-specific aspects of the environment for world models, effectively eliminating non-essential information. Joint training of representations, dynamics, and policy often leads to instabilities. To further address this issue, we develop a Hybrid Recurrent State-Space Model (HRSSM) structure, enhancing state representation robustness for effective policy learning. Our empirical evaluation demonstrates significant performance improvements over existing methods in a range of visually complex control tasks such as Maniskill with exogenous distractors from the Matterport environment. Our code is avaliable at https://github.com/bit1029public/HRSSM.

In recent years there has been a tremendous surge in the general capabilities of AI systems, mainly fuelled by training foundation models on internet-scale data. Nevertheless, the creation of open-ended, ever self-improving AI remains elusive. **In this position paper, we argue that the ingredients are now in place to achieve *open-endedness* in AI systems with respect to a human observer. Furthermore, we claim that such open-endedness is an essential property of any artificial superhuman intelligence (ASI).** We begin by providing a concrete formal definition of open-endedness through the lens of novelty and learnability. We then illustrate a path towards ASI via open-ended systems built on top of foundation models, capable of making novel, human-relevant discoveries. We conclude by examining the safety implications of generally-capable open-ended AI. We expect that open-ended foundation models will prove to be an increasingly fertile and safety-critical area of research in the near future.

Uncertainty Sampling is an Active Learning strategy that aims to improve the data efficiency of machine learning models by iteratively acquiring labels of data points with the highest uncertainty. While it has proven effective for independent data its applicability to graphs remains under-explored. We propose the first extensive study of Uncertainty Sampling for node classification: **(1)** We benchmark Uncertainty Sampling beyond predictive uncertainty and highlight a significant performance gap to other Active Learning strategies. **(2)** We develop ground-truth Bayesian uncertainty estimates in terms of the data generating process and prove their effectiveness in guiding Uncertainty Sampling toward optimal queries. We confirm our results on synthetic data and design an approximate approach that consistently outperforms other uncertainty estimators on real datasets. **(3)** Based on this analysis, we relate pitfalls in modeling uncertainty to existing methods. Our analysis enables and informs the development of principled uncertainty estimation on graphs.

Designing faithful yet accurate AI models is challenging, particularly in the field of individual treatment effect estimation (ITE). ITE prediction models deployed in critical settings such as healthcare should ideally be (i) accurate, and (ii) provide faithful explanations. However, current solutions are inadequate: state-of-the-art black-box models do not supply explanations, post-hoc explainers for black-box models lack faithfulness guarantees, and self-interpretable models greatly compromise accuracy. To address these issues, we propose DISCRET, a self-interpretable ITE framework that synthesizes faithful, rule-based explanations for each sample. A key insight behind DISCRET is that explanations can serve dually as *database queries* to identify similar subgroups of samples. We provide a novel RL algorithm to efficiently synthesize these explanations from a large search space. We evaluate DISCRET on diverse tasks involving tabular, image, and text data. DISCRET outperforms the best self-interpretable models and has accuracy comparable to the best black-box models while providing faithful explanations. DISCRET is available at https://github.com/wuyinjun-1993/DISCRET-ICML2024.

The $k$-Weisfeiler-Leman ($k$-WL) graph isomorphism test hierarchy is a common method for assessing the expressive power of graph neural networks (GNNs). Recently, GNNs whose expressive power is equivalent to the $2$-WL test were proven to be universal on weighted graphs which encode $3\mathrm{D}$ point cloud data, yet this result is limited to invariant continuous functions on point clouds. In this paper, we extend this result in three ways: Firstly, we show that PPGN can simulate $2$-WL uniformly on all point clouds with low complexity. Secondly, we show that $2$-WL tests can be extended to point clouds which include both positions and velocities, a scenario often encountered in applications. Finally, we provide a general framework for proving equivariant universality and leverage it to prove that a simple modification of this invariant PPGN architecture can be used to obtain a universal equivariant architecture that can approximate all continuous equivariant functions uniformly. Building on our results, we develop our WeLNet architecture, which sets new state-of-the-art results on the N-Body dynamics task and the GEOM-QM9 molecular conformation generation task.