ICML 2024 Papers

Retrieval pipelines are an integral component of many machine learning systems. However, they perform poorly in domains where documents are long (e.g., 10K tokens or more) and where identifying the relevant document requires synthesizing information across the entire text. Developing long-context retrieval encoders suitable for these domains raises three challenges: (1) how to evaluate long-context retrieval performance, (2) how to pretrain a base language model to represent both short contexts (corresponding to queries) and long contexts (corresponding to documents), and (3) how to finetune this model for retrieval under the batch size limitations imposed by GPU memory constraints. To address these challenges, we first introduce LoCoV1, a 12 task benchmark constructed to measure long-context retrieval where chunking is not possible or not effective. We next present the M2-BERT retrieval encoder, an 80M parameter state-space encoder model built from the Monarch Mixer architecture, capable of scaling to documents up to 32K tokens long. We describe a pretraining data mixture which allows this encoder to process both short and long context sequences, and a finetuning approach that adapts this base model to retrieval with only single-sample batches. Finally, we validate the M2-BERT retrieval encoder on LoCoV1, finding that it outperforms competitive Transformer-based models by at least 22.2 points, despite containing 90× fewer parameters.

A well-trained deep neural network on balanced datasets usually exhibits the Neural Collapse (NC) phenomenon, which is an informative indicator of the model achieving good performance. However, NC is usually hard to be achieved for a model trained on long-tailed datasets, leading to the deteriorated performance of test data. This work aims to induce the NC phenomenon in imbalanced learning from the perspective of distribution matching. By enforcing the distribution of last-layer representations to align the ideal distribution of the ETF structure, we develop a Distribution Alignment Optimization (DisA) loss, acting as a plug-and-play method can be combined with most of the existing long-tailed methods, we further instantiate it to the cases of fixing classifier and learning classifier. The extensive experiments show the effectiveness of DisA, providing a promising solution to the imbalanced issue. Our code is available at DisA.

Multimodal learning methods with targeted unimodal learning objectives have exhibited their superior efficacy in alleviating the imbalanced multimodal learning problem. However, in this paper, we identify the previously ignored gradient conflict between multimodal and unimodal learning objectives, potentially misleading the unimodal encoder optimization. To well diminish these conflicts, we observe the discrepancy between multimodal loss and unimodal loss, where both gradient magnitude and covariance of the easier-to-learn multimodal loss are smaller than the unimodal one. With this property, we analyze Pareto integration under our multimodal scenario and propose MMPareto algorithm, which could ensure a final gradient with direction that is common to all learning objectives and enhanced magnitude to improve generalization, providing innocent unimodal assistance. Finally, experiments across multiple types of modalities and frameworks with dense cross-modal interaction indicate our superior and extendable method performance. Our method is also expected to facilitate multi-task cases with a clear discrepancy in task difficulty, demonstrating its ideal scalability. The source code and dataset are available at https://github.com/GeWu-Lab/MMPareto_ICML2024.

The existing graph neural architecture search (GNAS) methods assume that the graph tasks are static during the search process, ignoring the ubiquitous scenarios where sequential graph tasks come in a continual fashion. Moreover, existing GNAS works resort to entangled graph factors during the architecture search process, resulting in the catastrophic forgetting problems. In this paper, we study the problem of continual graph neural architecture search that is expected to continually search the architecture to learn new graph tasks without forgetting the past, which remains largely unexplored in the literature. However, this problem poses the challenge of architecture conflicts, i.e., the optimal architecture for the new graph task may have performance deterioration and thus sub-optimal for past tasks. To address the challenge, we propose a novel Disentangled Continual Graph Neural Architecture Search with Invariant Modularization (GASIM) method, which is able to continually search the optimal architectures without forgetting past knowledge. Specifically, we first design a modular graph architecture super-network incorporating multiple modules to enable searching architecture with factor expertise. Second, we propose a factor-based task-module router that discovers the latent graph factors and routes the incoming task to the best suitable architecture module to alleviate the forgetting problem induced by architecture conflicts. Finally, we propose an invariant architecture search mechanism to capture the shared knowledge among tasks. Extensive experiments on real-world datasets demonstrate that the proposed method achieves state-of-the-art performance against baselines in continual graph neural architecture search.

Low-bit integer training emerges as a promising approach to mitigate the heavy burden during network training by quantizing the weights, activations, and gradients. However, existing methods cannot well achieve mixed-precision quantization for low-bit training and are commonly limited to INT8 precision. In this paper, we propose a novel low-bit integer training framework that, for the first time, achieves adaptive mixed-precision allocation (AMPA) for weights, activations, and gradients, and pushes the boundaries to a precision level below INT8. We develop a novel magnitude-based sensitivity measurement with regard to the quantization losses of weight, activation, and gradient quantization and the average gradient magnitudes, which is demonstrated as an upper bound of quantization influence in theory. We further design a layer-wise precision update strategy under observations on the quantization losses and their effects on model performance in low-bit training. Extensive experiments on different backbones and datasets show that, compared to INT8 quantization, the proposed method can achieve more than 38% BitOPs reduction with a tolerable loss below 2% in image classification, image segmentation, and language modeling.

Most works studying representation learning focus only on classification and neglect regression. Yet, the learning objectives and, therefore, the representation topologies of the two tasks are fundamentally different: classification targets class separation, leading to disconnected representations, whereas regression requires ordinality with respect to the target, leading to continuous representations. We thus wonder how the effectiveness of a regression representation is influenced by its topology, with evaluation based on the Information Bottleneck (IB) principle. The IB principle is an important framework that provides principles for learning effective representations. We establish two connections between it and the topology of regression representations. The first connection reveals that a lower intrinsic dimension of the feature space implies a reduced complexity of the representation $Z$. This complexity can be quantified as the conditional entropy of $Z$ on the target $Y$, and serves as an upper bound on the generalization error. The second connection suggests a feature space that is topologically similar to the target space will better align with the IB principle. Based on these two connections, we introduce PH-Reg, a regularizer specific to regression that matches the intrinsic dimension and topology of the feature space with the target space. Experiments on synthetic and real-world regression tasks demonstrate the benefits of PH-Reg. Code: https://github.com/needylove/PH-Reg.

Existing OCR engines or document image analysis systems typically rely on training separate models for text detection in varying scenarios and granularities, leading to significant computational complexity and resource demands. In this paper, we introduce "Detect Any Text" (DAT), an advanced paradigm that seamlessly unifies scene text detection, layout analysis, and document page detection into a cohesive, end-to-end model. This design enables DAT to efficiently manage text instances at different granularities, including *word*, *line*, *paragraph* and *page*. A pivotal innovation in DAT is the across-granularity interactive attention module, which significantly enhances the representation learning of text instances at varying granularities by correlating structural information across different text queries. As a result, it enables the model to achieve mutually beneficial detection performances across multiple text granularities. Additionally, a prompt-based segmentation module refines detection outcomes for texts of arbitrary curvature and complex layouts, thereby improving DAT's accuracy and expanding its real-world applicability. Experimental results demonstrate that DAT achieves state-of-the-art performances across a variety of text-related benchmarks, including multi-oriented/arbitrarily-shaped scene text detection, document layout analysis and page detection tasks.

In dynamical systems reconstruction (DSR) we seek to infer from time series measurements a generative model of the underlying dynamical process. This is a prime objective in any scientific discipline, where we are particularly interested in parsimonious models with a low parameter load. A common strategy here is parameter pruning, removing all parameters with small weights. However, here we find this strategy does not work for DSR, where even low magnitude parameters can contribute considerably to the system dynamics. On the other hand, it is well known that many natural systems which generate complex dynamics, like the brain or ecological networks, have a sparse topology with comparatively few links. Inspired by this, we show that *geometric pruning*, where in contrast to magnitude-based pruning weights with a low contribution to an attractor's geometrical structure are removed, indeed manages to reduce parameter load substantially without significantly hampering DSR quality. We further find that the networks resulting from geometric pruning have a specific type of topology, and that this topology, and not the magnitude of weights, is what is most crucial to performance. We provide an algorithm that automatically generates such topologies which can be used as priors for generative modeling of dynamical systems by RNNs, and compare it to other well studied topologies like small-world or scale-free networks.

Many applications of RCTs involve the presence of multiple treatment administrators---from field experiments to online advertising---that compete for the subjects' attention. In the face of competition, estimating a causal effect becomes difficult, as the position at which a subject sees a treatment influences their response, and thus the treatment effect. In this paper, we build a game-theoretic model of agents who wish to estimate causal effects in the presence of competition, through a bidding system and a utility function that minimizes estimation error. Our main technical result establishes an approximation with a tractable objective that maximizes the sample value obtained through strategically allocating budget on subjects. This allows us to find an equilibrium in our model: we show that the tractable objective has a pure Nash equilibrium, and that any Nash equilibrium is an approximate equilibrium for our general objective that minimizes estimation error under broad conditions. Conceptually, our work successfully combines elements from causal inference and game theory to shed light on the equilibrium behavior of experimentation under competition.

In recent years, there has been a growing research interest in decision-focused learning, which embeds optimization problems as a layer in learning pipelines and demonstrates a superior performance than the prediction-focused approach. However, for distributionally robust optimization (DRO), a popular paradigm for decision-making under uncertainty, it is still unknown how to embed it as a layer, i.e., how to differentiate decisions with respect to an ambiguity set. In this paper, we develop such differentiable DRO layers for generic mixed-integer DRO problems with parameterized second-order conic ambiguity sets and discuss its extension to Wasserstein ambiguity sets. To differentiate the mixed-integer decisions, we propose a novel dual-view methodology by handling continuous and discrete parts of decisions via different principles. Specifically, we construct a differentiable energy-based surrogate to implement the dual-view methodology and use importance sampling to estimate its gradient. We further prove that such a surrogate enjoys the asymptotic convergency under regularization. As an application of the proposed differentiable DRO layers, we develop a novel decision-focused learning pipeline for contextual distributionally robust decision-making tasks and compare it with the prediction-focused approach in experiments

The Weisfeiler--Leman algorithm (1-WL) is a well-studied heuristic for the graph isomorphism problem. Recently, the algorithm has played a prominent role in understanding the expressive power of message-passing graph neural networks (MPNNs) and being effective as a graph kernel. Despite its success, the 1-WL faces challenges in distinguishing non-isomorphic graphs, leading to the development of more expressive MPNN and kernel architectures. However, the relationship between enhanced expressivity and improved generalization performance remains unclear. Here, we show that an architecture's expressivity offers limited insights into its generalization performance when viewed through graph isomorphism. Moreover, we focus on augmenting 1-WL and MPNNs with subgraph information and employ classical margin theory to investigate the conditions under which an architecture's increased expressivity aligns with improved generalization performance. In addition, we introduce variations of expressive 1-WL-based kernel and MPNN architectures with provable generalization properties. Our empirical study confirms the validity of our theoretical findings.

Budget pacing is a popular service that has been offered by major internet advertising platforms since their inception. In the past few years, autobidding products that provide real-time bidding as a service to advertisers have seen a prominent rise in adoption. A popular autobidding stategy is value maximization subject to return-on-spend (ROS) constraints. For historical or business reasons, the systems that govern these two services, namely budget pacing and ROS pacing, are not necessarily always a single unified and coordinated entity that optimizes a global objective subject to both constraints. The purpose of this work is to theoretically and empirically compare algorithms with different degrees of coordination between these two pacing systems. In particular, we compare (a) a fully-decoupled sequential algorithm; (b) a minimally-coupled min-pacing algorithm; (c) a fully-coupled dual-based algorithm. Our main contribution is to theoretically analyze the min-pacing algorithm and show that it attains similar guarantees to the fully-coupled canonical dual-based algorithm. On the other hand, we show that the sequential algorithm, even though appealing by virtue of being fully decoupled, could badly violate the constraints. We validate our theoretical findings empirically by showing that the min-pacing algorithm performs almost as well as the canonical dual-based algorithm on a semi-synthetic dataset that was generated from a large online advertising platform's auction data.

This paper revisits the simple, long-studied, yet still unsolved problem of making image classifiers robust to imperceptible perturbations. Taking CIFAR10 as an example, SOTA clean accuracy is about $100$%, but SOTA robustness to $\ell_{\infty}$-norm bounded perturbations barely exceeds $70$%. To understand this gap, we analyze how model size, dataset size, and synthetic data quality affect robustness by developing the first scaling laws for adversarial training. Our scaling laws reveal inefficiencies in prior art and provide actionable feedback to advance the field. For instance, we discovered that SOTA methods diverge notably from compute-optimal setups, using excess compute for their level of robustness. Leveraging a compute-efficient setup, we surpass the prior SOTA with $20$% ($70$%) fewer training (inference) FLOPs. We trained various compute-efficient models, with our best achieving $74$% AutoAttack accuracy ($+3$% gain). However, our scaling laws also predict robustness slowly grows then plateaus at $90$%: dwarfing our new SOTA by scaling is impractical, and perfect robustness is impossible. To better understand this predicted limit, we carry out a small-scale human evaluation on the AutoAttack data that fools our top-performing model. Concerningly, we estimate that human performance also plateaus near $90$%, which we show to be attributable to $\ell_{\infty}$-constrained attacks' generation of invalid images not consistent with their original labels. Having characterized limiting roadblocks, we outline promising paths for future research.

We tackle the challenge of predicting models' Out-of-Distribution (OOD) performance using in-distribution (ID) measurements without requiring OOD data. Existing evaluations with ``Effective robustness'', which use ID accuracy as an indicator of OOD accuracy, encounter limitations when models are trained with diverse supervision and distributions, such as class labels (*Vision Models, VMs, on ImageNet*) and textual descriptions (*Visual-Language Models, VLMs, on LAION*). VLMs often generalize better to OOD data than VMs despite having similar or lower ID performance. To improve the prediction of models' OOD performance from ID measurements, we introduce the *Lowest Common Ancestor (LCA)-on-the-Line* framework. This approach revisits the established concept of LCA distance, which measures the hierarchical distance between labels and predictions within a predefined class hierarchy, such as WordNet. We assess 75 models using ImageNet as the ID dataset and five significantly shifted OOD variants, uncovering a strong linear correlation between ID LCA distance and OOD top-1 accuracy. Our method provides a compelling alternative for understanding why VLMs tend to generalize better. Additionally, we propose a technique to construct a taxonomic hierarchy on any dataset using $K$-means clustering, demonstrating that LCA distance is robust to the constructed taxonomic hierarchy. Moreover, we demonstrate that aligning model predictions with class taxonomies, through soft labels or prompt engineering, can enhance model generalization. Open source code in our [Project Page](https://elvishelvis.github.io/papers/lca/).

In complex environments with large discrete action spaces, effective decision-making is critical in reinforcement learning (RL). Despite the widespread use of value-based RL approaches like Q-learning, they come with a computational burden, necessitating the maximization of a value function over all actions in each iteration. This burden becomes particularly challenging when addressing large-scale problems and using deep neural networks as function approximators. In this paper, we present stochastic value-based RL approaches which, in each iteration, as opposed to optimizing over the entire set of $n$ actions, only consider a variable stochastic set of a sublinear number of actions, possibly as small as $\mathcal{O}(\log(n))$. The presented stochastic value-based RL methods include, among others, Stochastic Q-learning, StochDQN, and StochDDQN, all of which integrate this stochastic approach for both value-function updates and action selection. The theoretical convergence of Stochastic Q-learning is established, while an analysis of stochastic maximization is provided. Moreover, through empirical validation, we illustrate that the various proposed approaches outperform the baseline methods across diverse environments, including different control problems, achieving near-optimal average returns in significantly reduced time.

The study of machine learning under limited resources has gathered increasing attention, considering improving the learning efficiency and effectiveness with budgeted resources. However, previous efforts mainly focus on *single* learning task, and a common resource-limited scenario is less explored: to handle *multiple* time-constrained learning tasks concurrently with budgeted computational resources. In this paper, we point out that this is a very challenging task because it demands the learner to be concerned about not only the progress of the learning tasks but also the coordinative allocation of computational resources. We present the *Learning with Adaptive Resource Allocation* (LARA) approach, which comprises an efficient online estimator for learning progress prediction, an adaptive search method for computational resource allocation, and a balancing strategy for alleviating prediction-allocation compounding errors. Empirical studies validate the effectiveness of our proposed approach.

Bootstrap is a popular methodology for simulating input uncertainty. However, it can be computationally expensive when the number of samples is large. We propose a new approach called **Orthogonal Bootstrap** that reduces the number of required Monte Carlo replications. We decomposes the target being simulated into two parts: the *non-orthogonal part* which has a closed-form result known as Infinitesimal Jackknife and the *orthogonal part* which is easier to be simulated. We theoretically and numerically show that Orthogonal Bootstrap significantly reduces the computational cost of Bootstrap while improving empirical accuracy and maintaining the same width of the constructed interval.

We study inductive bias in Transformers in the infinitely over-parameterized Gaussian process limit and argue transformers tend to be biased towards more permutation symmetric functions in sequence space. We show that the representation theory of the symmetric group can be used to give quantitative analytical predictions when the dataset is symmetric to permutations between tokens. We present a simplified transformer block and solve the model at the limit, including accurate predictions for the learning curves and network outputs. We show that in common setups, one can derive tight bounds in the form of a scaling law for the learnability as a function of the context length. Finally, we argue WikiText dataset, does indeed possess a degree of permutation symmetry.

We focus on verifying relational properties defined over deep neural networks (DNNs) such as robustness against universal adversarial perturbations (UAP), certified worst-case hamming distance for binary string classifications, etc. Precise verification of these properties requires reasoning about multiple executions of the same DNN. However, most of the existing works in DNN verification only handle properties defined over single executions and as a result, are imprecise for relational properties. Though few recent works for relational DNN verification, capture linear dependencies between the inputs of multiple executions, they do not leverage dependencies between the outputs of hidden layers producing imprecise results. We develop a scalable relational verifier RACoon that utilizes cross-execution dependencies at all layers of the DNN gaining substantial precision over SOTA baselines on a wide range of datasets, networks, and relational properties.