ICML 2023 Papers

Compositional data with a large number of components and an abundance of zeros are frequently observed in many fields recently. Analyzing such sparse high-dimensional compositional data naturally calls for dimension reduction or, more preferably, variable selection. Most existing approaches lack interpretability or cannot handle zeros properly, as they rely on a log-ratio transformation. We approach this problem with sufficient dimension reduction (SDR), one of the most studied dimension reduction frameworks in statistics. Characterized by the conditional independence of the data to the response on the found subspace, the SDR framework has been effective for both linear and nonlinear dimension reduction problems. This work proposes a compositional SDR that can handle zeros naturally while incorporating the nonlinear nature and spurious negative correlations among components rigorously. A critical consideration of sub-composition versus amalgamation for compositional variable selection is discussed. The proposed compositional SDR is shown to be statistically consistent in constructing a sub-simplex consisting of true signal variables. Simulation and real microbiome data are used to demonstrate the performance of the proposed SDR compared to existing state-of-art approaches.

Pretraining on a large-scale corpus has become a standard method to build general language models (LMs). Adapting a model to new data distributions targeting different downstream tasks poses significant challenges. Naive fine-tuning may incur catastrophic forgetting when the over-parameterized LMs overfit the new data but fail to preserve the pretrained features. Lifelong learning (LLL) aims to enable information systems to learn from a continuous data stream across time. However, most prior work modifies the training recipe assuming a static fixed network architecture. We find that additional model capacity and proper regularization are key elements to achieving strong LLL performance. Thus, we propose Lifelong-MoE, an extensible MoE (Mixture-of-Experts) architecture that dynamically adds model capacity via adding experts with regularized pretaining. Our results show that by only introducing a limited number of extra experts while keeping the computation cost constant, our model can steadily adapt to data distribution shifts while preserving the previous knowledge. Compared to existing lifelong learning approaches, Lifelong-MoE achieves better few-shot performance on NLP tasks. More impressively, Lifelong-MoE surpasses multi-task learning on 19 downstream NLU tasks.

A directed acyclic graph (DAG) provides valuable prior knowledge that is often discarded in regression tasks in machine learning. We show that the independences arising from the presence of collider structures in DAGs provide meaningful inductive biases, which constrain the regression hypothesis space and improve predictive performance. We introduce collider regression, a framework to incorporate probabilistic causal knowledge from a collider in a regression problem. When the hypothesis space is a reproducing kernel Hilbert space, we prove a strictly positive generalisation benefit under mild assumptions and provide closed-form estimators of the empirical risk minimiser. Experiments on synthetic and climate model data demonstrate performance gains of the proposed methodology.

Learning a universal policy across different robot morphologies can significantly improve learning efficiency and generalization in continuous control. However, it poses a challenging multi-task reinforcement learning problem, as the optimal policy may be quite different across robots and critically depend on the morphology. Existing methods utilize graph neural networks or transformers to handle heterogeneous state and action spaces across different morphologies, but pay little attention to the dependency of a robot's control policy on its morphology context. In this paper, we propose a hierarchical architecture to better model this dependency via contextual modulation, which includes two key submodules: (1) Instead of enforcing hard parameter sharing across robots, we use hypernetworks to generate morphology-dependent control parameters; (2) We propose a fixed attention mechanism that solely depends on the morphology to modulate the interactions between different limbs in a robot. Experimental results show that our method not only improves learning performance on a diverse set of training robots, but also generalizes better to unseen morphologies in a zero-shot fashion. The code is publicly available at https://github.com/MasterXiong/ModuMorph.

Algorithms such as Differentially Private SGD enable training machine learning models with formal privacy guarantees. However, because these guarantees hold with respect to unrealistic adversaries, the protection afforded against practical attacks is typically much better. An emerging strand of work empirically estimates the protection afforded by differentially private training as a confidence interval for the privacy budget $\hat{\varepsilon}$ spent with respect to specific threat models. Existing approaches derive confidence intervals for $\hat{\varepsilon}$ from confidence intervals for false positive and false negative rates of membership inference attacks, which requires training an impractically large number of models to get intervals that can be acted upon. We propose a novel, more efficient Bayesian approach that brings privacy estimates within the reach of practitioners. Our approach reduces sample size by computing a posterior for $\hat{\varepsilon}$ (not just a confidence interval) from the joint posterior of the false positive and false negative rates of membership inference attacks. We implement an end-to-end system for privacy estimation that integrates our approach and state-of-the-art membership inference attacks, and evaluate it on text and vision classification tasks. For the same number of samples, we see a reduction in interval width of up to 40% compared to prior work.

Model fairness is an essential element for Trustworthy AI. While many techniques for model fairness have been proposed, most of them assume that the training and deployment data distributions are identical, which is often not true in practice. In particular, when the bias between labels and sensitive groups changes, the fairness of the trained model is directly influenced and can worsen. We make two contributions for solving this problem. First, we analytically show that existing in-processing fair algorithms have fundamental limits in accuracy and group fairness. We utilize the notion of correlation shifts between labels and groups, which can explicitly capture the change of the above bias. Second, we propose a novel pre-processing step that samples the input data to reduce correlation shifts and thus enables the in-processing approaches to overcome their limitations. We formulate an optimization problem for adjusting the data ratio among labels and sensitive groups to reflect the shifted correlation. A key benefit of our approach lies in decoupling the roles of pre- and in-processing approaches: correlation adjustment via pre-processing and unfairness mitigation on the processed data via in-processing. Experiments show that our framework effectively improves existing in-processing fair algorithms w.r.t. accuracy and fairness, both on synthetic and real datasets.

Prior knowledge and symbolic rules in machine learning are often expressed in the form of label constraints, especially in structured prediction problems. In this work, we compare two common strategies for encoding label constraints in a machine learning pipeline, *regularization with constraints* and *constrained inference*, by quantifying their impact on model performance. For regularization, we show that it narrows the generalization gap by precluding models that are inconsistent with the constraints. However, its preference for small violations introduces a bias toward a suboptimal model. For constrained inference, we show that it reduces the population risk by correcting a model's violation, and hence turns the violation into an advantage. Given these differences, we further explore the use of two approaches together and propose conditions for constrained inference to compensate for the bias introduced by regularization, aiming to improve both the model complexity and optimal risk.

Recently, unsupervised representation learning (URL) has improved the sample efficiency of Reinforcement Learning (RL) by pretraining a model from a large unlabeled dataset. The underlying principle of these methods is to learn temporally predictive representations by predicting future states in the latent space. However, an important challenge of this approach is the representational collapse, where the subspace of the latent representations collapses into a low-dimensional manifold. To address this issue, we propose a novel URL framework that causally predicts future states while increasing the dimension of the latent manifold by decorrelating the features in the latent space. Through extensive empirical studies, we demonstrate that our framework effectively learns predictive representations without collapse, which significantly improves the sample efficiency of state-of-the-art URL methods on the Atari 100k benchmark. The code is available at https://github.com/dojeon-ai/SimTPR.

Learning heuristics for vehicle routing problems (VRPs) has gained much attention due to the less reliance on hand-crafted rules. However, existing methods are typically trained and tested on the same task with a fixed size and distribution (of nodes), and hence suffer from limited generalization performance. This paper studies a challenging yet realistic setting, which considers generalization across both size and distribution in VRPs. We propose a generic meta-learning framework, which enables effective training of an initialized model with the capability of fast adaptation to new tasks during inference. We further develop a simple yet efficient approximation method to reduce the training overhead. Extensive experiments on both synthetic and benchmark instances of the traveling salesman problem (TSP) and capacitated vehicle routing problem (CVRP) demonstrate the effectiveness of our method. The code is available at: https://github.com/RoyalSkye/Omni-VRP.

Test-time adaptation (TTA) adapts a source model to the distribution shift in testing data without using any source data. There have been plenty of algorithms concentrated on covariate shift in the last decade, i.e., $\mathcal{D}_t(X)$, the distribution of the test data is different from the source data. Nonetheless, in real application scenarios, it is necessary to consider the influence of label distribution shift, i.e., both $\mathcal{D}_t(X)$ and $\mathcal{D}_t(Y)$ are shifted, which has not been sufficiently explored yet. To remedy this, we study a new problem setup, namely, TTA with Open-world Data Shift (AODS). The goal of AODS is simultaneously adapting a model to covariate and label distribution shifts in the test phase. In this paper, we first analyze the relationship between classification error and distribution shifts. Motivated by this, we hence propose a new framework, namely ODS, which decouples the mixed distribution shift and then addresses covariate and label distribution shifts accordingly. We conduct experiments on multiple benchmarks with different types of shifts, and the results demonstrate the superior performance of our method against the state of the arts. Moreover, ODS is suitable for many TTA algorithms.

Driven by the need from real-world applications, Auxiliary Modality Learning (AML) offers the possibility to utilize more information from auxiliary data in training, while only requiring data from one or fewer modalities in test, to save the overall computational cost and reduce the amount of input data for inferencing. In this work, we formally define ``Auxiliary Modality Learning'' (AML), systematically classify types of auxiliary modality (in visual computing) and architectures for AML, and analyze their performance. We also analyze the conditions under which AML works well from the optimization and data distribution perspectives. To guide various choices to achieve optimal performance using AML, we propose a novel method to assist in choosing the best auxiliary modality and estimating an upper bound performance before executing AML. In addition, we propose a new AML method using generalized curriculum distillation to enable more effective curriculum learning. Our method achieves the best performance compared to other SOTA methods.

How can agents learn internal models that veridically represent interactions with the real world is a largely open question. As machine learning is moving towards representations containing not just observational but also interventional knowledge, we study this problem using tools from representation learning and group theory. We propose methods enabling an agent acting upon the world to learn internal representations of sensory information that are consistent with actions that modify it. We use an autoencoder equipped with a group representation acting on its latent space, trained using an equivariance-derived loss in order to enforce a suitable homomorphism property on the group representation. In contrast to existing work, our approach does not require prior knowledge of the group and does not restrict the set of actions the agent can perform. We motivate our method theoretically, and show empirically that it can learn a group representation of the actions, thereby capturing the structure of the set of transformations applied to the environment. We further show that this allows agents to predict the effect of sequences of future actions with improved accuracy.

In machine learning applications, it is well known that carefully designed learning rate (step size) schedules can significantly improve the convergence of commonly used first-order optimization algorithms. Therefore how to set step size adaptively becomes an important research question. A popular and effective method is the Polyak step size, which sets step size adaptively for gradient descent or stochastic gradient descent without the need to estimate the smoothness parameter of the objective function. However, there has not been a principled way to generalize the Polyak step size for algorithms with momentum accelerations. This paper presents a general framework to set the learning rate adaptively for first-order optimization methods with momentum, motivated by the derivation of Polyak step size. It is shown that the resulting techniques are much less sensitive to the choice of momentum parameter and may avoid the oscillation of the heavy-ball method on ill-conditioned problems. These adaptive step sizes are further extended to the stochastic settings, which are attractive choices for stochastic gradient descent with momentum. Our methods are demonstrated to be more effective for stochastic gradient methods than prior adaptive step size algorithms in large-scale machine learning tasks.

The use of deep neural networks for modelling system dynamics is increasingly popular, but long-term prediction accuracy and out-of-distribution generalization still present challenges. In this study, we address these challenges by considering the parameters of dynamical systems as factors of variation of the data and leverage their ground-truth values to disentangle the representations learned by generative models. Our experimental results in phase-space and observation-space dynamics, demonstrate the effectiveness of latent-space supervision in producing disentangled representations, leading to improved long-term prediction accuracy and out-of-distribution robustness.

Prediction of a $\textbf{varying number}$ of $\textbf{ordered clusters}$ from sets of $\textbf{any cardinality}$ is a challenging task for neural networks, combining elements of set representation, clustering and learning to order. This task arises in many diverse areas, ranging from medical triage and early discharge, through machine part management and multi-channel signal analysis for petroleum exploration to product catalog structure prediction. This paper focuses on that last area, which exemplifies a number of challenges inherent to adaptive ordered clustering, referred to further as the eponymous $\textit{Catalog Problem}$. These include learning variable cluster constraints, exhibiting relational reasoning and managing combinatorial complexity. Despite progress in both neural clustering and set-to-sequence methods, no joint, fully differentiable model exists to-date. We develop such a modular architecture, referred to further as Neural Ordered Clusters (NOC), enhance it with a specific mechanism for learning cluster-level cardinality constraints, and provide a robust comparison of its performance in relation to alternative models. We test our method on three datasets, including synthetic catalog structures and PROCAT, a dataset of real-world catalogs consisting of over 1.5M products, achieving state-of-the-art results on a new, more challenging formulation of the underlying problem, which has not been addressed before. Additionally, we examine the network's ability to learn higher-order interactions.

Recent efforts at explaining the interplay of memorization and generalization in deep overparametrized networks have posited that neural networks *memorize* ``hard'' examples in the final few layers of the model. Memorization refers to the ability to correctly predict on *atypical* examples of the training set. In this work, we show that rather than being confined to individual layers, memorization is a phenomenon confined to a small set of neurons in various layers of the model. First, via three experimental sources of converging evidence, we find that most layers are redundant for the memorization of examples and the layers that contribute to example memorization are, in general, not the final layers. The three sources are *gradient accounting* (measuring the contribution to the gradient norms from memorized and clean examples), *layer rewinding* (replacing specific model weights of a converged model with previous training checkpoints), and *retraining* (training rewound layers only on clean examples). Second, we ask a more generic question: can memorization be localized *anywhere* in a model? We discover that memorization is often confined to a small number of neurons or channels (around 5) of the model. Based on these insights we propose a new form of dropout---*example-tied dropout* that enables us to direct the memorization of examples to an aprior determined set of neurons. By dropping out these neurons, we are able to reduce the accuracy on memorized examples from 100% to 3%, while also reducing the generalization gap.

Text-to-image synthesis has recently seen significant progress thanks to large pretrained language models, large-scale training data, and the introduction of scalable model families such as diffusion and autoregressive models. However, the best-performing models require iterative evaluation to generate a single sample. In contrast, generative adversarial networks (GANs) only need a single forward pass. They are thus much faster, but they currently remain far behind the state-of-the-art in large-scale text-to-image synthesis. This paper aims to identify the necessary steps to regain competitiveness. Our proposed model, StyleGAN-T, addresses the specific requirements of large-scale text-to-image synthesis, such as large capacity, stable training on diverse datasets, strong text alignment, and controllable variation vs. text alignment tradeoff. StyleGAN-T significantly improves over previous GANs and outperforms distilled diffusion models - the previous state-of-the-art in fast text-to-image synthesis - in terms of sample quality and speed.

A prominent paradigm for graph neural networks is based on the message-passing framework. In this framework, information communication is realized only between neighboring nodes. The challenge of approaches that use this paradigm is to ensure efficient and accurate long-distance communication between nodes, as deep convolutional networks are prone to over smoothing. In this paper, we present a novel method based on time derivative graph diffusion (TIDE) to overcome these structural limitations of the message-passing framework. Our approach allows for optimizing the spatial extent of diffusion across various tasks and network channels, thus enabling medium and long-distance communication efficiently. Furthermore, we show that our architecture design also enables local message-passing and thus inherits from the capabilities of local message-passing approaches. We show that on both widely used graph benchmarks and synthetic mesh and graph datasets, the proposed framework outperforms state-of-the-art methods by a significant margin.

With infinitely many high-quality data points, infinite computational power, an infinitely large foundation model with a perfect training algorithm and guaranteed zero generalization error on the pretext task, can the model be used for everything? This question cannot be answered by the existing theory of representation, optimization or generalization, because the issues they mainly investigate are assumed to be nonexistent here. In this paper, we show that category theory provides powerful machinery to answer this question. We have proved three results. The first one limits the power of prompt-based learning, saying that the model can solve a downstream task with prompts if and only if the task is representable. The second one says fine tuning does not have this limit, as a foundation model with the minimum required power (up to symmetry) can theoretically solve downstream tasks for the category defined by pretext task, with fine tuning and enough resources. Our final result can be seen as a new type of generalization theorem, showing that the foundation model can generate unseen objects from the target category (e.g., images) using the structural information from the source category (e.g., texts). Along the way, we provide a categorical framework for supervised and self-supervised learning, which might be of independent interest.