ICML 2022 论文列表

We study the problem of identifying cause and effect over two univariate continuous variables $X$ and $Y$ from a sample of their joint distribution. Our focus lies on the setting when the variance of the noise may be dependent on the cause. We propose to partition the domain of the cause into multiple segments where the noise indeed is dependent. To this end, we minimize a scale-invariant, penalized regression score, finding the optimal partitioning using dynamic programming. We show under which conditions this allows us to identify the causal direction for the linear setting with heteroscedastic noise, for the non-linear setting with homoscedastic noise, as well as empirically confirm that these results generalize to the non-linear and heteroscedastic case. Altogether, the ability to model heteroscedasticity translates into an improved performance in telling cause from effect on a wide range of synthetic and real-world datasets.

The inductive matrix completion (IMC) problem is to recover a low rank matrix from few observed entries while incorporating prior knowledge about its row and column subspaces. In this work, we make three contributions to the IMC problem: (i) we prove that under suitable conditions, the IMC optimization landscape has no bad local minima; (ii) we derive a simple scheme with theoretical guarantees to estimate the rank of the unknown matrix; and (iii) we propose GNIMC, a simple Gauss-Newton based method to solve the IMC problem, analyze its runtime and derive for it strong recovery guarantees. The guarantees for GNIMC are sharper in several aspects than those available for other methods, including a quadratic convergence rate, fewer required observed entries and stability to errors or deviations from low-rank. Empirically, given entries observed uniformly at random, GNIMC recovers the underlying matrix substantially faster than several competing methods.

Self-attention, an architectural motif designed to model long-range interactions in sequential data, has driven numerous recent breakthroughs in natural language processing and beyond. This work provides a theoretical analysis of the inductive biases of self-attention modules. Our focus is to rigorously establish which functions and long-range dependencies self-attention blocks prefer to represent. Our main result shows that bounded-norm Transformer networks "create sparse variables": a single self-attention head can represent a sparse function of the input sequence, with sample complexity scaling only logarithmically with the context length. To support our analysis, we present synthetic experiments to probe the sample complexity of learning sparse Boolean functions with Transformers.

In many areas, we have well-founded insights about causal structure that would be useful to bring into our trained models while still allowing them to learn in a data-driven fashion. To achieve this, we present the new method of interchange intervention training (IIT). In IIT, we (1) align variables in a causal model (e.g., a deterministic program or Bayesian network) with representations in a neural model and (2) train the neural model to match the counterfactual behavior of the causal model on a base input when aligned representations in both models are set to be the value they would be for a source input. IIT is fully differentiable, flexibly combines with other objectives, and guarantees that the target causal model is a causal abstraction of the neural model when its loss is zero. We evaluate IIT on a structural vision task (MNIST-PVR), a navigational language task (ReaSCAN), and a natural language inference task (MQNLI). We compare IIT against multi-task training objectives and data augmentation. In all our experiments, IIT achieves the best results and produces neural models that are more interpretable in the sense that they more successfully realize the target causal model.

In many real-world multi-agent systems, the sparsity of team rewards often makes it difficult for an algorithm to successfully learn a cooperative team policy. At present, the common way for solving this problem is to design some dense individual rewards for the agents to guide the cooperation. However, most existing works utilize individual rewards in ways that do not always promote teamwork and sometimes are even counterproductive. In this paper, we propose

Transformer-based models such as BERT have proven successful in information retrieval problem, which seek to identify relevant documents for a given query. There are two broad flavours of such models: cross-attention (CA) models, which learn a joint embedding for the query and document, and dual-encoder (DE) models, which learn separate embeddings for the query and document. Empirically, CA models are often found to be more accurate, which has motivated a series of works seeking to bridge this gap. However, a more fundamental question remains less explored: does this performance gap reflect an inherent limitation in the capacity of DE models, or a limitation in the training of such models? And does such an understanding suggest a principled means of improving DE models? In this paper, we study these questions, with three contributions. First, we establish theoretically that with a sufficiently large embedding dimension, DE models have the capacity to model a broad class of score distributions. Second, we show empirically that on real-world problems, DE models may overfit to spurious correlations in the training set, and thus under-perform on test samples. To mitigate this behaviour, we propose a suitable distillation strategy, and confirm its practical efficacy on the MSMARCO-Passage and Natural Questions benchmarks.

Multi-head attention is a driving force behind state-of-the-art transformers, which achieve remarkable performance across a variety of natural language processing (NLP) and computer vision tasks. It has been observed that for many applications, those attention heads learn redundant embedding, and most of them can be removed without degrading the performance of the model. Inspired by this observation, we propose Transformer with a Mixture of Gaussian Keys (Transformer-MGK), a novel transformer architecture that replaces redundant heads in transformers with a mixture of keys at each head. These mixtures of keys follow a Gaussian mixture model and allow each attention head to focus on different parts of the input sequence efficiently. Compared to its conventional transformer counterpart, Transformer-MGK accelerates training and inference, has fewer parameters, and requires fewer FLOPs to compute while achieving comparable or better accuracy across tasks. Transformer-MGK can also be easily extended to use with linear attention. We empirically demonstrate the advantage of Transformer-MGK in a range of practical applications, including language modeling and tasks that involve very long sequences. On the Wikitext-103 and Long Range Arena benchmark, Transformer-MGKs with 4 heads attain comparable or better performance to the baseline transformers with 8 heads.

Task-free continual learning (CL) aims to learn a non-stationary data stream without explicit task definitions and not forget previous knowledge. The widely adopted memory replay approach could gradually become less effective for long data streams, as the model may memorize the stored examples and overfit the memory buffer. Second, existing methods overlook the high uncertainty in the memory data distribution since there is a big gap between the memory data distribution and the distribution of all the previous data examples. To address these problems, for the first time, we propose a principled memory evolution framework to dynamically evolve the memory data distribution by making the memory buffer gradually harder to be memorized with distributionally robust optimization (DRO). We then derive a family of methods to evolve the memory buffer data in the continuous probability measure space with Wasserstein gradient flow (WGF). The proposed DRO is w.r.t the worst-case evolved memory data distribution, thus guarantees the model performance and learns significantly more robust features than existing memory-replay-based methods. Extensive experiments on existing benchmarks demonstrate the effectiveness of the proposed methods for alleviating forgetting. As a by-product of the proposed framework, our method is more robust to adversarial examples than existing task-free CL methods.

Many selection processes such as finding patients qualifying for a medical trial or retrieval pipelines in search engines consist of multiple stages, where an initial screening stage focuses the resources on shortlisting the most promising candidates. In this paper, we investigate what guarantees a screening classifier can provide, independently of whether it is constructed manually or trained. We find that current solutions do not enjoy distribution-free theoretical guarantees and we show that, in general, even for a perfectly calibrated classifier, there always exist specific pools of candidates for which its shortlist is suboptimal. Then, we develop a distribution-free screening algorithm—called Calibrated Subsect Selection (CSS)—that, given any classifier and some amount of calibration data, finds near-optimal shortlists of candidates that contain a desired number of qualified candidates in expectation. Moreover, we show that a variant of CSS that calibrates a given classifier multiple times across specific groups can create shortlists with provable diversity guarantees. Experiments on US Census survey data validate our theoretical results and show that the shortlists provided by our algorithm are superior to those provided by several competitive baselines.

The reliability of neural networks is essential for their use in safety-critical applications. Existing approaches generally aim at improving the robustness of neural networks to either real-world distribution shifts (e.g., common corruptions and perturbations, spatial transformations, and natural adversarial examples) or worst-case distribution shifts (e.g., optimized adversarial examples). In this work, we propose the Decision Region Quantification (DRQ) algorithm to improve the robustness of any differentiable pre-trained model against both real-world and worst-case distribution shifts in the data. DRQ analyzes the robustness of local decision regions in the vicinity of a given data point to make more reliable predictions. We theoretically motivate the DRQ algorithm by showing that it effectively smooths spurious local extrema in the decision surface. Furthermore, we propose an implementation using targeted and untargeted adversarial attacks. An extensive empirical evaluation shows that DRQ increases the robustness of adversarially and non-adversarially trained models against real-world and worst-case distribution shifts on several computer vision benchmark datasets.

Generalization in deep reinforcement learning over unseen environment variations usually requires policy learning over a large set of diverse training variations. We empirically observe that an agent trained on many variations (a generalist) tends to learn faster at the beginning, yet its performance plateaus at a less optimal level for a long time. In contrast, an agent trained only on a few variations (a specialist) can often achieve high returns under a limited computational budget. To have the best of both worlds, we propose a novel generalist-specialist training framework. Specifically, we first train a generalist on all environment variations; when it fails to improve, we launch a large population of specialists with weights cloned from the generalist, each trained to master a selected small subset of variations. We finally resume the training of the generalist with auxiliary rewards induced by demonstrations of all specialists. In particular, we investigate the timing to start specialist training and compare strategies to learn generalists with assistance from specialists. We show that this framework pushes the envelope of policy learning on several challenging and popular benchmarks including Procgen, Meta-World and ManiSkill.

Machine learning algorithms typically assume that training and test examples are drawn from the same distribution. However, distribution shift is a common problem in real-world applications and can cause models to perform dramatically worse at test time. In this paper, we specifically consider the problems of subpopulation shifts (e.g., imbalanced data) and domain shifts. While prior works often seek to explicitly regularize internal representations or predictors of the model to be domain invariant, we instead aim to learn invariant predictors without restricting the model’s internal representations or predictors. This leads to a simple mixup-based technique which learns invariant predictors via selective augmentation called LISA. LISA selectively interpolates samples either with the same labels but different domains or with the same domain but different labels. Empirically, we study the effectiveness of LISA on nine benchmarks ranging from subpopulation shifts to domain shifts, and we find that LISA consistently outperforms other state-of-the-art methods and leads to more invariant predictors. We further analyze a linear setting and theoretically show how LISA leads to a smaller worst-group error.

Mini-batch optimal transport (m-OT) has been widely used recently to deal with the memory issue of OT in large-scale applications. Despite their practicality, m-OT suffers from misspecified mappings, namely, mappings that are optimal on the mini-batch level but are partially wrong in the comparison with the optimal transportation plan between the original measures. Motivated by the misspecified mappings issue, we propose a novel mini-batch method by using partial optimal transport (POT) between mini-batch empirical measures, which we refer to as mini-batch partial optimal transport (m-POT). Leveraging the insight from the partial transportation, we explain the source of misspecified mappings from the m-OT and motivate why limiting the amount of transported masses among mini-batches via POT can alleviate the incorrect mappings. Finally, we carry out extensive experiments on various applications such as deep domain adaptation, partial domain adaptation, deep generative model, color transfer, and gradient flow to demonstrate the favorable performance of m-POT compared to current mini-batch methods.

We enhance auto-regressive language models by conditioning on document chunks retrieved from a large corpus, based on local similarity with preceding tokens. With a 2 trillion token database, our Retrieval-Enhanced Transformer (RETRO) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25{\texttimes} fewer parameters. After fine-tuning, RETRO performance translates to downstream knowledge-intensive tasks such as question answering. RETRO combines a frozen Bert retriever, a differentiable encoder and a chunked cross-attention mechanism to predict tokens based on an order of magnitude more data than what is typically consumed during training. We typically train RETRO from scratch, yet can also rapidly RETROfit pre-trained transformers with retrieval and still achieve good performance. Our work opens up new avenues for improving language models through explicit memory at unprecedented scale.

Ensembles of deep neural networks have demonstrated superior performance, but their heavy computational cost hinders applying them for resource-limited environments. It motivates distilling knowledge from the ensemble teacher into a smaller student network, and there are two important design choices for this ensemble distillation: 1) how to construct the student network, and 2) what data should be shown during training. In this paper, we propose a weight averaging technique where a student with multiple subnetworks is trained to absorb the functional diversity of ensemble teachers, but then those subnetworks are properly averaged for inference, giving a single student network with no additional inference cost. We also propose a perturbation strategy that seeks inputs from which the diversities of teachers can be better transferred to the student. Combining these two, our method significantly improves upon previous methods on various image classification tasks.

Deep neural networks (DNNs) are found to be vulnerable to adversarial noise. They are typically misled by adversarial samples to make wrong predictions. To alleviate this negative effect, in this paper, we investigate the dependence between outputs of the target model and input adversarial samples from the perspective of information theory, and propose an adversarial defense method. Specifically, we first measure the dependence by estimating the mutual information (MI) between outputs and the natural patterns of inputs (called natural MI) and MI between outputs and the adversarial patterns of inputs (called adversarial MI), respectively. We find that adversarial samples usually have larger adversarial MI and smaller natural MI compared with those w.r.t. natural samples. Motivated by this observation, we propose to enhance the adversarial robustness by maximizing the natural MI and minimizing the adversarial MI during the training process. In this way, the target model is expected to pay more attention to the natural pattern that contains objective semantics. Empirical evaluations demonstrate that our method could effectively improve the adversarial accuracy against multiple attacks.

Inverting an image onto the latent space of pre-trained generators, e.g., StyleGAN-v2, has emerged as a popular strategy to leverage strong image priors for ill-posed restoration. Several studies have showed that this approach is effective at inverting images similar to the data used for training. However, with out-of-distribution (OOD) data that the generator has not been exposed to, existing inversion techniques produce sub-optimal results. In this paper, we propose SPHInX (StyleGAN with Projection Heads for Inverting X), an approach for accurately embedding OOD images onto the StyleGAN latent space. SPHInX optimizes a style projection head using a novel training strategy that imposes a vicinal regularization in the StyleGAN latent space. To further enhance OOD inversion, SPHInX can additionally optimize a content projection head and noise variables in every layer. Our empirical studies on a suite of OOD data show that, in addition to producing higher quality reconstructions over the state-of-the-art inversion techniques, SPHInX is effective for ill-posed restoration tasks while offering semantic editing capabilities.

We study privacy-preserving exploration in sequential decision-making for environments that rely on sensitive data such as medical records. In particular, we focus on solving the problem of reinforcement learning (RL) subject to the constraint of (joint) differential privacy in the linear MDP setting, where both dynamics and rewards are given by linear functions. Prior work on this problem due to (Luyo et al., 2021) achieves a regret rate that has a dependence of O(K^{3/5}) on the number of episodes K. We provide a private algorithm with an improved regret rate with an optimal dependence of O($\sqrt{}$K) on the number of episodes. The key recipe for our stronger regret guarantee is the adaptivity in the policy update schedule, in which an update only occurs when sufficient changes in the data are detected. As a result, our algorithm benefits from low switching cost and only performs O(log(K)) updates, which greatly reduces the amount of privacy noise. Finally, in the most prevalent privacy regimes where the privacy parameter \epsilon is a constant, our algorithm incurs negligible privacy cost{—}in comparison with the existing non-private regret bounds, the additional regret due to privacy appears in lower-order terms.

Kernel-based models such as kernel ridge regression and Gaussian processes are ubiquitous in machine learning applications for regression and optimization. It is well known that a major downside for kernel-based models is the high computational cost; given a dataset of $n$ samples, the cost grows as $\mathcal{O}(n^3)$. Existing sparse approximation methods can yield a significant reduction in the computational cost, effectively reducing the actual cost down to as low as $\mathcal{O}(n)$ in certain cases. Despite this remarkable empirical success, significant gaps remain in the existing results for the analytical bounds on the error due to approximation. In this work, we provide novel confidence intervals for the Nyström method and the sparse variational Gaussian process approximation method, which we establish using novel interpretations of the approximate (surrogate) posterior variance of the models. Our confidence intervals lead to improved performance bounds in both regression and optimization problems.