ICLR 2023 论文列表

We consider two biologically plausible structures, the Spiking Neural Network (SNN) and the self-attention mechanism. The former offers an energy-efficient and event-driven paradigm for deep learning, while the latter has the ability to capture feature dependencies, enabling Transformer to achieve good performance. It is intuitively promising to explore the marriage between them. In this paper, we consider leveraging both self-attention capability and biological properties of SNNs, and propose a novel Spiking Self Attention (SSA) as well as a powerful framework, named Spiking Transformer (Spikformer). The SSA mechanism in Spikformer models the sparse visual feature by using spike-form Query, Key, and Value without softmax. Since its computation is sparse and avoids multiplication, SSA is efficient and has low computational energy consumption. It is shown that Spikformer with SSA can outperform the state-of-the-art SNNs-like frameworks in image classification on both neuromorphic and static datasets. Spikformer (66.3M parameters) with comparable size to SEW-ResNet-152 (60.2M,69.26%) can achieve 74.81% top1 accuracy on ImageNet using 4 time steps, which is the state-of-the-art in directly trained SNNs models. Code is avaiable at https://github.com/ZK-Zhou/spikformer.

Gaussian processes (GPs) are an attractive class of machine learning models because of their simplicity and flexibility as building blocks of more complex Bayesian models. Meanwhile, graph neural networks (GNNs) emerged recently as a promising class of models for graph-structured data in semi-supervised learning and beyond. Their competitive performance is often attributed to a proper capturing of the graph inductive bias. In this work, we introduce this inductive bias into GPs to improve their predictive performance for graph-structured data. We show that a prominent example of GNNs, the graph convolutional network, is equivalent to some GP when its layers are infinitely wide; and we analyze the kernel universality and the limiting behavior in depth. We further present a programmable procedure to compose covariance kernels inspired by this equivalence and derive example kernels corresponding to several interesting members of the GNN family. We also propose a computationally efficient approximation of the covariance matrix for scalable posterior inference with large-scale data. We demonstrate that these graph-based kernels lead to competitive classification and regression performance, as well as advantages in computation time, compared with the respective GNNs.

The domain generalization (DG) setup considers the problem where models are trained on data sampled from multiple domains and evaluated on test domains unseen during training. In this paper, we formulate DG as a sample selection problem where each domain is sampled from a common underlying population through non-random sampling probabilities that correlate with both the features and the outcome. Under this setting, the fundamental iid assumption of the empirical risk minimization (ERM) is violated, so it often performs worse on test domains whose non-random sampling probabilities differ from the domains in the training dataset. We propose a Selection-Guided DG (SGDG) framework to learn the selection probability of each domain and the joint distribution of the outcome and domain selection variables. The proposed SGDG is domain generalizable as it intends to minimize the risk under the population distribution. We theoretically proved that, under certain regular conditions, SGDG can achieve smaller risk than ERM. Furthermore, we present a class of parametric SGDG (HeckmanDG) estimators applicable to continuous, binary, and multinomial outcomes. We also demonstrated its efficacy empirically through simulations and experiments on a set of benchmark datasets comparing with other well-known DG methods.

Lottery Ticket Hypothesis (LTH) claims the existence of a winning ticket (i.e., a properly pruned sub-network together with original weight initialization) that can achieve competitive performance to the original dense network. A recent work, called UGS, extended LTH to prune graph neural networks (GNNs) for effectively accelerating GNN inference. UGS simultaneously prunes the graph adjacency matrix and the model weights using the same masking mechanism, but since the roles of the graph adjacency matrix and the weight matrices are very different, we find that their sparsifications lead to different performance characteristics. Specifically, we find that the performance of a sparsified GNN degrades significantly when the graph sparsity goes beyond a certain extent. Therefore, we propose two techniques to improve GNN performance when the graph sparsity is high. First, UGS prunes the adjacency matrix using a loss formulation which, however, does not properly involve all elements of the adjacency matrix; in contrast, we add a new auxiliary loss head to better guide the edge pruning by involving the entire adjacency matrix. Second, by regarding unfavorable graph sparsification as adversarial data perturbations, we formulate the pruning process as a min-max optimization problem to gain the robustness of lottery tickets when the graph sparsity is high. We further investigate the question: Can the ``retrainable'' winning ticket of a GNN be also effective for graph transferring learning? We call it the transferable graph lottery ticket (GLT) hypothesis. Extensive experiments were conducted which demonstrate the superiority of our proposed sparsification method over UGS, and which empirically verified our transferable GLT hypothesis.

In this paper, we systematically study the audio-visual speech separation task in a multi-speaker scenario. Given the facial information of each speaker, the goal of this task is to separate the corresponding speech from the mixed speech. The existing works are designed for speech separation in a controlled setting with a fixed number of speakers (mostly 2 or 3 speakers), which seems to be impractical for real applications. As a result, we try to utilize a single model to separate the voices with a variable number of speakers. Based on the observation, there are two prominent issues for multi-speaker separation: 1) There are some noisy voice pieces belonging to other speakers in the separation results; 2) Part of the target speech is missing after separation. Accordingly, we propose \textbf{BFRNet}, including a {\bf B}asic audio-visual speech separator and a Filter-Recovery Network (\textbf{FRNet}). FRNet can refine the coarse audio separated by basic audio-visual speech separator. To have fair comparisons, we build a comprehensive benchmark for multi-speaker audio-visual speech separation to verify the performance of various methods. Experimental results show that our method is able to achieve the state-of-the-art performance. Furthermore, we also find that FRNet can boost the performance of other off-the-shelf speech separators, which exhibits its ability of generalization.

With the adoption of recent laws ensuring the ``right to be forgotten'', the problem of machine unlearning has become of significant importance. This is particularly the case for graph-structured data, and learning tools specialized for such data, including graph neural networks (GNNs). This work introduces the first known approach for \emph{approximate graph unlearning} with provable theoretical guarantees. The challenges in addressing the problem are two-fold. First, there exist multiple different types of unlearning requests that need to be considered, including node feature, edge and node unlearning. Second, to establish provable performance guarantees, one needs to carefully evaluate the process of feature mixing during propagation. We focus on analyzing Simple Graph Convolutions (SGC) and their generalized PageRank (GPR) extensions, thereby laying the theoretical foundations for unlearning GNNs. Empirical evaluations of six benchmark datasets demonstrate excellent performance/complexity/privacy trade-offs of our approach compared to complete retraining and general methods that do not leverage graph information. For example, unlearning $200$ out of $1208$ training nodes of the Cora dataset only leads to a $0.1\%$ loss in test accuracy, but offers a $4$-fold speed-up compared to complete retraining with a $(\epsilon,\delta)=(1,10^{-4})$ ``privacy cost''. We also exhibit a $12\%$ increase in test accuracy for the same dataset when compared to unlearning methods that do not leverage graph information, with comparable time complexity and the same privacy guarantee.

In-Context Learning (ICL), which formulates target tasks as prompt completion conditioned on in-context demonstrations, has become the prevailing utilization of LLMs. In this paper, we first disclose an actual predicament for this typical usage that it can not scale up with training data due to context length restriction. Besides, existing works have shown that ICL also suffers from various biases and requires delicate calibration treatment. To address both challenges, we advocate a simple and effective solution, $k$NN Prompting, which first queries LLM with training data for distributed representations, then predicts test instances by simply referring to nearest neighbors. We conduct comprehensive experiments to demonstrate its two-fold superiority: 1) Calibration-Free: $k$NN Prompting does not directly align LLM output distribution with task-specific label space, instead leverages such distribution to align test and training instances. It significantly outperforms state-of-the-art calibration-based methods under comparable few-shot scenario. 2) Beyond-Context: $k$NN Prompting can further scale up effectively with as many training data as are available, continually bringing substantial improvements. The scaling trend holds across 10 orders of magnitude ranging from 2 shots to 1024 shots as well as different LLMs scales ranging from 0.8B to 30B. It successfully bridges data scaling into model scaling, and brings new potentials for the gradient-free paradigm of LLM deployment. Code is publicly available at https://github.com/BenfengXu/KNNPrompting

The classical Perceptron algorithm of Rosenblatt can be used to find a linear threshold function to correctly classify $n$ linearly separable data points, assuming the classes are separated by some margin $\gamma > 0$. A foundational result is that Perceptron converges after $\Omega(1/\gamma^{2})$ iterations. There have been several recent works that managed to improve this rate by a quadratic factor, to $\Omega(\sqrt{\log n}/\gamma)$, with more sophisticated algorithms. In this paper, we unify these existing results under one framework by showing that they can all be described through the lens of solving min-max problems using modern acceleration techniques, mainly through \emph{optimistic} online learning. We then show that the proposed framework also leads to improved results for a series of problems beyond the standard Perceptron setting. Specifically, a) for the margin maximization problem, we improve the state-of-the-art result from $O(\log t/t^2)$ to $O(1/t^2)$, where $t$ is the number of iterations; b) we provide the first result on identifying the implicit bias property of the classical Nesterov's accelerated gradient descent (NAG) algorithm, and show NAG can maximize the margin with an $O(1/t^2)$ rate; c) for the classical $p$-norm Perceptron problem, we provide an algorithm with $\Omega(\sqrt{(p-1)\log n}/\gamma)$ convergence rate, while existing algorithms suffer the $\Omega({(p-1)}/\gamma^2)$ convergence rate.

Pre-training is prevalent in nowadays deep learning to improve the learned model's performance. However, in the literature on federated learning (FL), neural networks are mostly initialized with random weights. These attract our interest in conducting a systematic study to explore pre-training for FL. Across multiple visual recognition benchmarks, we found that pre-training can not only improve FL, but also close its accuracy gap to the counterpart centralized learning, especially in the challenging cases of non-IID clients' data. To make our findings applicable to situations where pre-trained models are not directly available, we explore pre-training with synthetic data or even with clients' data in a decentralized manner, and found that they can already improve FL notably. Interestingly, many of the techniques we explore are complementary to each other to further boost the performance, and we view this as a critical result toward scaling up deep FL for real-world applications. We conclude our paper with an attempt to understand the effect of pre-training on FL. We found that pre-training enables the learned global models under different clients' data conditions to converge to the same loss basin, and makes global aggregation in FL more stable. Nevertheless, pre-training seems to not alleviate local model drifting, a fundamental problem in FL under non-IID data.

We first elucidate various fundamental properties of optimal adversarial predictors: the structure of optimal adversarial convex predictors in terms of optimal adversarial zero-one predictors, bounds relating the adversarial convex loss to the adversarial zero-one loss, and the fact that continuous predictors can get arbitrarily close to the optimal adversarial error for both convex and zero-one losses. Applying these results along with new Rademacher complexity bounds for adversarial training near initialization, we prove that for general data distributions and perturbation sets, adversarial training on shallow networks with early stopping and an idealized optimal adversary is able to achieve optimal adversarial test error. By contrast, prior theoretical work either considered specialized data distributions or only provided training error guarantees.

Optimizing multiple competing objectives is a common problem across science and industry. The inherent inextricable trade-off between those objectives leads one to the task of exploring their Pareto front. A meaningful quantity for the purpose of the latter is the hypervolume indicator, which is used in Bayesian Optimization (BO) and Evolutionary Algorithms (EAs). However, the computational complexity for the calculation of the hypervolume scales unfavorably with increasing number of objectives and data points, which restricts its use in those common multi-objective optimization frameworks. To overcome these restrictions, previous work has focused on approximating the hypervolume using deep learning. In this work, we propose a novel deep learning architecture to approximate the hypervolume function, which we call DeepHV. For better sample efficiency and generalization, we exploit the fact that the hypervolume is scale equivariant in each of the objectives as well as permutation invariant w.r.t. both the objectives and the samples, by using a deep neural network that is equivariant w.r.t. the combined group of scalings and permutations. We show through an ablation study that including these symmetries leads to significantly improved model accuracy. We evaluate our method against exact, and approximate hypervolume methods in terms of accuracy, computation time, and generalization. We also apply and compare our methods to state-of-the-art multi-objective BO methods and EAs on a range of synthetic and real-world benchmark test cases. The results show that our methods are promising for such multi-objective optimization tasks.

We evaluate the reasoning abilities of large language models in multilingual settings. We introduce the Multilingual Grade School Math (MGSM) benchmark, by manually translating 250 grade-school math problems from the GSM8K dataset (Cobbe et al., 2021) into ten typologically diverse languages. We find that the ability to solve MGSM problems via chain-of-thought prompting emerges with increasing model scale, and that models have strikingly strong multilingual reasoning abilities, even in underrepresented languages such as Bengali and Swahili. Finally, we show that multilingual reasoning abilities of language models extend to other tasks such as commonsense reasoning and word-in-context semantic judgment. The MGSM benchmark is publicly available at AnonymousLink and the supplementary material.

Deep generative models learn highly complex and non-linear representations to generate realistic synthetic data. While they have achieved notable success in computer vision and natural language processing, similar advances have been less demonstrable in the tabular domain. This is partially because generative modelling of tabular data entails a particular set of challenges, including heterogeneous relationships, limited number of samples, and difficulties in incorporating prior knowledge. Additionally, unlike their counterparts in image and sequence domain, deep generative models for tabular data almost exclusively employ fully-connected layers, which encode weak inductive biases about relationships between inputs. Real-world data generating processes can often be represented using relational structures, which encode sparse, heterogeneous relationships between variables. In this work, we learn and exploit relational structure underlying tabular data to better model variable dependence, and as a natural means to introduce regularization on relationships and include prior knowledge. Specifically, we introduce GOGGLE, an end-to-end message passing scheme that jointly learns the relational structure and corresponding functional relationships as the basis of generating synthetic samples. Using real-world datasets, we provide empirical evidence that the proposed method is effective in generating realistic synthetic data and exploiting domain knowledge for downstream tasks.

Object rearrangement is a challenge for embodied agents because solving these tasks requires generalizing across a combinatorially large set of configurations of entities and their locations. Worse, the representations of these entities are unknown and must be inferred from sensory percepts. We present a hierarchical abstraction approach to uncover these underlying entities and achieve combinatorial generalization from unstructured visual inputs. By constructing a factorized transition graph over clusters of entity representations inferred from pixels, we show how to learn a correspondence between intervening on states of entities in the agent's model and acting on objects in the environment. We use this correspondence to develop a method for control that generalizes to different numbers and configurations of objects, which outperforms current offline deep RL methods when evaluated on simulated rearrangement tasks.

In this work, we attempt to bridge the two fields of finite-agent and infinite-agent games, by studying how the optimal policies of agents evolve with the number of agents (population size) in mean-field games, an agent-centric perspective in contrast to the existing works focusing typically on the convergence of the empirical distribution of the population. To this end, the premise is to obtain the optimal policies of a set of finite-agent games with different population sizes. However, either deriving the closed-form solution for each game is theoretically intractable, training a distinct policy for each game is computationally intensive, or directly applying the policy trained in a game to other games is sub-optimal. We address these challenges through the \textbf{P}opulation-size-\textbf{A}ware \textbf{P}olicy \textbf{O}ptimization (PAPO). Our contributions are three-fold. First, to efficiently generate efficient policies for games with different population sizes, we propose PAPO, which unifies two natural options (augmentation and hypernetwork) and achieves significantly better performance. PAPO consists of three components: i) the population-size encoding which transforms the original value of population size to an equivalent encoding to avoid training collapse, ii) a hypernetwork to generate a distinct policy for each game conditioned on the population size, and iii) the population size as an additional input to the generated policy. Next, we construct a multi-task-based training procedure to efficiently train the neural networks of PAPO by sampling data from multiple games with different population sizes. Finally, extensive experiments on multiple environments show the significant superiority of PAPO over baselines, and the analysis of the evolution of the generated policies further deepens our understanding of the two fields of finite-agent and infinite-agent games.

Knowledge-intensive tasks, such as open-domain question answering (QA), require access to a large amount of world or domain knowledge. A common approach for knowledge-intensive tasks is to employ a retrieve-then-read pipeline that first retrieves a handful of relevant contextual documents from an external corpus such as Wikipedia and then predicts an answer conditioned on the retrieved documents. In this paper, we present a novel perspective for solving knowledge-intensive tasks by replacing document retrievers with large language model generators. We call our method generate-then-read (GenRead), which first prompts a large language model to generate contextual documents based on a given question, and then reads the generated documents to produce the final answer. Furthermore, we propose a novel clustering-based prompting method that selects distinct prompts, in order to generate diverse documents that cover different perspectives, leading to better recall over acceptable answers. We conduct extensive experiments on three different knowledge-intensive tasks, including open-domain QA, fact checking, and dialogue system. Notably, GenRead achieves 71.6 and 54.4 exact match scores on TriviaQA and WebQ, significantly outperforming the state-of-the-art retrieve-then-read pipeline DPR-FiD by +4.0 and +3.9, without retrieving any documents from any external knowledge source. Lastly, we demonstrate the model performance can be further improved by combining retrieval and generation. Our code and generated documents can be found at https://github.com/wyu97/GenRead.

Graph-structured physical mechanisms are ubiquitous in real-world scenarios, thus revealing underneath formulas is of great importance for scientific discovery. However, classical symbolic regression methods fail on this task since they can only handle input-output pairs that are not graph-structured. In this paper, we propose a new approach that generalizes symbolic regression to graph-structured physical mechanisms. The essence of our method is to model the formula skeleton with a message-passing flow, which helps transform the discovery of the skeleton into the search for the message-passing flow. Such a transformation guarantees that we are able to search a message-passing flow, which is efficient and Pareto-optimal in terms of both accuracy and simplicity. Subsequently, the underneath formulas can be identified by interpreting component functions of the searched message-passing flow, reusing classical symbolic regression methods. We conduct extensive experiments on datasets from different physical domains, including mechanics, electricity, and thermology, and on real-world datasets of pedestrian dynamics without ground-truth formulas. The experimental results not only verify the rationale of our design but also demonstrate that the proposed method can automatically learn precise and interpretable formulas for graph-structured physical mechanisms.

Indiscriminate data poisoning attacks are quite effective against supervised learning. However, not much is known about their impact on unsupervised contrastive learning (CL). This paper is the first to consider indiscriminate poisoning attacks of contrastive learning. We propose Contrastive Poisoning (CP), the first effective such attack on CL. We empirically show that Contrastive Poisoning, not only drastically reduces the performance of CL algorithms, but also attacks supervised learning models, making it the most generalizable indiscriminate poisoning attack. We also show that CL algorithms with a momentum encoder are more robust to indiscriminate poisoning, and propose a new countermeasure based on matrix completion. Code is available at: https://github.com/kaiwenzha/contrastive-poisoning.

Recently, there has been a growing surge of interest in enabling machine learning systems to generalize well to Out-of-Distribution (OOD) data. Most efforts are devoted to advancing optimization objectives that regularize models to capture the underlying invariance; however, there often are compromises in the optimization process of these OOD objectives: i) Many OOD objectives have to be relaxed as penalty terms of Empirical Risk Minimization (ERM) for the ease of optimization, while the relaxed forms can weaken the robustness of the original objective; ii) The penalty terms also require careful tuning of the penalty weights due to the intrinsic conflicts between ERM and OOD objectives. Consequently, these compromises could easily lead to suboptimal performance of either the ERM or OOD objective. To address these issues, we introduce a multi-objective optimization (MOO) perspective to understand the OOD optimization process, and propose a new optimization scheme called PAreto Invariant Risk Minimization (PAIR). PAIR improves the robustness of OOD objectives by cooperatively optimizing with other OOD objectives, thereby bridging the gaps caused by the relaxations. Then PAIR approaches a Pareto optimal solution that trades off the ERM and OOD objectives properly. Extensive experiments on challenging benchmarks, WILDS, show that PAIR alleviates the compromises and yields top OOD performances.