ICLR 2023 Papers

We tackle the problem of Selective Classification where the objective is to achieve the best performance on a predetermined ratio (coverage) of the dataset. Recent state-of-the-art selective methods come with architectural changes either via introducing a separate selection head or an extra abstention logit. In this paper, we challenge the aforementioned methods. The results suggest that the superior performance of state-of-the-art methods is owed to training a more generalizable classifier rather than their proposed selection mechanisms. We argue that the best performing selection mechanism should instead be rooted in the classifier itself. Our proposed selection strategy uses the classification scores and achieves better results by a significant margin, consistently, across all coverages and all datasets, without any added compute cost. Furthermore, inspired by semi-supervised learning, we propose an entropy-based regularizer that improves the performance of selective classification methods. Our proposed selection mechanism with the proposed entropy-based regularizer achieves new state-of-the-art results.

The expressive power of Graph Neural Networks (GNNs) is fundamental for understanding their capabilities and limitations, i.e., what graph properties can or cannot be learnt by a GNN. Since standard GNNs have been characterised to be upper-bounded by the Weisfeiler-Lehman (1-WL) algorithm, recent attempts concentrated on developing more expressive GNNs in terms of the $k$-WL hierarchy, a well-established framework for graph isormorphism tests. In this work we show that, contrary to the widely accepted view, the $k$-WL hierarchy is not well-suited for measuring expressive GNNs. This is due to limitations that are inherent to high-dimensional WL algorithms such as the lack of a natural interpretation and high computational costs, which makes it difficult to draw any firm conclusions about the expressive power of GNNs beyond 1-WL. Thus, we propose a novel hierarchy of graph isomorphism tests, namely Neighbourhood WL ($\mathscr{N}$-WL), and also establish a new theorem on the equivalence of expressivity between induced connected subgraphs and induced subgraphs within this hierarchy. Further, we design a GNN model upon $\mathscr{N}$-WL, Graph Neighbourhood Neural Network (G3N), and empirically verify its expressive power on synthetic and real-world benchmarks.

Comprehensive studies of synthetic optical flow datasets have attempted to reveal what properties lead to accuracy improvement in learning-based optical flow estimation. However, manually identifying and verifying the properties that contribute to accurate optical flow estimation require large-scale trial-and-error experiments with iteratively generating whole synthetic datasets and training on them, \ie, impractical. To address this challenge, we propose a differentiable optical flow data generation pipeline and a loss function to drive the pipeline, called DFlow. DFlow efficiently synthesizes a dataset effective for a target domain without the need for cumbersome try-and-errors. This favorable property is achieved by proposing an efficient dataset comparison method that uses neural networks to approximately encode each dataset and compares the proxy networks instead of explicitly comparing datasets in a pairwise way. Our experiments show the competitive performance of our DFlow against the prior arts in pre-training. Furthermore, compared to competing datasets, DFlow achieves the best fine-tuning performance on the Sintel public benchmark with RAFT.

Empirical studies suggest that machine learning models trained with empirical risk minimization (ERM) often rely on attributes that may be spuriously correlated with the class labels. Such models typically lead to poor performance during inference for data lacking such correlations. In this work, we explicitly consider a situation where potential spurious correlations are present in the majority of training data. In contrast with existing approaches, which use the ERM model outputs to detect the samples without spurious correlations and either heuristically upweight or upsample those samples, we propose the logit correction (LC) loss, a simple yet effective improvement on the softmax cross-entropy loss, to correct the sample logit. We demonstrate that minimizing the LC loss is equivalent to maximizing the group-balanced accuracy, so the proposed LC could mitigate the negative impacts of spurious correlations. Our extensive experimental results further reveal that the proposed LC loss outperforms state-of-the-art solutions on multiple popular benchmarks by a large margin, an average 5.5% absolute improvement, without access to spurious attribute labels. LC is also competitive with oracle methods that make use of the attribute labels.

The great success of deep learning heavily relies on increasingly larger training data, which comes at a price of huge computational and infrastructural costs. This poses crucial questions that, do all training data contribute to model's performance? How much does each individual training sample or a sub-training-set affect the model's generalization, and how to construct the smallest subset from the entire training data as a proxy training set without significantly sacrificing the model's performance? To answer these, we propose dataset pruning, an optimization-based sample selection method that can (1) examine the influence of removing a particular set of training samples on model's generalization ability with theoretical guarantee, and (2) construct the smallest subset of training data that yields strictly constrained generalization gap. The empirically observed generalization gap of dataset pruning is substantially consistent with our theoretical expectations. Furthermore, the proposed method prunes 40% training examples on the CIFAR-10 dataset, halves the convergence time with only 1.3% test accuracy decrease, which is superior to previous score-based sample selection methods.

Neural message passing is a basic feature extraction unit for graph-structured data considering neighboring node features in network propagation from one layer to the next. We model such process by an interacting particle system with attractive and repulsive forces and the Allen-Cahn force arising in the modeling of phase transition. The dynamics of the system is a reaction-diffusion process which can separate particles without blowing up. This induces an Allen-Cahn message passing (ACMP) for graph neural networks where the numerical iteration for the particle system solution constitutes the message passing propagation. ACMP which has a simple implementation with a neural ODE solver can propel the network depth up to one hundred of layers with theoretically proven strictly positive lower bound of the Dirichlet energy. It thus provides a deep model of GNNs circumventing the common GNN problem of oversmoothing. GNNs with ACMP achieve state of the art performance for real-world node classification tasks on both homophilic and heterophilic datasets. Codes are available at https://github.com/ykiiiiii/ACMP

There has been growing interest in automatically predicting missing type annotations in programs written in Python and JavaScript. While prior methods have achieved impressive accuracy when predicting the most common types, they often perform poorly on rare or complex types. In this paper, we present a new type inference method that treats type prediction as a code infilling task by leveraging CodeT5, a state-of-the-art seq2seq pre-trained language model for code. Our method uses static analysis to construct dynamic contexts for each code element whose type signature is to be predicted by the model. We also propose an iterative decoding scheme that incorporates previous type predictions in the model's input context, allowing information exchange between related code elements. Our evaluation shows that the proposed approach, TypeT5, not only achieves a higher overall accuracy (particularly on rare and complex types) but also produces more coherent results with fewer type errors---while enabling easy user intervention.

Pretrained language models have shown superior performance on many natural language processing tasks, yet they still struggle at multi-step formal reasoning tasks like grade school math problems. One key challenge of finetuning them to solve such math reasoning problems is that many existing datasets only contain one reference solution for each problem, despite the fact that there are often alternative solutions resembling different reasoning paths to the final answer. This way, the finetuned models are biased towards the limited reference solutions, which limits their generalization to unseen examples. To mitigate this issue, we propose to let the model perform sampling during training and learn from both self-sampled fully-correct solutions, which yield the correct answer upon execution, and partially-correct solutions, whose intermediate state matches an intermediate state of a known correct solution. We show that our use of self-sampled correct and partially-correct solutions can benefit learning and help guide the sampling process, leading to more efficient exploration of the solution space. Additionally, we explore various training objectives to support learning from multiple solutions per example and find they greatly affect the performance. Experiments on two math reasoning datasets show the effectiveness of our method compared to learning from a single reference solution with MLE, where we improve PASS@100 from 35.5% to 44.5% for GSM8K, and 27.6% to 36.2% PASS@80 for MathQA. Such improvements are also consistent across different model sizes.

Despite the empirical advances of deep learning across a variety of learning tasks, our theoretical understanding of its success is still very restricted. One of the key challenges is the overparametrized nature of modern models, enabling complete overfitting of the data even if the labels are randomized, i.e. networks can completely \textit{memorize} all given patterns. While such a memorization capacity seems worrisome, in this work we show that under training protocols that include \textit{data augmentation}, neural networks learn to memorize entirely random labels in a benign way, i.e. they learn embeddings that lead to highly non-trivial performance under nearest neighbour probing. We demonstrate that deep models have the surprising ability to separate noise from signal by distributing the task of memorization and feature learning to different layers. As a result, only the very last layers are used for memorization, while preceding layers encode performant features which remain largely unaffected by the label noise. We explore the intricate role of the augmentations used for training and identify a memorization-generalization trade-off in terms of their diversity, marking a clear distinction to all previous works. Finally, we give a first explanation for the emergence of benign memorization by showing that \textit{malign} memorization under data augmentation is infeasible due to the insufficient capacity of the model for the increased sample size. As a consequence, the network is forced to leverage the correlated nature of the augmentations and as a result learns meaningful features. To complete the picture, a better theory of feature learning in deep neural networks is required to fully understand the origins of this phenomenon.

In representation learning, a common approach is to seek representations which disentangle the underlying factors of variation. Eastwood & Williams (2018) proposed three metrics for quantifying the quality of such disentangled representations: disentanglement (D), completeness (C) and informativeness (I). In this work, we first connect this DCI framework to two common notions of linear and nonlinear identifiability, thereby establishing a formal link between disentanglement and the closely-related field of independent component analysis. We then propose an extended DCI-ES framework with two new measures of representation quality—explicitness (E) and size (S)—and point out how D and C can be computed for black-box predictors. Our main idea is that the functional capacity required to use a representation is an important but thus-far neglected aspect of representation quality, which we quantify using explicitness or ease-of-use (E). We illustrate the relevance of our extensions on the MPI3D and Cars3D datasets.

The success of machine learning relies heavily on massive amounts of data, which are usually generated and stored across a range of diverse and distributed data sources. Decentralized learning has thus been advocated and widely deployed to make efficient use of distributed datasets, with an extensive focus on supervised learning (SL) problems. Unfortunately, the majority of real-world data are unlabeled and can be highly heterogeneous across sources. In this work, we carefully study decentralized learning with unlabeled data through the lens of self-supervised learning (SSL), specifically contrastive visual representation learning. We study the effectiveness of a range of contrastive learning algorithms under a decentralized learning setting, on relatively large-scale datasets including ImageNet-100, MS-COCO, and a new real-world robotic warehouse dataset. Our experiments show that the decentralized SSL (Dec-SSL) approach is robust to the heterogeneity of decentralized datasets, and learns useful representation for object classification, detection, and segmentation tasks, even when combined with the simple and standard decentralized learning algorithm of Federated Averaging (FedAvg). This robustness makes it possible to significantly reduce communication and to reduce the participation ratio of data sources with only minimal drops in performance. Interestingly, using the same amount of data, the representation learned by Dec-SSL can not only perform on par with that learned by centralized SSL which requires communication and excessive data storage costs, but also sometimes outperform representations extracted from decentralized SL which requires extra knowledge about the data labels. Finally, we provide theoretical insights into understanding why data heterogeneity is less of a concern for Dec-SSL objectives, and introduce feature alignment and clustering techniques to develop a new Dec-SSL algorithm that further improves the performance, in the face of highly non-IID data. Our study presents positive evidence to embrace unlabeled data in decentralized learning, and we hope to provide new insights into whether and why decentralized SSL is effective and/or even advantageous.

Heterogeneity of data distributed across clients limits the performance of global models trained through federated learning, especially in the settings with highly imbalanced class distributions of local datasets. In recent years, personalized federated learning (pFL) has emerged as a potential solution to the challenges presented by heterogeneous data. However, existing pFL methods typically enhance performance of local models at the expense of the global model's accuracy. We propose FedHKD (Federated Hyper-Knowledge Distillation), a novel FL algorithm in which clients rely on knowledge distillation (KD) to train local models. In particular, each client extracts and sends to the server the means of local data representations and the corresponding soft predictions -- information that we refer to as ``hyper-knowledge". The server aggregates this information and broadcasts it to the clients in support of local training. Notably, unlike other KD-based pFL methods, FedHKD does not rely on a public dataset nor it deploys a generative model at the server. We analyze convergence of FedHKD and conduct extensive experiments on visual datasets in a variety of scenarios, demonstrating that FedHKD provides significant improvement in both personalized as well as global model performance compared to state-of-the-art FL methods designed for heterogeneous data settings.

Neural sequence models, especially transformers, exhibit a remarkable capacity for in-context learning. They can construct new predictors from sequences of labeled examples $(x, f(x))$ presented in the input without further parameter updates. We investigate the hypothesis that transformer-based in-context learners implement standard learning algorithms implicitly, by encoding context-specific parametric models in their hidden representations, and updating these implicit models as new examples appear in the context. Using linear regression as a model problem, we offer three sources of evidence for this hypothesis. First, we prove by construction that transformers can implement learning algorithms for linear models based on gradient descent and closed-form computation of regression parameters. Second, we show that trained in-context learners closely match the predictors computed by gradient descent, ridge regression, and exact least-squares regression, transitioning between different predictors as transformer depth and dataset noise vary. Third, we present preliminary evidence that in-context learners share algorithmic features with these predictors: learners' late layers encode weight vectors and moment matrices. These results suggest that in-context learning is understandable in algorithmic terms, and that (at least in the linear case) learners may work by rediscovering standard estimation algorithms.

In supervised learning, it has been shown that label noise in the data can be interpolated without penalties on test accuracy. We show that interpolating label noise induces adversarial vulnerability, and prove the first theorem showing the relationship between label noise and adversarial risk for any data distribution. Our results are almost tight if we do not make any assumptions on the inductive bias of the learning algorithm. We then investigate how different components of this problem affect this result including properties of the distribution. We also discuss non-uniform label noise distributions; and prove a new theorem showing uniform label noise induces nearly as large an adversarial risk as the worst poisoning with the same noise rate. Then, we provide theoretical and empirical evidence that uniform label noise is more harmful than typical real-world label noise. Finally, we show how inductive biases amplify the effect of label noise and argue the need for future work in this direction.

In this paper we present a novel method to estimate 3D human pose and shape from monocular videos. This task requires directly recovering pixel-alignment 3D human pose and body shape from monocular images or videos, which is challenging due to its inherent ambiguity. To improve precision, existing methods highly rely on the initialized mean pose and shape as prior estimates and parameter regression with an iterative error feedback manner. In addition, video-based approaches model the overall change over the image-level features to temporally enhance the single-frame feature, but fail to capture the rotational motion at the joint level, and cannot guarantee local temporal consistency. To address these issues, we propose a novel Transformer-based model with a design of independent tokens. First, we introduce three types of tokens independent of the image feature: \textit{joint rotation tokens, shape token, and camera token}. By progressively interacting with image features through Transformer layers, these tokens learn to encode the prior knowledge of human 3D joint rotations, body shape, and position information from large-scale data, and are updated to estimate SMPL parameters conditioned on a given image. Second, benefiting from the proposed token-based representation, we further use a temporal model to focus on capturing the rotational temporal information of each joint, which is empirically conducive to preventing large jitters in local parts. Despite being conceptually simple, the proposed method attains superior performances on the 3DPW and Human3.6M datasets. Using ResNet-50 and Transformer architectures, it obtains 42.0 mm error on the PA-MPJPE metric of the challenging 3DPW, outperforming state-of-the-art counterparts by a large margin. Code will be publicly available\footnote{\url{https://github.com/yangsenius/INT_HMR_Model}}.

When learning task-oriented dialogue (ToD) agents, reinforcement learning (RL) techniques can naturally be utilized to train dialogue strategies to achieve user-specific goals. Prior works mainly focus on adopting advanced RL techniques to train the ToD agents, while the design of the reward function is not well studied. This paper aims at answering the question of how to efficiently learn and leverage a reward function for training end-to-end (E2E) ToD agents. Specifically, we introduce two generalized objectives for reward-function learning, inspired by the classical learning-to-rank literature. Further, we utilize the learned reward function to guide the training of the E2E ToD agent. With the proposed techniques, we achieve competitive results on the E2E response-generation task on the Multiwoz 2.0 dataset. Source code and checkpoints are publicly released at https://github.com/Shentao-YANG/Fantastic_Reward_ICLR2023.

Monocular scene reconstruction from posed images is challenging due to the complexity of a large environment. Recent volumetric methods learn to directly predict the TSDF volume and have demonstrated promising results in this task. However, most methods focus on how to extract and fuse the 2D features to a 3D feature volume, but none of them improve the way how the 3D volume is aggregated. In this work, we propose an SDF transformer network, which replaces the role of 3D CNN for better 3D feature aggregation. To reduce the explosive computation complexity of the 3D multi-head attention, we propose a sparse window attention module, where the attention is only calculated between the non-empty voxels within a local window. Then a top-down-bottom-up 3D attention network is built for 3D feature aggregation, where a dilate-attention structure is proposed to prevent geometry degeneration, and two global modules are employed to equip with global receptive fields. The experiments on multiple datasets show that this 3D transformer network generates a more accurate and complete reconstruction, which outperforms previous methods by a large margin. Remarkably, the mesh accuracy is improved by 41.8%, and the mesh completeness is improved by 25.3% on the ScanNet dataset. The code of our method will be made public.

With more people publishing their personal data online, unauthorized data usage has become a serious concern. The unlearnable examples strategies have been introduced to prevent third parties from training on the data without permission. They add perturbations to the users’ data before publishing, so as to make the models trained on the perturbed published dataset invalidated. These perturbations have been generated for a specific training setting and a target dataset. However, their unlearnable effects significantly decrease when used in other training settings or datasets. To tackle this issue, we propose a novel unlearnable strategy based on Class-wise Separability Discriminant (CSD), which boosts the transferability of the unlearnable perturbations by enhancing the linear separability. Extensive experiments demonstrate the transferability of the unlearnable examples crafted by our proposed method across training settings and datasets.

1-bit gradient compression and local steps are two representative techniques that enable drastic communication reduction in distributed SGD. Their benefits, however, remain an open question on Adam-based large model pre-training (e.g. BERT and GPT). In this paper, we demonstrate the non-linearity in Adam causes slow convergence even when 1-bit compression or local steps are individually applied. To alleviate this limitation, we propose \textbf{0/1 Adam} that linearizes each Adam step via approximating its optimizer states using their stale estimates and linear correlation. \textbf{0/1 Adam} performs an Adam-like step to preserve the adaptivity, while its linearity allows utilizing 1-bit compression and local steps simultaneously for wall-clock time speed up. We provide convergence guarantee for \textbf{0/1 Adam} on smooth non-convex objectives. On various large-scale benchmarks such as BERT-Base, BERT-Large, GPT-2 pre-training and ImageNet, we demonstrate on up to 128 GPUs that \textbf{0/1 Adam} is able to reduce up to 87\% of data volume, 54\% of communication rounds, and achieve up to 2$\times$ higher training throughput and end-to-end training time reduction compared to the state-of-the-art baseline 1-bit Adam; while enjoying the same statistical convergence speed and end task model accuracy on GLUE dataset and ImageNet validation set.