ICLR 2023 Papers

In recent years, contrastive learning achieves impressive results on self-supervised visual representation learning, but there still lacks a rigorous understanding of its learning dynamics. In this paper, we show that if we cast a contrastive objective equivalently into the feature space, then its learning dynamics admits an interpretable form. Specifically, we show that its gradient descent corresponds to a specific message passing scheme on the corresponding augmentation graph. Based on this perspective, we theoretically characterize how contrastive learning gradually learns discriminative features with the alignment update and the uniformity update. Meanwhile, this perspective also establishes an intriguing connection between contrastive learning and Message Passing Graph Neural Networks (MP-GNNs). This connection not only provides a unified understanding of many techniques independently developed in each community, but also enables us to borrow techniques from MP-GNNs to design new contrastive learning variants, such as graph attention, graph rewiring, jumpy knowledge techniques, etc. We believe that our message passing perspective not only provides a new theoretical understanding of contrastive learning dynamics, but also bridges the two seemingly independent areas together, which could inspire more interleaving studies to benefit from each other. The code is available at https://github.com/PKU-ML/Message-Passing-Contrastive-Learning.

Inspired by the success of self-supervised autoregressive representation learning in natural language (GPT and its variants), and advances in recent visual architecture design with Vision Transformers (ViTs), in this paper, we explore the effects various design choices have on the success of applying such training strategies for visual feature learning. Specifically, we introduce a novel strategy that we call Random Segments with Autoregressive Coding (RandSAC). In RandSAC, we group patch representations (image tokens) into hierarchically arranged segments; within each segment, tokens are predicted in parallel, similar to BERT, while across segment predictions are sequential, similar to GPT. We illustrate that randomized serialization of the segments significantly improves the performance and results in distribution over spatially-long (across-segments) and -short (within-segment) predictions which are effective for feature learning. We illustrate the pertinence of these design choices and explore alternatives on a number of datasets (e.g., CIFAR10, ImageNet). While our pre-training strategy works with vanilla Transformer, we also propose a conceptually simple, but highly effective, addition to the decoder that allows learnable skip-connections to encoder feature layers, which further improves the performance.

Symbolic regression (SR) is an important technique for discovering hidden mathematical expressions from observed data. Transformer-based approaches have been widely used for machine translation due to their high performance, and are recently highly expected to be used for SR. They input the data points, then output the expression skeleton, and finally optimize the coefficients. However, recent transformer-based methods for SR focus more attention on large scale training data and ignore the ill-posed problem: the lack of sufficient supervision, i.e., expressions that may be completely different have the same supervision because of their same skeleton, which makes it challenging to deal with data that may be from the same expression skeleton but with different coefficients. Therefore, we present a transformer-based model for SR with the ability to alleviate this problem. Specifically, we leverage a feature extractor based on pure residual MLP networks to obtain more information about data points. Furthermore, the core idea is that we propose a joint learning mechanism combining supervised contrastive learning, which makes features of data points from expressions with the same skeleton more similar so as to effectively alleviates the ill-posed problem. The benchmark results show that the proposed method is up to 25% higher with respect to the recovery rate of skeletons than typical transformer-based methods. Moreover, our method outperforms state-of-the-art SR methods based on reinforcement learning and genetic programming in terms of the coefficient of determination ($R^2$).

Recently, it has been shown that neural networks not only approximate the ground-state wave functions of a single molecular system well but can also generalize to multiple geometries. While such generalization significantly speeds up training, each energy evaluation still requires Monte Carlo integration which limits the evaluation to a few geometries. In this work, we address the inference shortcomings by proposing the Potential learning from ab-initio Networks (PlaNet) framework, in which we simultaneously train a surrogate model in addition to the neural wave function. At inference time, the surrogate avoids expensive Monte-Carlo integration by directly estimating the energy, accelerating the process from hours to milliseconds. In this way, we can accurately model high-resolution multi-dimensional energy surfaces for larger systems that previously were unobtainable via neural wave functions. Finally, we explore an additional inductive bias by introducing physically-motivated restricted neural wave function models. We implement such a function with several additional improvements in the new PESNet++ model. In our experimental evaluation, PlaNet accelerates inference by 7 orders of magnitude for larger molecules like ethanol while preserving accuracy. Compared to previous energy surface networks, PESNet++ reduces energy errors by up to 74%.

Unsupervised meta-learning aims to learn generalizable knowledge across a distribution of tasks constructed from unlabeled data. Here, the main challenge is how to construct diverse tasks for meta-learning without label information; recent works have proposed to create, e.g., pseudo-labeling via pretrained representations or creating synthetic samples via generative models. However, such a task construction strategy is fundamentally limited due to heavy reliance on the immutable pseudo-labels during meta-learning and the quality of the representations or the generated samples. To overcome the limitations, we propose a simple yet effective unsupervised meta-learning framework, coined Pseudo-supervised Contrast (PsCo), for few-shot classification. We are inspired by the recent self-supervised learning literature; PsCo utilizes a momentum network and a queue of previous batches to improve pseudo-labeling and construct diverse tasks in a progressive manner. Our extensive experiments demonstrate that PsCo outperforms existing unsupervised meta-learning methods under various in-domain and cross-domain few-shot classification benchmarks. We also validate that PsCo is easily scalable to a large-scale benchmark, while recent prior-art meta-schemes are not.

Semi-supervised learning (SSL) provides an effective means of leveraging unlabelled data to improve a model’s performance. Even though the domain has received a considerable amount of attention in the past years, most methods present the common drawback of lacking theoretical guarantees. Our starting point is to notice that the estimate of the risk that most discriminative SSL methods minimise is biased, even asymptotically. This bias impedes the use of standard statistical learning theory and can hurt empirical performance. We propose a simple way of removing the bias. Our debiasing approach is straightforward to implement and applicable to most deep SSL methods. We provide simple theoretical guarantees on the trustworthiness of these modified methods, without having to rely on the strong assumptions on the data distribution that SSL theory usually requires. In particular, we provide generalisation error bounds for the proposed methods. We evaluate debiased versions of different existing SSL methods, such as the Pseudo-label method and Fixmatch, and show that debiasing can compete with classic deep SSL techniques in various settings by providing better calibrated models. Additionally, we provide a theoretical explanation of the intuition of the popular SSL methods. An implementation of a debiased version of Fixmatch is available at https://github.com/HugoSchmutz/DeFixmatch

Learning from large amounts of unsupervised data and a small amount of supervision is an important open problem in computer vision. We propose a new semi-supervised learning method, Semantic Positives via Pseudo-Labels (SEMPPL), that combines labelled and unlabelled data to learn informative representations. Our method extends self-supervised contrastive learning—where representations are shaped by distinguishing whether two samples represent the same underlying datum (positives) or not (negatives)—with a novel approach to selecting positives. To enrich the set of positives, we leverage the few existing ground-truth labels to predict the missing ones through a k-nearest neighbors classifier by using the learned embeddings of the labelled data. We thus extend the set of positives with datapoints having the same pseudo-label and call these semantic positives. We jointly learn the representation and predict bootstrapped pseudo-labels. This creates a reinforcing cycle. Strong initial representations enable better pseudo-label predictions which then improve the selection of semantic positives and lead to even better representations. SEMPPL outperforms competing semi-supervised methods setting new state-of-the-art performance of 76% and 68.5% top-1accuracy when using a ResNet-50 and training on 10% and 1% of labels on ImageNet, respectively. Furthermore, when using selective kernels, SEMPPL significantly outperforms previous state-of-the-art achieving 72.3% and 78.3% top-1accuracy on ImageNet with 1% and 10% labels, respectively, which improves absolute +7.8% and +6.2% over previous work. SEMPPL also exhibits state-of-the-art performance over larger ResNet models as well as strong robustness, out-of-distribution and transfer performance. We release the checkpoints and the evaluation code at https://github.com/deepmind/semppl.

There are sequence generation tasks where the best order to generate the target sequence is not left-to-right. For example, an answer to the Sudoku game, a structured code like s-expression, and even a logical natural language answer where the analysis may be generated after the decision. We define the target sequences of those tasks as complex sequences. Obviously, a complex sequence should be constructed with multiple logical steps, and has dependencies among each part of itself (e.g. decisions depend on analyses). It's a great challenge for the classic left-to-right autoregressive generation system to generate complex sequences. Current approaches improve one-pass left-to-right generation on NLG tasks by generating different heuristic intermediate sequences in multiple stages. However, for complex sequences, the heuristic rules to break down them may hurt performance, and increase additional exposure bias. To tackle these challenges, we propose a PLM-friendly autoregressive self-boost refinement framework, CASR. When training, CASR inputs the predictions generated by the model itself at the previous refinement step (instead of those produced by heuristic rules). To find an optimal design, we also discuss model architecture, parameter efficiency and initialization strategy. By evaluating CASR on Sudoku, WebQSP, MTOP and KVRET through controlled experiments and empirical studies, we find that CASR produces high-quality outputs. CASR also improves Accuracy on Sudoku (70.93% --> 97.28%) and achieves state-of-the-art performance on KVRET with Micro F1 score (67.88% --> 70.00%).

Simplicial Embeddings (SEM) are representations learned through self-supervised learning (SSL), wherein a representation is projected into $L$ simplices of $V$ dimensions each using a \texttt{softmax} operation. This procedure conditions the representation onto a constrained space during pretraining and imparts an inductive bias for group sparsity. For downstream classification, we formally prove that the SEM representation leads to better generalization than an unnormalized representation. Furthermore, we empirically demonstrate that SSL methods trained with SEMs have improved generalization on natural image datasets such as CIFAR-100 and ImageNet. Finally, when used in a downstream classification task, we show that SEM features exhibit emergent semantic coherence where small groups of learned features are distinctly predictive of semantically-relevant classes.

In this paper, we first extend the recent Masked Auto-Encoder (MAE) model from a single modality to audio-visual multi-modalities. Subsequently, we propose the Contrastive Audio-Visual Masked Auto-Encoder (CAV-MAE) by combining contrastive learning and masked data modeling, two major self-supervised learning frameworks, to learn a joint and coordinated audio-visual representation. Our experiments show that the contrastive audio-visual correspondence learning objective not only enables the model to perform audio-visual retrieval tasks, but also helps the model learn a better joint representation. As a result, our fully self-supervised pretrained CAV-MAE achieves a new SOTA accuracy of 65.9% on VGGSound, and is comparable with the previous best supervised pretrained model on AudioSet in the audio-visual event classification task. Code and pretrained models are at https://github.com/yuangongnd/cav-mae.

Several recent works on self-supervised learning are trained by mapping different augmentations of the same image to the same feature representation. The data augmentations used are of crucial importance to the quality of learned feature representations. In this paper, we analyze how the color jitter traditionally used in data augmentation negatively impacts the quality of the color features in learned feature representations. To address this problem, we propose a more realistic, physics-based color data augmentation - which we call Planckian Jitter - that creates realistic variations in chromaticity and produces a model robust to illumination changes that can be commonly observed in real life, while maintaining the ability to discriminate image content based on color information. Experiments confirm that such a representation is complementary to the representations learned with the currently-used color jitter augmentation and that a simple concatenation leads to significant performance gains on a wide range of downstream datasets. In addition, we present a color sensitivity analysis that documents the impact of different training methods on model neurons and shows that the performance of the learned features is robust with respect to illuminant variations. Official code available at: https://github.com/TheZino/PlanckianJitter

Many datasets are biased, namely they contain easy-to-learn features that are highly correlated with the target class only in the dataset but not in the true underlying distribution of the data. For this reason, learning unbiased models from biased data has become a very relevant research topic in the last years. In this work, we tackle the problem of learning representations that are robust to biases. We first present a margin-based theoretical framework that allows us to clarify why recent contrastive losses (InfoNCE, SupCon, etc.) can fail when dealing with biased data. Based on that, we derive a novel formulation of the supervised contrastive loss ($\epsilon$-SupInfoNCE), providing more accurate control of the minimal distance between positive and negative samples. Furthermore, thanks to our theoretical framework, we also propose FairKL, a new debiasing regularization loss, that works well even with extremely biased data. We validate the proposed losses on standard vision datasets including CIFAR10, CIFAR100, and ImageNet, and we assess the debiasing capability of FairKL with $\epsilon$-SupInfoNCE, reaching state-of-the-art performance on a number of biased datasets, including real instances of biases "in the wild".

In large-scale retrieval, the lexicon-weighting paradigm, learning weighted sparse representations in vocabulary space, has shown promising results with high quality and low latency. Despite it deeply exploiting the lexicon-representing capability of pre-trained language models, a crucial gap remains between language modeling and lexicon-weighting retrieval -- the former preferring certain or low-entropy words whereas the latter favoring pivot or high-entropy words -- becoming the main barrier to lexicon-weighting performance for large-scale retrieval. To bridge this gap, we propose a brand-new pre-training framework, lexicon-bottlenecked masked autoencoder (LexMAE), to learn importance-aware lexicon representations. Essentially, we present a lexicon-bottlenecked module between a normal language modeling encoder and a weakened decoder, where a continuous bag-of-words bottleneck is constructed to learn a lexicon-importance distribution in an unsupervised fashion. The pre-trained LexMAE is readily transferred to the lexicon-weighting retrieval via fine-tuning. On the ad-hoc retrieval benchmark, MS-Marco, it achieves 42.6% MRR@10 with 45.8 QPS for the passage dataset and 44.4% MRR@100 with 134.8 QPS for the document dataset, by a CPU machine. And LexMAE shows state-of-the-art zero-shot transfer capability on BEIR benchmark with 12 datasets.

Semi-supervised Learning (SSL) has witnessed great success owing to the impressive performances brought by various methods based on pseudo labeling and consistency regularization. However, we argue that existing methods might fail to utilize the unlabeled data more effectively since they either use a pre-defined / fixed threshold or an ad-hoc threshold adjusting scheme, resulting in inferior performance and slow convergence. We first analyze a motivating example to obtain intuitions on the relationship between the desirable threshold and model's learning status. Based on the analysis, we hence propose FreeMatch to adjust the confidence threshold in a self-adaptive manner according to the model's learning status. We further introduce a self-adaptive class fairness regularization penalty to encourage the model for diverse predictions during the early training stage. Extensive experiments indicate the superiority of FreeMatch especially when the labeled data are extremely rare. FreeMatch achieves 5.78%, 13.59%, and 1.28% error rate reduction over the latest state-of-the-art method FlexMatch on CIFAR-10 with 1 label per class, STL-10 with 4 labels per class, and ImageNet with 100 labels per class, respectively. Moreover, FreeMatch can also boost the performance of imbalanced SSL. The codes can be found at https://github.com/microsoft/Semi-supervised-learning.

Recent symbolic music generative models have achieved significant improvements in the quality of the generated samples. Nevertheless, it remains hard for users to control the output in such a way that it matches their expectation. To address this limitation, high-level, human-interpretable conditioning is essential. In this work, we release FIGARO, a Transformer-based conditional model trained to generate symbolic music based on a sequence of high-level control codes. To this end, we propose description-to-sequence learning, which consists of automatically extracting fine-grained, human-interpretable features (the description) and training a sequence-to-sequence model to reconstruct the original sequence given only the description as input. FIGARO achieves state-of-the-art performance in multi-track symbolic music generation both in terms of style transfer and sample quality. We show that performance can be further improved by combining human-interpretable with learned features. Our extensive experimental evaluation shows that FIGARO is able to generate samples that closely adhere to the content of the input descriptions, even when they deviate significantly from the training distribution.

We identify and overcome two key obstacles in extending the success of BERT-style pre-training, or masked image modeling, to convolutional networks (convnets): (i) convolution operation cannot handle irregular, randomly masked input images; (ii) the single-scale nature of BERT pre-training is inconsistent with convnet’s hierarchical structure. For (i), we treat unmasked pixels as sparse voxels of 3D point clouds and use sparse convolution to encode. This is the first use of sparse convolution for 2D masked modeling. For (ii), we develop a hierarchical decoder to reconstruct images from multi-scale encoded features. Our method, called Sparse masKed modeling (SparK), is general: it can be used directly on any convolutional model without backbone modifications. We validate it on both classical (ResNet) and modern (ConvNeXt) models: on three downstream tasks, it surpasses both state-of-the-art contrastive learning and transformer-based masked modeling by similarly large margins (around +1.0%). The improvements on object detection and instance segmentation are more significant (up to +3.5%), validating the strong transferability of features learned. We also find SparK’s favorable scaling behavior by observing more gains on larger networks. All of these findings support the promising future of generative pre-training on convnets. Both codes and pre-trained models have been released at https://github.com/keyu-tian/SparK.

Skip connections and normalisation layers form two standard architectural components that are ubiquitous for the training of Deep Neural Networks (DNNs), but whose precise roles are poorly understood. Recent approaches such as Deep Kernel Shaping have made progress towards reducing our reliance on them, using insights from wide NN kernel theory to improve signal propagation in vanilla DNNs (which we define as networks without skips or normalisation). However, these approaches are incompatible with the self-attention layers present in transformers, whose kernels are intrinsically more complicated to analyse and control. And so the question remains: \emph{is it possible to train deep vanilla transformers?} We answer this question in the affirmative by designing several approaches that use combinations of parameter initialisations, bias matrices and location-dependent rescaling to achieve faithful signal propagation in vanilla transformers. Our methods address various intricacies specific to signal propagation in transformers, including the interaction with positional encoding and causal masking. In experiments on WikiText-103 and C4, our approaches enable deep transformers without normalisation to train at speeds matching their standard counterparts, and deep vanilla transformers to reach the same performance as standard ones after about 5 times more iterations.

Standard empirical risk minimization (ERM) training can produce deep neural network (DNN) models that are accurate on average but under-perform in under-represented population subgroups, especially when there are imbalanced group distributions in the long-tailed training data. Therefore, approaches that improve the accuracy - group robustness trade-off frontier of a DNN model (i.e. improving worst-group accuracy without sacrificing average accuracy, or vice versa) is of crucial importance. Uncertainty-based active learning (AL) can potentially improve the frontier by preferentially sampling underrepresented subgroups to create a more balanced training dataset. However, the quality of uncertainty estimates from modern DNNs tend to degrade in the presence of spurious correlations and dataset bias, compromising the effectiveness of AL for sampling tail groups. In this work, we propose Introspective Self-play (ISP), a simple approach to improve the uncertainty estimation of a deep neural network under dataset bias, by adding an auxiliary introspection task requiring a model to predict the bias for each data point in addition to the label. We show that ISP provably improves the bias-awareness of the model representation and the resulting uncertainty estimates. On two real-world tabular and language tasks,ISP serves as a simple “plug-in” for AL model training, consistently improving both the tail-group sampling rate and the final accuracy-fairness trade-off frontier of popular AL methods.

Deep neural networks based on layer-stacking architectures have historically suffered from poor inherent interpretability. Meanwhile, symbolic probabilistic models function with clear interpretability, but how to combine them with neural networks to enhance their performance remains to be explored. In this paper, we try to marry these two systems for text classification via a structured language model. We propose a Symbolic-Neural model that can learn to explicitly predict class labels of text spans from a constituency tree without requiring any access to span-level gold labels. As the structured language model learns to predict constituency trees in a self-supervised manner, only raw texts and sentence-level labels are required as training data, which makes it essentially a general constituent-level self-interpretable classification model. Our experiments demonstrate that our approach could achieve good prediction accuracy in downstream tasks. Meanwhile, the predicted span labels are consistent with human rationales to a certain degree.