ICLR 2022 Papers

Supervised learning is a fundamental framework used to train machine learning systems. A supervised learning problem is often formulated using an i.i.d. assumption that restricts model attention to a single relevant signal at a time when predicting. This contrasts with the human ability to actively use related samples as reference when making decisions. We hypothesize that the restriction to a single signal for each prediction in the standard i.i.d. framework contributes to well-known drawbacks of supervised learning: making overconfident predictions and vulnerability to overfitting, adversarial attacks, and out-of-distribution data. To address these limitations, we propose a new supervised learning paradigm called self-joint learning that generalizes the standard approach by modeling the joint conditional distribution of two observed samples, where each sample is an image and its label. Rather than assuming samples are independent, our models explicitly learn the sample-to-sample relation of conditional independence. Our framework can naturally incorporate auxiliary unlabeled data to further improve the performance. Experiments on benchmark image datasets show our method offers significant improvement over standard supervised learning in terms of accuracy, robustness against adversarial attacks, out-of-distribution detection, and overconfidence mitigation.

We advocate for a practical Maximum Likelihood Estimation (MLE) approach towards designing loss functions for regression and forecasting, as an alternative to the typical approach of direct empirical risk minimization on a specific target metric. The MLE approach is better suited to capture inductive biases such as prior domain knowledge in datasets, and can output post-hoc estimators at inference time that can optimize different types of target metrics. We present theoretical results to demonstrate that our approach is competitive with any estimator for the target metric under some general conditions. In two example practical settings, Poisson and Pareto regression, we show that our competitive results can be used to prove that the MLE approach has better excess risk bounds than directly minimizing the target metric. We also demonstrate empirically that our method instantiated with a well-designed general purpose mixture likelihood family can obtain superior performance for a variety of tasks across time-series forecasting and regression datasets with different data distributions.

Recently, large-scale Contrastive Language-Image Pre-training (CLIP) has attracted unprecedented attention for its impressive zero-shot recognition ability and excellent transferability to downstream tasks. However, CLIP is quite data-hungry and requires 400M image-text pairs for pre-training, thereby restricting its adoption. This work proposes a novel training paradigm, Data efficient CLIP (DeCLIP), to alleviate this limitation. We demonstrate that by carefully utilizing the widespread supervision among the image-text pairs, our De-CLIP can learn generic visual features more efficiently. Instead of using the single image-text contrastive supervision, we fully exploit data potential through the use of (1) self-supervision within each modality; (2) multi-view supervision across modalities; (3) nearest-neighbor supervision from other similar pairs. Benefiting from intrinsic supervision, our DeCLIP-ResNet50 can achieve 60.4% zero-shot top1 accuracy on ImageNet, which is 0.8% above the CLIP-ResNet50 while using 7.1×fewer data. Our DeCLIP-ResNet50 outperforms its counterpart in 8 out of 11 visual datasets when transferred to downstream tasks. Moreover, Scaling up the model and computing also works well in our framework.

State-of-the-art maximum entropy models for texture synthesis are built from statistics relying on image representations defined by convolutional neural networks (CNN). Such representations capture rich structures in texture images, outperforming wavelet-based representations in this regard. However, conversely to neural networks, wavelets offer meaningful representations, as they are known to detect structures at multiple scales (e.g. edges) in images. In this work, we propose a family of statistics built upon non-linear wavelet based representations, that can be viewed as a particular instance of a one-layer CNN, using a generalized rectifier non-linearity. These statistics significantly improve the visual quality of previous classical wavelet-based models, and allow one to produce syntheses of similar quality to state-of-the-art models, on both gray-scale and color textures. We further provide insights on memorization effects in these models.

Most existing Vision-and-Language (V&L) models rely on pre-trained visual encoders, using a relatively small set of manually-annotated data (as compared to web-crawled data), to perceive the visual world. However, it has been observed that large-scale pretraining usually can result in better generalization performance, e.g., CLIP (Contrastive Language-Image Pre-training), trained on a massive amount of image-caption pairs, has shown a strong zero-shot capability on various vision tasks. To further study the advantage brought by CLIP, we propose to use CLIP as the visual encoder in various V&L models in two typical scenarios: 1) plugging CLIP into task-specific fine-tuning; 2) combining CLIP with V&L pre-training and transferring to downstream tasks. We show that CLIP significantly outperforms widely-used visual encoders trained with in-domain annotated data, such as BottomUp-TopDown. We achieve competitive or better results on diverse V&L tasks, while establishing new state-of-the-art results on Visual Question Answering, Visual Entailment, and V&L Navigation tasks.

Most current few-shot learning methods train a model from abundantly labeled base category data and then transfer and adapt the model to sparsely labeled novel category data. These methods mostly generalize well on novel categories from the same domain as the base categories but perform poorly for distant domain categories. In this paper, we propose a framework for few-shot learning coined as ConFeSS (Contrastive Learning and Feature Selection System) that tackles large domain shift between base and novel categories. The first step of our framework trains a feature extracting backbone with the contrastive loss on the base category data. Since the contrastive loss does not use supervision, the features can generalize better to distant target domains. For the second step, we train a masking module to select relevant features that are more suited to target domain classification. Finally, a classifier is fine-tuned along with the backbone such that the backbone produces features similar to the relevant ones. To evaluate our framework, we tested it on a recently introduced cross-domain few-shot learning benchmark. Experimental results demonstrate that our framework outperforms all meta-learning approaches and produces competitive results against recent cross-domain methods. Additional analyses are also performed to better understand our framework.

The success of state-of-the-art deep neural networks heavily relies on the presence of large-scale labelled datasets, which are extremely expensive and time-consuming to annotate. This paper focuses on tackling semi-supervised part segmentation tasks by generating high-quality images with a pre-trained GAN and labelling the generated images with an automatic annotator. In particular, we formulate the annotator learning as a learning-to-learn problem. Given a pre-trained GAN, the annotator learns to label object parts in a set of randomly generated images such that a part segmentation model trained on these synthetic images with their predicted labels obtains low segmentation error on a small validation set of manually labelled images. We further reduce this nested-loop optimization problem to a simple gradient matching problem and efficiently solve it with an iterative algorithm. We show that our method can learn annotators from a broad range of labelled images including real images, generated images, and even analytically rendered images. Our method is evaluated with semi-supervised part segmentation tasks and significantly outperforms other semi-supervised competitors when the amount of labelled examples is extremely limited.

How to learn an effective reinforcement learning-based model for control tasks from high-level visual observations is a practical and challenging problem. A key to solving this problem is to learn low-dimensional state representations from observations, from which an effective policy can be learned. In order to boost the learning of state encoding, recent works are focused on capturing behavioral similarities between state representations or applying data augmentation on visual observations. In this paper, we propose a novel meta-learner-based framework for representation learning regarding behavioral similarities for reinforcement learning. Specifically, our framework encodes the high-dimensional observations into two decomposed embeddings regarding reward and dynamics in a Markov Decision Process (MDP). A pair of meta-learners are developed, one of which quantifies the reward similarity and the other quantifies dynamics similarity over the correspondingly decomposed embeddings. The meta-learners are self-learned to update the state embeddings by approximating two disjoint terms in on-policy bisimulation metric. To incorporate the reward and dynamics terms, we further develop a strategy to adaptively balance their impacts based on different tasks or environments. We empirically demonstrate that our proposed framework outperforms state-of-the-art baselines on several benchmarks, including conventional DM Control Suite, Distracting DM Control Suite and a self-driving task CARLA.

Large-scale pre-training has proven to be effective for visual representation learning on downstream tasks, especially for improving robustness and generalization. However, the recently developed detection transformers only employ pre-training on its backbone while leaving the key component, i.e., a 12-layer transformer, being trained from scratch, which prevents the model from above benefits. This separated training paradigm is mainly caused by the discrepancy between the upstream and downstream tasks. To mitigate the issue, we propose FP-DETR, a new method that Fully Pre-Trains an encoder-only transformer and smoothly fine-tunes it for object detection via a task adapter. Inspired by the success of textual prompts in NLP, we treat query positional embeddings as visual prompts to help the model attend to the target area (prompting) and recognize the object. To this end, we propose the task adapter which leverages self-attention to model the contextual relation between object query embedding. Experiments on the challenging COCO dataset demonstrate that our FP-DETR achieves competitive performance. Moreover, it enjoys better robustness to common corruptions and generalization to small-size datasets than state-of-the-art detection transformers. Code will be made publicly available at $\url{https://github.com/encounter1997/FP-DETR}$.

We present implicit displacement fields, a novel representation for detailed 3D geometry. Inspired by a classic surface deformation technique, displacement mapping, our method represents a complex surface as a smooth base surface plus a displacement along the base's normal directions, resulting in a frequency-based shape decomposition, where the high-frequency signal is constrained geometrically by the low-frequency signal. Importantly, this disentanglement is unsupervised thanks to a tailored architectural design that has an innate frequency hierarchy by construction. We explore implicit displacement field surface reconstruction and detail transfer and demonstrate superior representational power, training stability, and generalizability.

Self-supervised protein language models have proved their effectiveness in learning the proteins representations. With the increasing computational power, current protein language models pre-trained with millions of diverse sequences can advance the parameter scale from million-level to billion-level and achieve remarkable improvement. However, those prevailing approaches rarely consider incorporating knowledge graphs (KGs), which can provide rich structured knowledge facts for better protein representations. We argue that informative biology knowledge in KGs can enhance protein representation with external knowledge. In this work, we propose OntoProtein, the first general framework that makes use of structure in GO (Gene Ontology) into protein pre-training models. We construct a novel large-scale knowledge graph that consists of GO and its related proteins, and gene annotation texts or protein sequences describe all nodes in the graph. We propose novel contrastive learning with knowledge-aware negative sampling to jointly optimize the knowledge graph and protein embedding during pre-training. Experimental results show that OntoProtein can surpass state-of-the-art methods with pre-trained protein language models in TAPE benchmark and yield better performance compared with baselines in protein-protein interaction and protein function prediction.

The success of language Transformers is primarily attributed to the pretext task of masked language modeling (MLM), where texts are first tokenized into semantically meaningful pieces. In this work, we study masked image modeling (MIM) and indicate the necessity and challenges of using a semantically meaningful visual tokenizer. We present a self-supervised framework iBOT that can perform masked prediction with an online tokenizer. Specifically, we perform self-distillation on masked patch tokens and take the teacher network as the online tokenizer, along with self-distillation on the class token to acquire visual semantics. The online tokenizer is jointly learnable with the MIM objective and dispenses with a multi-stage training pipeline where the tokenizer needs to be pre-trained beforehand. We show the prominence of iBOT by achieving an 82.3% linear probing accuracy and an 87.8% fine-tuning accuracy evaluated on ImageNet-1K. Beyond the state-of-the-art image classification results, we underline emerging local semantic patterns, which helps the models to obtain strong robustness against common corruptions and achieve leading results on dense downstream tasks, e.g., object detection, instance segmentation, and semantic segmentation.

Variational Autoencoders (VAEs) are one of the most commonly used generative models, particularly for image data. A prominent difficulty in training VAEs is data that is supported on a lower dimensional manifold. Recent work by Dai and Wipf (2020) proposes a two-stage training algorithm for VAEs, based on a conjecture that in standard VAE training the generator will converge to a solution with 0 variance which is correctly supported on the ground truth manifold. They gave partial support for this conjecture by showing that some optima of the VAE loss do satisfy this property, but did not analyze the training dynamics. In this paper, we show that for linear encoders/decoders, the conjecture is true—that is the VAE training does recover a generator with support equal to the ground truth manifold—and does so due to an implicit bias of gradient descent rather than merely the VAE loss itself. In the nonlinear case, we show that VAE training frequently learns a higher-dimensional manifold which is a superset of the ground truth manifold.

Recently, post-training quantization (PTQ) has driven much attention to produce efficient neural networks without long-time retraining. Despite the low cost, current PTQ works always fail under the extremely low-bit setting. In this study, we pioneeringly confirm that properly incorporating activation quantization into the PTQ reconstruction benefits the final accuracy. To deeply understand the inherent reason, a theoretical framework is established, which inspires us that the flatness of the optimized low-bit model on calibration and test data is crucial. Based on the conclusion, a simple yet effective approach dubbed as \textsc{QDrop} is proposed, which randomly drops the quantization of activations during reconstruction. Extensive experiments on various tasks including computer vision (image classification, object detection) and natural language processing (text classification and question answering) prove its superiority. With \textsc{QDrop}, the limit of PTQ is pushed to the 2-bit activation for the first time and the accuracy boost can be up to 51.49\%. Without bells and whistles, \textsc{QDrop} establishes a new state of the art for PTQ.

Deep equilibrium layers (DEQs) have demonstrated promising performance and are competitive with standard explicit models on many benchmarks. However, little is known about certifying robustness for these models. Inspired by interval bound propagation (IBP), we propose the IBP-MonDEQ layer, a DEQ layer whose robustness can be verified by computing upper and lower interval bounds on the output. Our key insights are that these interval bounds can be obtained as the fixed-point solution to an IBP-inspired equilibrium equation, and furthermore, that this solution always exists and is unique when the layer obeys a certain parameterization. This fixed point can be interpreted as the result of applying IBP to an infinitely deep, weight-tied neural network, which may be of independent interest, as IBP bounds are typically unstable for deeper networks. Our empirical comparison reveals that models with IBP-MonDEQ layers can achieve comparable $\ell_{\infty}$ certified robustness to similarly-sized fully explicit networks.

Our world is full of asymmetries. Gravity and wind can make reaching a place easier than coming back. Social artifacts such as genealogy charts and citation graphs are inherently directed. In reinforcement learning and control, optimal goal-reaching strategies are rarely reversible (symmetrical). Distance functions supported on these asymmetrical structures are called quasimetrics. Despite their common appearance, little research has been done on the learning of quasimetrics. Our theoretical analysis reveals that a common class of learning algorithms, including unconstrained multilayer perceptrons (MLPs), provably fails to learn a quasimetric consistent with training data. In contrast, our proposed Poisson Quasimetric Embedding (PQE) is the first quasimetric learning formulation that both is learnable with gradient-based optimization and enjoys strong performance guarantees. Experiments on random graphs, social graphs, and offline Q-learning demonstrate its effectiveness over many common baselines.

Recent success in cooperative multi-agent reinforcement learning (MARL) relies on centralized training and policy sharing. Centralized training eliminates the issue of non-stationarity MARL yet induces large communication costs, and policy sharing is empirically crucial to efficient learning in certain tasks yet lacks theoretical justification. In this paper, we formally characterize a subclass of cooperative Markov games where agents exhibit a certain form of homogeneity such that policy sharing provably incurs no suboptimality. This enables us to develop the first consensus-based decentralized actor-critic method where the consensus update is applied to both the actors and the critics while ensuring convergence. We also develop practical algorithms based on our decentralized actor-critic method to reduce the communication cost during training, while still yielding policies comparable with centralized training.

We study neural contextual bandits, a general class of contextual bandits, where each context-action pair is associated with a raw feature vector, but the specific reward generating function is unknown. We propose a novel learning algorithm that transforms the raw feature vector using the last hidden layer of a deep ReLU neural network (deep representation learning), and uses an upper confidence bound (UCB) approach to explore in the last linear layer (shallow exploration). We prove that under standard assumptions, our proposed algorithm achieves $\tilde{O}(\sqrt{T})$ finite-time regret, where $T$ is the learning time horizon. Compared with existing neural contextual bandit algorithms, our approach is computationally much more efficient since it only needs to explore in the last layer of the deep neural network.

Recent self-supervised methods for image representation learning maximize the agreement between embedding vectors produced by encoders fed with different views of the same image. The main challenge is to prevent a collapse in which the encoders produce constant or non-informative vectors. We introduce VICReg (Variance-Invariance-Covariance Regularization), a method that explicitly avoids the collapse problem with two regularizations terms applied to both embeddings separately: (1) a term that maintains the variance of each embedding dimension above a threshold, (2) a term that decorrelates each pair of variables. Unlike most other approaches to the same problem, VICReg does not require techniques such as: weight sharing between the branches, batch normalization, feature-wise normalization, output quantization, stop gradient, memory banks, etc., and achieves results on par with the state of the art on several downstream tasks. In addition, we show that our variance regularization term stabilizes the training of other methods and leads to performance improvements.