ICLR 2023 Papers

Learning energy-based models (EBMs) is known to be difficult especially on discrete data where gradient-based learning strategies cannot be applied directly. Although ratio matching is a sound method to learn discrete EBMs, it suffers from expensive computation and excessive memory requirements, thereby resulting in difficulties in learning EBMs on high-dimensional data. Motivated by these limitations, in this study, we propose ratio matching with gradient-guided importance sampling (RMwGGIS). Particularly, we use the gradient of the energy function w.r.t. the discrete data space to approximately construct the provably optimal proposal distribution, which is subsequently used by importance sampling to efficiently estimate the original ratio matching objective. We perform experiments on density modeling over synthetic discrete data, graph generation, and training Ising models to evaluate our proposed method. The experimental results demonstrate that our method can significantly alleviate the limitations of ratio matching, perform more effectively in practice, and scale to high-dimensional problems. Our implementation is available at https://github.com/divelab/RMwGGIS.

Large language models (LLMs) show impressive abilities via few-shot prompting. Commercialized APIs such as OpenAI GPT-3 further increase their use in real-world language applications. However, the crucial problem of how to improve the reliability of GPT-3 is still under-explored. While reliability is a broad and vaguely defined term, we decompose reliability into four main facets that correspond to the existing framework of ML safety and are well-recognized to be important: generalizability, social biases, calibration, and factuality. Our core contribution is to establish simple and effective prompts that improve GPT-3’s reliability as it: 1) generalizes out-of-distribution, 2) balances demographic distribution and uses natural language instructions to reduce social biases, 3) calibrates output probabilities, and 4) updates the LLM’s factual knowledge and reasoning chains. With appropriate prompts, GPT-3 is more reliable than smaller-scale supervised models on all these facets. We release all processed datasets, evaluation scripts, and model predictions. Our systematic empirical study not only sheds new insights on the reliability of prompting LLMs, but more importantly, our prompting strategies can help practitioners more reliably use LLMs like GPT-3.

By conditioning on natural language instructions, large language models (LLMs) have displayed impressive capabilities as general-purpose computers. However, task performance depends significantly on the quality of the prompt used to steer the model, and most effective prompts have been handcrafted by humans. Inspired by classical program synthesis and the human approach to prompt engineering, we propose Automatic Prompt Engineer (APE) for automatic instruction generation and selection. In our method, we treat the instruction as the "program," optimized by searching over a pool of instruction candidates proposed by an LLM in order to maximize a chosen score function. To evaluate the quality of the selected instruction, we evaluate the zero-shot performance of another LLM following the selected instruction. Experiments on 24 NLP tasks show that our automatically generated instructions outperform the prior LLM baseline by a large margin and achieve better or comparable performance to the instructions generated by human annotators on 21/24 tasks. We conduct extensive qualitative and quantitative analyses to explore the performance of APE. We show that APE-engineered prompts can be applied to steer models toward truthfulness and/or informativeness, as well as to improve few-shot learning performance by simply prepending them to standard in-context learning prompts.

We present Masked Frequency Modeling (MFM), a unified frequency-domain-based approach for self-supervised pre-training of visual models. Instead of randomly inserting mask tokens to the input embeddings in the spatial domain, in this paper, we shift the perspective to the frequency domain. Specifically, MFM first masks out a portion of frequency components of the input image and then predicts the missing frequencies on the frequency spectrum. Our key insight is that predicting masked components in the frequency domain is more ideal to reveal underlying image patterns rather than predicting masked patches in the spatial domain, due to the heavy spatial redundancy. Our findings suggest that with the right configuration of mask-and-predict strategy, both the structural information within high-frequency components and the low-level statistics among low-frequency counterparts are useful in learning good representations. For the first time, MFM demonstrates that, for both ViT and CNN, a simple non-Siamese framework can learn meaningful representations even using none of the following: (i) extra data, (ii) extra model, (iii) mask token. Experimental results on image classification and semantic segmentation, as well as several robustness benchmarks show the competitive performance and advanced robustness of MFM compared with recent masked image modeling approaches. Furthermore, we also comprehensively investigate the effectiveness of classical image restoration tasks for representation learning from a unified frequency perspective and reveal their intriguing relations with our MFM approach. Project page: https://www.mmlab-ntu.com/project/mfm/index.html.

Stochastic gradient descent (SGD) and its variants are considered as the de-facto methods to train deep neural networks (DNNs). While recent improvements to SGD mainly focus on the descent algorithm itself, few works pay attention to utilizing the historical solutions---as an iterative method, SGD has gone through substantial explorations before convergence. Recently, an interesting attempt is stochastic weight averaging (SWA), which significantly improves the generalization by simply averaging the solutions at the tail stage of training. In this paper, we realize that the averaging coefficients could be determined in a trainable manner and propose Trainable Weight Averaging (TWA), a novel optimization method in the reduced subspace spanned by historical solutions. TWA has much greater flexibility and can be applied to the head stage of training to achieve training efficiency while preserving good generalization capability. Further, we propose a distributed training scheme to resolve the memory burden of large-scale training with efficient parallel computation. In the extensive numerical experiments, (i) TWA achieves consistent improvements over SWA with less sensitivity to learning rate; (ii) applying TWA in the head stage of training largely speeds up the convergence, resulting in over $40\%$ time saving on CIFAR and $30\%$ on ImageNet with improved generalization compared with regular training.

The extent to which text-only language models (LMs) learn to represent the physical, non-linguistic world is an open question. Prior work has shown that pretrained LMs can be taught to ``understand'' visual inputs when the models' parameters are updated on image captioning tasks. We test a stronger hypothesis: that the conceptual representations learned by text-only models are functionally equivalent (up to a linear transformation) to those learned by models trained on vision tasks. Specifically, we show that the image representations from vision models can be transferred as continuous prompts to frozen LMs by training only a single linear projection. Using these to prompt the LM achieves competitive performance on captioning and visual question answering tasks compared to models that tune both the image encoder and text decoder (such as the MAGMA model). We compare three image encoders with increasing amounts of linguistic supervision seen during pretraining: BEIT (no linguistic information), NF-ResNET (lexical category information), and CLIP (full natural language descriptions). We find that all three encoders perform equally well at transferring visual property information to the language model (e.g., whether an animal is large or small), but that image encoders pretrained with linguistic supervision more saliently encode category information (e.g., distinguishing hippo vs.\ elephant) and thus perform significantly better on benchmark language-and-vision tasks. Our results indicate that LMs encode conceptual information structurally similarly to vision-based models, even those that are solely trained on images.

Denoising diffusion models are a popular class of generative models providing state-of-the-art results in many domains. One adds gradually noise to data using a diffusion to transform the data distribution into a Gaussian distribution. Samples from the generative model are then obtained by simulating an approximation of the time-reversal of this diffusion initialized by Gaussian samples. Practically, the intractable score terms appearing in the time-reversed process are approximated using score matching techniques. We explore here a similar idea to sample approximately from unnormalized probability density functions and estimate their normalizing constants. We consider a process where the target density diffuses towards a Gaussian. Denoising Diffusion Samplers (DDS) are obtained by approximating the corresponding time-reversal. While score matching is not applicable in this context, we can leverage many of the ideas introduced in generative modeling for Monte Carlo sampling. Existing theoretical results from denoising diffusion models also provide theoretical guarantees for DDS. We discuss the connections between DDS, optimal control and Schr\"odinger bridges and finally demonstrate DDS experimentally on a variety of challenging sampling tasks.

For long-tailed classification tasks, most works often pretrain a big model on a large-scale (unlabeled) dataset, and then fine-tune the whole pretrained model for adapting to long-tailed data. Though promising, fine-tuning the whole pretrained model tends to suffer from high cost in computation and deployment of different models for different tasks, as well as weakened generalization capability for overfitting to certain features of long-tailed data. To alleviate these issues, we propose an effective Long-tailed Prompt Tuning (LPT) method for long-tailed classification tasks. LPT introduces several trainable prompts into a frozen pretrained model to adapt it to long-tailed data. For better effectiveness, we divide prompts into two groups: 1) a shared prompt for the whole long-tailed dataset to learn general features and to adapt a pretrained model into the target long-tailed domain; and 2) group-specific prompts to gather group-specific features for the samples which have similar features and also to empower the pretrained model with fine-grained discrimination ability. Then we design a two-phase training paradigm to learn these prompts. In the first phase, we train the shared prompt via conventional supervised prompt tuning to adapt a pretrained model to the desired long-tailed domain. In the second phase, we use the learnt shared prompt as query to select a small best matched set for a group of similar samples from the group-specific prompt set to dig the common features of these similar samples, and then optimize these prompts with a dual sampling strategy and the asymmetric Gaussian Clouded Logit loss. By only fine-tuning a few prompts while fixing the pretrained model, LPT can reduce training cost and deployment cost by storing a few prompts, and enjoys a strong generalization ability of the pretrained model. Experiments show that on various long-tailed benchmarks, with only $\sim$1.1\% extra trainable parameters, LPT achieves comparable or higher performance than previous whole model fine-tuning methods, and is more robust to domain-shift.

Reinforcement learning in partially observed Markov decision processes (POMDPs) faces two challenges. (i) It often takes the full history to predict the future, which induces a sample complexity that scales exponentially with the horizon. (ii) The observation and state spaces are often continuous, which induces a sample complexity that scales exponentially with the extrinsic dimension. Addressing such challenges requires learning a minimal but sufficient representation of the observation and state histories by exploiting the structure of the POMDP. To this end, we propose a reinforcement learning algorithm named Represent to Control (RTC), which learns the representation at two levels while optimizing the policy.~(i) For each step, RTC learns to represent the state with a low-dimensional feature, which factorizes the transition kernel. (ii) Across multiple steps, RTC learns to represent the full history with a low-dimensional embedding, which assembles the per-step feature. We integrate (i) and (ii) in a unified framework that allows a variety of estimators (including maximum likelihood estimators and generative adversarial networks). For a class of POMDPs with a low-rank structure in the transition kernel, RTC attains an $O(1/\epsilon^2)$ sample complexity that scales polynomially with the horizon and the intrinsic dimension (that is, the rank). Here $\epsilon$ is the optimality gap. To our best knowledge, RTC is the first sample-efficient algorithm that bridges representation learning and policy optimization in POMDPs with infinite observation and state spaces.

Despite the popularity and efficacy of knowledge distillation, there is limited understanding of why it helps. In order to study the generalization behavior of a distilled student, we propose a new theoretical framework that leverages supervision complexity: a measure of alignment between teacher-provided supervision and the student's neural tangent kernel. The framework highlights a delicate interplay among the teacher's accuracy, the student's margin with respect to the teacher predictions, and the complexity of the teacher predictions. Specifically, it provides a rigorous justification for the utility of various techniques that are prevalent in the context of distillation, such as early stopping and temperature scaling. Our analysis further suggests the use of online distillation, where a student receives increasingly more complex supervision from teachers in different stages of their training. We demonstrate efficacy of online distillation and validate the theoretical findings on a range of image classification benchmarks and model architectures.

The sequence alignment problem is one of the most fundamental problems in bioinformatics and a plethora of methods were devised to tackle it. Here we introduce BetaAlign, a methodology for aligning sequences using an NLP approach. BetaAlign accounts for the possible variability of the evolutionary process among different datasets by using an ensemble of transformers, each trained on millions of samples generated from a different evolutionary model. Our approach leads to alignment accuracy that is similar and often better than commonly used methods, such as MAFFT, DIALIGN, ClustalW, T-Coffee, PRANK, and MUSCLE.

Dynamic graphs are ubiquitous across disciplines where observations usually change over time. Regressions on dynamic graphs often contribute to diverse critical tasks, such as climate early-warning and traffic controlling. Existing homophily Graph Neural Networks (GNNs) adopt physical connections or feature similarity as adjacent matrix to perform node-level aggregations. However, on dynamic graphs with diverse node-wise relations, exploiting a pre-defined fixed topology for message passing inevitably leads to the aggregations of target-deviated neighbors. We designate such phenomenon as the topology-task discordance, which naturally challenges the homophily assumption. In this work, we revisit node-wise relationships and explore novel homophily measurements on dynamic graphs with both signs and distances, capturing multiple node-level spatial relations and temporal evolutions. We discover that advancing homophily aggregations to signed target-oriented message passing can effectively resolve the discordance and promote aggregation capacity. Therefore, a GReTo is proposed, which performs signed message passing in immediate neighborhood, and exploits both local environments and target awareness to realize high-order message propagation. Empirically, our solution achieves significant improvements against best baselines, notably improving 24.79% on KnowAir and 3.60% on Metr-LA.

Implicit processes (IPs) are a generalization of Gaussian processes (GPs). IPs may lack a closed-form expression but are easy to sample from. Examples include, among others, Bayesian neural networks or neural samplers. IPs can be used as priors over functions, resulting in flexible models with well-calibrated prediction uncertainty estimates. Methods based on IPs usually carry out function-space approximate inference, which overcomes some of the difficulties of parameter-space approximate inference. Nevertheless, the approximations employed often limit the expressiveness of the final model, resulting, e.g., in a Gaussian predictive distribution, which can be restrictive. We propose here a multi-layer generalization of IPs called the Deep Variational Implicit process (DVIP). This generalization is similar to that of deep GPs over GPs, but it is more flexible due to the use of IPs as the prior distribution over the latent functions. We describe a scalable variational inference algorithm for training DVIP and show that it outperforms previous IP-based methods and also deep GPs. We support these claims via extensive regression and classification experiments. We also evaluate DVIP on large datasets with up to several million data instances to illustrate its good scalability and performance.

Visual attention does not always capture the essential object representation desired for robust predictions. Attention modules tend to underline not only the target object but also the common co-occurring context that the module thinks helpful in the training. The problem is rooted in the confounding effect of the context leading to incorrect causalities between objects and predictions, which is further exacerbated by visual attention. In this paper, to learn causal object features robust for contextual bias, we propose a novel attention module named Interventional Dual Attention (IDA) for visual recognition. Specifically, IDA adopts two attention layers with multiple sampling intervention, which compensates the attention against the confounder context. Note that our method is model-agnostic and thus can be implemented on various backbones. Extensive experiments show our model obtains significant improvements in classification and detection with lower computation. In particular, we achieve the state-of-the-art results in multi-label classification on MS-COCO and PASCAL-VOC.

We present a novel graph neural network we call AgentNet, which is designed specifically for graph-level tasks. AgentNet is inspired by sublinear algorithms, featuring a computational complexity that is independent of the graph size. The architecture of AgentNet differs fundamentally from the architectures of traditional graph neural networks. In AgentNet, some trained \textit{neural agents} intelligently walk the graph, and then collectively decide on the output. We provide an extensive theoretical analysis of AgentNet: We show that the agents can learn to systematically explore their neighborhood and that AgentNet can distinguish some structures that are even indistinguishable by 2-WL. Moreover, AgentNet is able to separate any two graphs which are sufficiently different in terms of subgraphs. We confirm these theoretical results with synthetic experiments on hard-to-distinguish graphs and real-world graph classification tasks. In both cases, we compare favorably not only to standard GNNs but also to computationally more expensive GNN extensions.

Recently, a few self-supervised representation learning (SSL) methods have outperformed the ImageNet classification pre-training for vision tasks such as object detection. However, its effects on 3D human body pose and shape estimation (3DHPSE) are open to question, whose target is fixed to a unique class, the human, and has an inherent task gap with SSL. We empirically study and analyze the effects of SSL and further compare it with other pre-training alternatives for 3DHPSE. The alternatives are 2D annotation-based pre-training and synthetic data pre-training, which share the motivation of SSL that aims to reduce the labeling cost. They have been widely utilized as a source of weak-supervision or fine-tuning, but have not been remarked as a pre-training source. SSL methods underperform the conventional ImageNet classification pre-training on multiple 3DHPSE benchmarks by 7.7% on average. In contrast, despite a much less amount of pre-training data, the 2D annotation-based pre-training improves accuracy on all benchmarks and shows faster convergence during fine-tuning. Our observations challenge the naive application of the current SSL pre-training to 3DHPSE and relight the value of other data types in the pre-training aspect.

As a typical self-supervised learning strategy, Masked Image Modeling (MIM) is driven by recovering all masked patches from visible ones. However, patches from the same image are highly correlated and it is redundant to reconstruct all the masked patches. We find that this redundancy is neglected by existing MIM based methods and causes non-negligible overheads in computation that do not necessarily benefit self-supervised representation. In this paper, we present a novel approach named PCAE, short for Progressively Compressed AutoEncoder, to address the redundant reconstruction issue by progressively compacting tokens and only retaining necessary information for forward propagation and reconstruction. In particular, we identify those redundant tokens in an image via a simple yet effective similarity metric between each token with the mean of the token sequence. Those redundant tokens that other ones can probably represent are progressively dropped accordingly during the forward propagation, and importantly, we only focus on reconstructing these retained tokens. As a result, we are able to achieve a better trade-off between performance and efficiency for pre-training. Besides, benefitting from the flexible strategy, PCAE can be also directly employed for downstream fine-tuning tasks and enable scalable deployment. Experiments show that PCAE achieves comparable performance to MAE with only 1/8 GPU days. The code is available at https://github.com/caddyless/PCAE/.

The success of deep learning heavily relies on large-scale data with comprehensive labels, which is more expensive and time-consuming to fetch in 3D compared to 2D images or natural languages. This promotes the potential of utilizing models pretrained with data more than 3D as teachers for cross-modal knowledge transferring. In this paper, we revisit masked modeling in a unified fashion of knowledge distillation, and we show that foundational Transformers pretrained with 2D images or natural languages can help self-supervised 3D representation learning through training Autoencoders as Cross-Modal Teachers (ACT). The pretrained Transformers are transferred as cross-modal 3D teachers using discrete variational autoencoding self-supervision, during which the Transformers are frozen with prompt tuning for better knowledge inheritance. The latent features encoded by the 3D teachers are used as the target of masked point modeling, wherein the dark knowledge is distilled to the 3D Transformer students as foundational geometry understanding. Our ACT pretrained 3D learner achieves state-of-the-art generalization capacity across various downstream benchmarks, e.g., 88.21% overall accuracy on ScanObjectNN. Codes have been released at https://github.com/RunpeiDong/ACT.

The graph Transformer emerges as a new architecture and has shown superior performance on various graph mining tasks. In this work, we observe that existing graph Transformers treat nodes as independent tokens and construct a single long sequence composed of all node tokens so as to train the Transformer model, causing it hard to scale to large graphs due to the quadratic complexity on the number of nodes for the self-attention computation. To this end, we propose a Neighborhood Aggregation Graph Transformer (NAGphormer) that treats each node as a sequence containing a series of tokens constructed by our proposed Hop2Token module. For each node, Hop2Token aggregates the neighborhood features from different hops into different representations and thereby produces a sequence of token vectors as one input. In this way, NAGphormer could be trained in a mini-batch manner and thus could scale to large graphs. Moreover, we mathematically show that as compared to a category of advanced Graph Neural Networks (GNNs), the decoupled Graph Convolutional Network, NAGphormer could learn more informative node representations from the multi-hop neighborhoods. Extensive experiments on benchmark datasets from small to large are conducted to demonstrate that NAGphormer consistently outperforms existing graph Transformers and mainstream GNNs. Code is available at https://github.com/JHL-HUST/NAGphormer.