ICLR 2024 Papers

Self-supervised learning converts raw perceptual data such as images to a compact space where simple Euclidean distances measure meaningful variations in data. In this paper, we extend this formulation by adding additional geometric structure to the embedding space by enforcing transformations of input space to correspond to simple (i.e., linear) transformations of embedding space. Specifically, in the contrastive learning setting, we introduce an equivariance objective and theoretically prove that its minima force augmentations on input space to correspond to rotations on the spherical embedding space. We show that merely combining our equivariant loss with a non-collapse term results in non-trivial representations, without requiring invariance to data augmentations. Optimal performance is achieved by also encouraging approximate invariance, where input augmentations correspond to small rotations. Our method, CARE: Contrastive Augmentation-induced Rotational Equivariance, leads to improved performance on downstream tasks and ensures sensitivity in embedding space to important variations in data (e.g., color) that standard contrastive methods do not achieve. Code is available at https://github.com/Sharut/CARE

This paper addresses the problem of anomaly detection in tabular data, which is usually implemented in an one-class classification setting where the training set only contains normal samples. Inspired by the success of masked image/language modeling in vision and natural language domains, we extend masked modeling methods to address this problem by capturing intrinsic correlations between features in training set. Thus, a sample deviate from such correlations is related to a high possibility of anomaly. To obtain multiple and diverse correlations, we propose a novel masking strategy which generates multiple masks by learning, and design a diversity loss to reduce the similarity of different masks. Extensive experiments show our method achieves state-of-the-art performance. We also discuss the interpretability from the perspective of each individual feature and correlations between features.

Deep representations have shown promising performance when transferred to downstream tasks in a black-box manner. Yet, their inherent lack of interpretability remains a significant challenge, as these features are often opaque to human understanding. In this paper, we propose Non-negative Contrastive Learning (NCL), a renaissance of Non-negative Matrix Factorization (NMF) aimed at deriving interpretable features. The power of NCL lies in its enforcement of non-negativity constraints on features, reminiscent of NMF's capability to extract features that align closely with sample clusters. NCL not only aligns mathematically well with an NMF objective but also preserves NMF's interpretability attributes, resulting in a more sparse and disentangled representation compared to standard contrastive learning (CL). Theoretically, we establish guarantees on the identifiability and downstream generalization of NCL. Empirically, we show that these advantages enable NCL to outperform CL significantly on feature disentanglement, feature selection, as well as downstream classification tasks. At last, we show that NCL can be easily extended to other learning scenarios and benefit supervised learning as well. Code is available at https://github.com/PKU-ML/non_neg.

Hypergraph Neural Networks (HGNNs) have recently attracted much attention and exhibited satisfactory performance due to their superiority in high-order correlation modeling. However, it is noticed that the high-order modeling capability of hypergraph also brings increased computation complexity, which hinders its practical industrial deployment. In practice, we find that one key barrier to the efficient deployment of HGNNs is the high-order structural dependencies during inference. In this paper, we propose to bridge the gap between the HGNNs and inference-efficient Multi-Layer Perceptron (MLPs) to eliminate the hypergraph dependency of HGNNs and thus reduce computational complexity as well as improve inference speed. Specifically, we introduce LightHGNN and LightHGNN$^+$ for fast inference with low complexity. LightHGNN directly distills the knowledge from teacher HGNNs to student MLPs via soft labels, and LightHGNN$^+$ further explicitly injects reliable high-order correlations into the student MLPs to achieve topology-aware distillation and resistance to over-smoothing. Experiments on eight hypergraph datasets demonstrate that even without hypergraph dependency, the proposed LightHGNNs can still achieve competitive or even better performance than HGNNs and outperform vanilla MLPs by $16.3$ on average. Extensive experiments on three graph datasets further show the average best performance of our LightHGNNs compared with all other methods. Experiments on synthetic hypergraphs with 5.5w vertices indicate LightHGNNs can run $100\times$ faster than HGNNs, showcasing their ability for latency-sensitive deployments.

The scientific scale-up of large language models (LLMs) necessitates a comprehensive understanding of their scaling properties. However, the existing literature on the scaling properties only yields an incomplete answer: optimization loss decreases predictably as the model size increases, in line with established scaling law; yet no scaling law for task has been established and the task performances are far from predictable during scaling. Task performances typically show minor gains on small models until they improve dramatically once models exceed a size threshold, exemplifying the ''emergent abilities''. In this study, we discover that small models, although they exhibit minor performance, demonstrate critical and consistent task performance improvements that are not captured by conventional evaluation strategies due to insufficient measurement resolution. To measure such improvements, we introduce PassUntil, an evaluation strategy with theoretically infinite resolution, through massive sampling in the decoding phase. With PassUntil, we conduct a quantitative investigation into the scaling law of task performance. The investigation contains two parts. Firstly, a strict task scaling law that is not conventionally known to exist, is identified, enhancing the predictability of task performances. Remarkably, we are able to predict the performance of the 2.4B model on code generation with merely 0.05\% deviation before training starts, which is the first systematic attempt to verify predictable scaling proposed by GPT-4's report. Secondly, underpinned by PassUntil, we are able to study emergent abilities quantitatively. We identify a kind of accelerated emergence whose scaling curve cannot be fitted by standard scaling law function and has a increasing speed. We then examine two hypothesis and imply that the ``multiple circuits hypothesis'' might be responsible for the accelerated emergence.

Sampling complex distributions is an important but difficult objective in various fields, including physics, chemistry, and statistics. An improvement of standard Monte Carlo (MC) methods, intensively used in particular in the context of disordered systems, is Parallel Tempering, also called replica exchange MC, in which a sequence of MC Markov chains at decreasing temperatures are run in parallel and can swap their configurations. In this work we apply the ideas of parallel tempering in the context of restricted Boltzmann machines (RBM), a paradigm of unsupervised architectures, capable to learn complex, multimodal distributions. Inspired by Deep Tempering, an approach introduced for deep belief networks, we show how to learn on top of the first RBM a stack of nested RBMs, using the representations of a RBM as ’data’ for the next one along the stack. In our Stacked Tempering approach the hidden configurations of a machine can be exchanged with the visible configurations of the next one in the stack. Replica exchanges between the different RBMs is facilitated by the increasingly clustered representations learnt by deeper RBMs, allowing for fast transitions between the different modes of the data distribution. Analytical calculations of mixing times in a simplified theoretical setting shed light on why Stacked Tempering works, and how hyperparameters, such as the aspect ratios of the RBMs and weight regularization should be chosen. We illustrate the efficiency of the Stacked Tempering method with respect to standard and replica exchange MC on several datasets: MNIST, in-silico Lattice Proteins, and the 2D-Ising model.

Existing Self-Supervised Learning (SSL) models for speech typically process speech signals at a fixed resolution of 20 milliseconds. This approach overlooks the varying informational content present at different resolutions in speech signals. In contrast, this paper aims to incorporate multi-resolution information into speech self-supervised representation learning. We introduce an SSL model that leverages a hierarchical Transformer architecture, complemented by HuBERT-style masked prediction objectives, to process speech at multiple resolutions. Experimental results indicate that the proposed model not only achieves more efficient inference but also exhibits superior or comparable performance to the original HuBERT model over various tasks. Specifically, significant performance improvements over the original HuBERT have been observed in fine-tuning experiments on the LibriSpeech speech recognition benchmark as well as in evaluations using the Speech Universal PERformance Benchmark (SUPERB) and Multilingual SUPERB (ML-SUPERB).

Sparse mixture of expert architectures (MoEs) scale model capacity without significant increases in training or inference costs. Despite their success, MoEs suffer from a number of issues: training instability, token dropping, inability to scale the number of experts, or ineffective finetuning. In this work, we propose Soft MoE, a fully-differentiable sparse Transformer that addresses these challenges, while maintaining the benefits of MoEs. Soft MoE performs an implicit soft assignment by passing different weighted combinations of all input tokens to each expert. As in other MoEs, experts in Soft MoE only process a subset of the (combined) tokens, enabling larger model capacity (and performance) at lower inference cost. In the context of visual recognition, Soft MoE greatly outperforms dense Transformers (ViTs) and popular MoEs (Tokens Choice and Experts Choice). Soft MoE scales well: Soft MoE Huge/14 with 128 experts in 16 MoE layers has over 40x more parameters than ViT Huge/14, with only 2% increased inference time, and substantially better quality.

Merging multi-exposure images is a common approach for obtaining high dynamic range (HDR) images, with the primary challenge being the avoidance of ghosting artifacts in dynamic scenes. Recent methods have proposed using deep neural networks for deghosting. However, the methods typically rely on sufficient data with HDR ground-truths, which are difficult and costly to collect. In this work, to eliminate the need for labeled data, we propose SelfHDR, a self-supervised HDR reconstruction method that only requires dynamic multi-exposure images during training. Specifically, SelfHDR learns a reconstruction network under the supervision of two complementary components, which can be constructed from multi-exposure images and focus on HDR color as well as structure, respectively. The color component is estimated from aligned multi-exposure images, while the structure one is generated through a structure-focused network that is supervised by the color component and an input reference (\eg, medium-exposure) image. During testing, the learned reconstruction network is directly deployed to predict an HDR image. Experiments on real-world images demonstrate our SelfHDR achieves superior results against the state-of-the-art self-supervised methods, and comparable performance to supervised ones. Codes are available at https://github.com/cszhilu1998/SelfHDR

As two prominent strategies for representation learning, Contrastive Learning (CL) and Masked Image Modeling (MIM) have witnessed significant progress. Previous studies have demonstrated the advantages of each approach in specific scenarios. CL, resembling supervised pre-training, excels at capturing longer-range global patterns and enhancing feature discrimination, while MIM is adept at introducing local and diverse attention across transformer layers. Considering the respective strengths, previous studies utilize feature distillation to inherit both discrimination and diversity. In this paper, we thoroughly examine previous feature distillation methods and observe that the increase in diversity mainly stems from asymmetric designs, which may in turn compromise the discrimination ability. To strike a balance between the two properties, we propose a simple yet effective strategy termed Hybrid Distill, which leverages both the CL and MIM teachers to jointly guide the student model. Hybrid Distill emulates the token relations of the MIM teacher at intermediate layers for diversity, while simultaneously distilling the final features of the CL teacher to enhance discrimination. A progressive redundant token masking strategy is employed to reduce the expenses associated with distillation and aid in preventing the model from converging to local optima. Experimental results demonstrate that Hybrid Distill achieves superior performance on various benchmark datasets.

This paper introduces Least Volume---a simple yet effective regularization inspired by geometric intuition---that can reduce the necessary number of latent dimensions needed by an autoencoder without requiring any prior knowledge of the intrinsic dimensionality of the dataset. We show that the Lipschitz continuity of the decoder is the key to making it work, provide a proof that PCA is just a linear special case of it, and reveal that it has a similar PCA-like importance ordering effect when applied to nonlinear models. We demonstrate the intuition behind the regularization on some pedagogical toy problems, and its effectiveness on several benchmark problems, including MNIST, CIFAR-10 and CelebA.

Recent studies have shown that Graph Transformers (GTs) can be effective for specific graph-level tasks. However, when it comes to node classification, training GTs remains challenging, especially in semi-supervised settings with a severe scarcity of labeled data. Our paper aims to address this research gap by focusing on semi-supervised node classification. To accomplish this, we develop a curriculum-enhanced attention distillation method that involves utilizing a Local GT teacher and a Global GT student. Additionally, we introduce the concepts of in-class and out-of-class and then propose two improvements, out-of-class entropy and top-k pruning, to facilitate the student's out-of-class exploration under the teacher's in-class guidance. Taking inspiration from human learning, our method involves a curriculum mechanism for distillation that initially provides strict guidance to the student and gradually allows for more out-of-class exploration by a dynamic balance. Extensive experiments show that our method outperforms many state-of-the-art approaches on seven public graph benchmarks, proving its effectiveness.

While large language models based on the transformer architecture have demonstrated remarkable in-context learning (ICL) capabilities, understandings of such capabilities are still in an early stage, where existing theory and mechanistic understanding focus mostly on simple scenarios such as learning simple function classes. This paper takes initial steps on understanding ICL in more complex scenarios, by studying learning with \emph{representations}. Concretely, we construct synthetic in-context learning problems with a compositional structure, where the label depends on the input through a possibly complex but \emph{fixed} representation function, composed with a linear function that \emph{differs} in each instance. By construction, the optimal ICL algorithm first transforms the inputs by the representation function, and then performs linear ICL on top of the transformed dataset. We show theoretically the existence of transformers that approximately implement such algorithms with mild depth and size. Empirically, we find trained transformers consistently achieve near-optimal ICL performance in this setting, and exhibit the desired dissection where lower layers transforms the dataset and upper layers perform linear ICL. Through extensive probing and a new pasting experiment, we further reveal several mechanisms within the trained transformers, such as concrete copying behaviors on both the inputs and the representations, linear ICL capability of the upper layers alone, and a post-ICL representation selection mechanism in a harder mixture setting. These observed mechanisms align well with our theory and may shed light on how transformers perform ICL in more realistic scenarios.

Contrastive learning typically matches pairs of related views among a number of unrelated negative views. Views can be generated (e.g. by augmentations) or be observed. We investigate matching when there are more than two related views which we call poly-view tasks, and derive new representation learning objectives using information maximization and sufficient statistics. We show that with unlimited computation, one should maximize the number of related views, and with a fixed compute budget, it is beneficial to decrease the number of unique samples whilst increasing the number of views of those samples. In particular, poly-view contrastive models trained for 128 epochs with batch size 256 outperform SimCLR trained for 1024 epochs at batch size 4096 on ImageNet1k, challenging the belief that contrastive models require large batch sizes and many training epochs.

Contrastive learning is a powerful self-supervised learning method, but we have a limited theoretical understanding of how it works and why it works. In this paper, we prove that contrastive learning with the standard InfoNCE loss is equivalent to spectral clustering on the similarity graph. Using this equivalence as the building block, we extend our analysis to the CLIP model and rigorously characterize how similar multi-modal objects are embedded together. Motivated by our theoretical insights, we introduce the Kernel-InfoNCE loss, incorporating mixtures of kernel functions that outperform the standard Gaussian kernel on several vision datasets.

Deep clustering, which learns representation and semantic clustering without labels information, poses a great challenge for deep learning-based approaches. Despite significant progress in recent years, most existing methods focus on uniformly distributed datasets, significantly limiting the practical applicability of their methods. In this paper, we first introduce a more practical problem setting named deep imbalanced clustering, where the underlying classes exhibit an imbalance distribution. To tackle this problem, we propose a novel pseudo-labeling-based learning framework. Our framework formulates pseudo-label generation as a progressive partial optimal transport problem, which progressively transports each sample to imbalanced clusters under prior distribution constraints, thus generating imbalance-aware pseudo-labels and learning from high-confident samples. In addition, we transform the initial formulation into an unbalanced optimal transport problem with augmented constraints, which can be solved efficiently by a fast matrix scaling algorithm. Experiments on various datasets, including a human-curated long-tailed CIFAR100, challenging ImageNet-R, and large-scale subsets of fine-grained iNaturalist2018 datasets, demonstrate the superiority of our method.

Dataset distillation aims to optimize a small set so that a model trained on the set achieves performance similar to that of a model trained on the full dataset. While many supervised methods have achieved remarkable success in distilling a large dataset into a small set of representative samples, however, they are not designed to produce a distilled dataset that can be effectively used to facilitate self-supervised pre-training. To this end, we propose a novel problem of distilling an unlabeled dataset into a set of small synthetic samples for efficient self-supervised learning (SSL). We first prove that a gradient of synthetic samples with respect to a SSL objective in naive bilevel optimization is \textit{biased} due to the randomness originating from data augmentations or masking for inner optimization. To address this issue, we propose to minimize the mean squared error (MSE) between a model's representations of the synthetic examples and their corresponding learnable target feature representations for the inner objective, which does not introduce any randomness. Our primary motivation is that the model obtained by the proposed inner optimization can mimic the \textit{self-supervised target model}. To achieve this, we also introduce the MSE between representations of the inner model and the self-supervised target model on the original full dataset for outer optimization. We empirically validate the effectiveness of our method on transfer learning. Our code is available at https://github.com/db-Lee/selfsup_dd

Large language models (LLMs) have demonstrated outstanding performance in various tasks, such as text summarization, text question-answering, and etc. While their performance is impressive, the computational footprint due to their vast number of parameters can be prohibitive. Existing solutions such as SparseGPT and Wanda attempt to alleviate this issue through weight pruning. However, their layer-wise approach results in significant perturbation to the model's output and requires meticulous hyperparameter tuning, such as the pruning rate, which can adversely affect overall model performance. To address this, this paper introduces a novel LLM pruning technique dubbed blockwise parameter-efficient sparsity allocation (BESA) by applying a blockwise reconstruction loss. In contrast to the typical layer-wise pruning techniques, BESA is characterized by two distinctive attributes: i) it targets the overall pruning error with respect to individual transformer blocks, and ii) it allocates layer-specific sparsity in a differentiable manner, both of which ensure reduced performance degradation after pruning. Our experiments show that BESA achieves state-of-the-art performance, efficiently pruning LLMs like LLaMA1, and LLaMA2 with 7B to 70B parameters on a single A100 GPU in just five hours. Code is available at [here](https://github.com/LinkAnonymous/BESA).

We focus on the self-supervised discovery of manipulation concepts that can be adapted and reassembled to address various robotic tasks. We propose that the decision to conceptualize a physical procedure should not depend on how we name it (semantics) but rather on the significance of the informativeness in its representation regarding the low-level physical state and state changes. We model manipulation concepts -- discrete symbols -- as generative and discriminative goals and derive metrics that can autonomously link them to meaningful sub-trajectories from noisy, unlabeled demonstrations. Specifically, we employ a trainable codebook containing encodings --symbols -- capable of synthesizing the end-state of a sub-trajectory given the current state (generative informativeness). Moreover, the encoding corresponding to a particular sub-trajectory should differentiate the state within and outside it and confidently predict the subsequent action based on the gradient of its discriminative score (discriminative informativeness). These metrics, which do not rely on human annotation, can be seamlessly integrated into a VQ-VAE framework, enabling the partitioning of demonstrations into semantically consistent sub-trajectories, fulfilling the purpose of discovering manipulation concepts and the corresponding (sub)-goal states. We evaluate the effectiveness of the learned concepts by training policies that utilize them as guidance, demonstrating superior performance compared to other baselines. Additionally, our discovered manipulation concepts compare favorably to human-annotated ones, while saving much manual effort. The code and trained models will be made public.