ICLR 2024 Papers

Uniformity plays an important role in evaluating learned representations, providing insights into self-supervised learning. In our quest for effective uniformity metrics, we pinpoint four principled properties that such metrics should possess. Namely, an effective uniformity metric should remain invariant to instance permutations and sample replications while accurately capturing feature redundancy and dimensional collapse. Surprisingly, we find that the uniformity metric proposed by \citet{Wang2020UnderstandingCR} fails to satisfy the majority of these properties. Specifically, their metric is sensitive to sample replications, and can not account for feature redundancy and dimensional collapse correctly. To overcome these limitations, we introduce a new uniformity metric based on the Wasserstein distance, which satisfies all the aforementioned properties. Integrating this new metric in existing self-supervised learning methods effectively mitigates dimensional collapse and consistently improves their performance on downstream tasks involving CIFAR-10 and CIFAR-100 datasets. Code is available at \url{https://github.com/statsle/WassersteinSSL}.

Scaling of deep neural networks, especially Transformers, is pivotal for their surging performance and has further led to the emergence of sophisticated reasoning capabilities in foundation models. Such scaling generally requires training large models from scratch with random initialization, failing to leverage the knowledge acquired by their smaller counterparts, which are already resource-intensive to obtain. To tackle this inefficiency, we present $\textbf{L}$ossl$\textbf{E}$ss $\textbf{MO}$del Expansio$\textbf{N}$ (LEMON), a recipe to initialize scaled models using the weights of their smaller but pre-trained counterparts. This is followed by model training with an optimized learning rate scheduler tailored explicitly for the scaled models, substantially reducing the training time compared to training from scratch. Notably, LEMON is versatile, ensuring compatibility with various network structures, including models like Vision Transformers and BERT. Our empirical results demonstrate that LEMON reduces computational costs by 56.7\% for Vision Transformers and 33.2\% for BERT when compared to training from scratch.

A key property of deep neural networks (DNNs) is their ability to learn new features during training. This intriguing aspect of deep learning stands out most clearly in recently reported Grokking phenomena. While mainly reflected as a sudden increase in test accuracy, Grokking is also believed to be a beyond lazy-learning/Gaussian Process (GP) phenomenon involving feature learning. Here we apply a recent development in the theory of feature learning, the adaptive kernel approach, to two teacher-student models with cubic-polynomial and modular addition teachers. We provide analytical predictions on feature learning and Grokking properties of these models and demonstrate a mapping between Grokking and the theory of phase transitions. We show that after Grokking, the state of the DNN is analogous to the mixed phase following a first-order phase transition. In this mixed phase, the DNN generates useful internal representations of the teacher that are sharply distinct from those before the transition.

Machine learning algorithms learned from data with skewed distributions usually suffer from poor generalization, especially when minority classes matter as much as, or even more than majority ones. This is more challenging on class-balanced data that has some hidden imbalanced subpopulations, since prevalent techniques mainly conduct class-level calibration and cannot perform subpopulation-level adjustments without subpopulation annotations. Regarding implicit subpopulation imbalance, we reveal that the key to alleviating the detrimental effect lies in effective subpopulation discovery with proper rebalancing. We then propose a novel subpopulation-imbalanced learning method called Scatter and HarmonizE (SHE). Our method is built upon the guiding principle of optimal data partition, which involves assigning data to subpopulations in a manner that maximizes the predictive information from inputs to labels. With theoretical guarantees and empirical evidences, SHE succeeds in identifying the hidden subpopulations and encourages subpopulation-balanced predictions. Extensive experiments on various benchmark datasets show the effectiveness of SHE.

Self-supervised heterogeneous graph learning has achieved promising results in various real applications, but it still suffers from the following issues: (i) meta-paths can be employed to capture the homophily in the heterogeneous graph, but meta-paths are human-defined, requiring substantial expert knowledge and computational costs; and (ii) the heterogeneity in the heterogeneous graph is usually underutilized, leading to the loss of task-related information. To solve these issues, this paper proposes to capture both homophily and heterogeneity in the heterogeneous graph without pre-defined meta-paths. Specifically, we propose to learn a self-expressive matrix to capture the homophily from the subspace and nearby neighbors. Meanwhile, we propose to capture the heterogeneity by aggregating the information of nodes from different types. We further design a consistency loss and a specificity loss, respectively, to extract the consistent information between homophily and heterogeneity and to preserve their specific task-related information. We theoretically analyze that the learned homophilous representations exhibit the grouping effect to capture the homophily, and considering both homophily and heterogeneity introduces more task-related information. Extensive experimental results verify the superiority of the proposed method on different downstream tasks.

Large Language models (LLMs) have demonstrated impressive in-context learning (ICL) capabilities, where a LLM makes predictions for a given test input together with a few input-output pairs (demonstrations). Nevertheless, the inclusion of demonstrations poses a challenge, leading to a quadratic increase in the computational overhead of the self-attention mechanism. Existing solutions attempt to condense lengthy demonstrations into compact vectors. However, they often require task-specific retraining or compromise LLM's in-context learning performance. To mitigate these challenges, we present Meta Demonstration Distillation (MEND), where a language model learns to distill any lengthy demonstrations into vectors without retraining for a new downstream task. We exploit the knowledge distillation to enhance alignment between MEND and MEND, achieving both efficiency and effectiveness concurrently. MEND is endowed with the meta-knowledge of distilling demonstrations through a two-stage training process, which includes meta-distillation pretraining and fine-tuning. Comprehensive evaluations across seven diverse ICL settings using decoder-only (GPT-2) and encoder-decoder (T5) attest to MEND's prowess. It not only matches but often outperforms the Vanilla ICL as well as other state-of-the-art distillation models, while significantly reducing the computational demands. This innovation promises enhanced scalability and efficiency for the practical deployment of large language models.

Overfitting negatively impacts the generalization ability of deep neural networks (DNNs) in both natural and adversarial training. Existing methods struggle to consistently address different types of overfitting, typically designing strategies that focus separately on either natural or adversarial patterns. In this work, we adopt a unified perspective by solely focusing on natural patterns to explore different types of overfitting. Specifically, we examine the memorization effect in DNNs and reveal a shared behaviour termed over-memorization, which impairs their generalization capacity. This behaviour manifests as DNNs suddenly becoming high-confidence in predicting certain training patterns and retaining a persistent memory for them. Furthermore, when DNNs over-memorize an adversarial pattern, they tend to simultaneously exhibit high-confidence prediction for the corresponding natural pattern. These findings motivate us to holistically mitigate different types of overfitting by hindering the DNNs from over-memorization training patterns. To this end, we propose a general framework, $\textit{Distraction Over-Memorization}$ (DOM), which explicitly prevents over-memorization by either removing or augmenting the high-confidence natural patterns. Extensive experiments demonstrate the effectiveness of our proposed method in mitigating overfitting across various training paradigms.

Fact verification aims to verify a claim using evidence from a trustworthy knowledge base. To address this challenge, algorithms must produce features for every claim that are both semantically meaningful, and compact enough to find a semantic alignment with the source information. In contrast to previous work, which tackled the alignment problem by learning over annotated corpora of claims and their corresponding labels, we propose SFAVEL ($\underline{S}$elf-supervised $\underline{Fa}$ct $\underline{Ve}$rification via $\underline{L}$anguage Model Distillation), a novel unsupervised pretraining framework that leverages pre-trained language models to distil self-supervised features into high-quality claim-fact alignments without the need for annotations. This is enabled by a novel contrastive loss function that encourages features to attain high-quality claim and evidence alignments whilst preserving the semantic relationships across the corpora. Notably, we present results that achieve a new state-of-the-art on FB15k-237 (+5.3\% Hits@1) and FEVER (+8\% accuracy) with linear evaluation.

The infinitely wide neural network has been proven a useful and manageable mathematical model that enables the understanding of many phenomena appearing in deep learning. One example is the convergence of random deep networks to Gaussian processes that enables a rigorous analysis of the way the choice of activation function and network weights impacts the training dynamics. In this paper, we extend the seminal proof of Matthews et al., 2018 to a larger class of initial weight distributions (which we call pseudo-iid), including the established cases of iid and orthogonal weights, as well as the emerging low-rank and structured sparse settings celebrated for their computational speed-up benefits. We show that fully-connected and convolutional networks initialised with pseudo-iid distributions are all effectively equivalent up to their variance. Using our results, one can identify the Edge of Chaos for a broader class of neural networks and tune them at criticality in order to enhance their training. Moreover, they enable the posterior distribution of Bayesian Neural Networks to be tractable across these various initialization schemes.

The massive interest in deep neural networks (DNNs) for both computer vision and natural language processing has been sparked by the growth in computational power. However, this led to an increase in the memory footprint, to a point where it can be challenging to simply load a model on commodity devices such as mobile phones. To address this limitation, quantization is a favored solution as it maps high precision tensors to a low precision, memory efficient format. In terms of memory footprint reduction, its most effective variants are based on codebooks. These methods, however, suffer from two limitations. First, they either define a single codebook for each tensor, or use a memory-expensive mapping to multiple codebooks. Second, gradient descent optimization of the mapping favors jumps toward extreme values, hence not defining a proximal search. In this work, we propose to address these two limitations. First, we initially group similarly distributed neurons and leverage the re-ordered structure to either apply different scale factors to the different groups, or map weights that fall in these groups to several codebooks, without any mapping overhead. Second, stemming from this initialization, we propose a joint learning of the codebook and weight mappings that bears similarities with recent gradient-based post-training quantization techniques. Third, drawing estimation from straight-through estimation techniques, we introduce a novel gradient update definition to enable a proximal search of the codebooks and their mappings. The proposed jointly learnable codebooks and mappings (JLCM) method allows a very efficient approximation of any DNN: as such, a Llama 7B can be compressed down to 2Go and loaded on 5-year-old smartphones.

A novel clustering algorithm for orderable data, called unsupervised order learning (UOL), is proposed in this paper. First, we develop the ordered $k$-means to group objects into ordered clusters by reducing the deviation of an object from consecutive clusters. Then, we train a network to construct an embedding space, in which objects are sorted compactly along a chain of line segments, determined by the cluster centroids. We alternate the clustering and the network training until convergence. Moreover, we perform unsupervised rank estimation via a simple nearest neighbor search in the embedding space. Extensive experiments on various orderable datasets demonstrate that UOL provides reliable ordered clustering results and decent rank estimation performances with no supervision. The source codes are available at https://github.com/seon92/UOL.

Recurrent Spiking Neural Networks (RSNNs) have emerged as a computationally efficient and brain-inspired machine learning model. The design of sparse RSNNs with fewer neurons and synapses helps reduce the computational complexity of RSNNs. Traditionally, sparse SNNs are obtained by first training a dense and complex SNN for a target task and, next, eliminating neurons with low activity (activity-based pruning) while maintaining task performance. In contrast, this paper presents a task-agnostic methodology for designing sparse RSNNs by pruning an untrained (arbitrarily initialized) large model. We introduce a novel Lyapunov Noise Pruning (LNP) algorithm that uses graph sparsification methods and utilizes Lyapunov exponents to design a stable sparse RSNN from an untrained RSNN. We show that the LNP can leverage diversity in neuronal timescales to design a sparse Heterogeneous RSNN (HRSNN). Further, we show that the same sparse HRSNN model can be trained for different tasks, such as image classification and time-series prediction. The experimental results show that, in spite of being task-agnostic, LNP increases computational efficiency (fewer neurons and synapses) and prediction performance of RSNNs compared to traditional activity-based pruning of trained dense models.

Node representation learning, such as Graph Neural Networks (GNNs), has become one of the important learning methods in machine learning, and the demand for reliable explanation generation is growing. Despite extensive research on explanation generation for supervised node representation learning, explaining unsupervised models has been less explored. To address this gap, we propose a method for generating counterfactual (CF) explanations in unsupervised node representation learning, aiming to identify the most important subgraphs that cause a significant change in the $k$-nearest neighbors of a node of interest in the learned embedding space upon perturbation. The $k$-nearest neighbor-based CF explanation method provides simple, yet pivotal, information for understanding unsupervised downstream tasks, such as top-$k$ link prediction and clustering. Furthermore, we introduce a Monte Carlo Tree Search (MCTS)-based explainability method for generating expressive CF explanations for **U**nsupervised **N**ode **R**epresentation learning methods, which we call **UNR-Explainer**. The proposed method demonstrates improved performance on six datasets for both unsupervised GraphSAGE and DGI.

The rapid evolution of Multi-modality Large Language Models (MLLMs) has catalyzed a shift in computer vision from specialized models to general-purpose foundation models. Nevertheless, there is still an inadequacy in assessing the abilities of MLLMs on **low-level visual perception and understanding**. To address this gap, we present **Q-Bench**, a holistic benchmark crafted to systematically evaluate potential abilities of MLLMs on three realms: low-level visual perception, low-level visual description, and overall visual quality assessment. **_a)_** To evaluate the low-level **_perception_** ability, we construct the **LLVisionQA** dataset, consisting of 2,990 diverse-sourced images, each equipped with a human-asked question focusing on its low-level attributes. We then measure the correctness of MLLMs on answering these questions. **_b)_** To examine the **_description_** ability of MLLMs on low-level information, we propose the **LLDescribe** dataset consisting of long expert-labelled *golden* low-level text descriptions on 499 images, and a GPT-involved comparison pipeline between outputs of MLLMs and the *golden* descriptions. **_c)_** Besides these two tasks, we further measure their visual quality **_assessment_** ability to align with human opinion scores. Specifically, we design a softmax-based strategy that enables MLLMs to predict *quantifiable* quality scores, and evaluate them on various existing image quality assessment (IQA) datasets. Our evaluation across the three abilities confirms that MLLMs possess preliminary low-level visual skills. However, these skills are still unstable and relatively imprecise, indicating the need for specific enhancements on MLLMs towards these abilities. We hope that our benchmark can encourage the research community to delve deeper to discover and enhance these untapped potentials of MLLMs.

Large language models (LLMs) excel in most NLP tasks but also require expensive cloud servers for deployment due to their size, while smaller models that can be deployed on lower cost (e.g., edge) devices, tend to lag behind in terms of response quality. Therefore in this work we propose a hybrid inference approach which combines their respective strengths to save cost and maintain quality. Our approach uses a router that assigns queries to the small or large model based on the predicted query difficulty and the desired quality level. The desired quality level can be tuned dynamically at test time to seamlessly trade quality for cost as per the scenario requirements. In experiments our approach allows us to make up to 40% fewer calls to the large model, with no drop in response quality.