ICLR 2023 Papers

Existing feature distillation methods commonly adopt the One-to-one Representation Matching between any pre-selected teacher-student layer pair. In this paper, we present $N$-to-$O$ne $R$epresentation $M$atching (NORM), a new two-stage knowledge distillation method, which relies on a simpleFeature Transform (FT) module consisting of two linear layers. In view of preserving the intact information learnt by the teacher network, during training, our FT module is merely inserted after the last convolutional layer of the student network. The first linear layer projects the student representation to a feature space having $N$ times feature channels than the teacher representation from the last convolutional layer, and the second linear layer contracts the expanded output back to the original feature space. By sequentially splitting the expanded student representation into $N$ non-overlapping feature segments having the same number of feature channels as the teacher's, they can be readily forced to approximate the intact teacher representation simultaneously, formulating a novel many-to-one representation matching mechanism conditioned on a single teacher-student layer pair. After training, such an FT module will be naturally merged into the subsequent fully connected layer thanks to its linear property, introducing no extra parameters or architectural modifications to the student network at inference. Extensive experiments on different visual recognition benchmarks demonstrate the leading performance of our method. For instance, the ResNet18|MobileNet|ResNet50-1/4 model trained by NORM reaches 72.14%|74.26%|68.03% top-1 accuracy on the ImageNet dataset when using a pre-trained ResNet34|ResNet50|ResNet50 model as the teacher, achieving an absolute improvement of 2.01%|4.63%|3.03% against the individually trained counterpart. Code is available at https://github.com/OSVAI/NORM.

Neural fields are receiving increased attention as a geometric representation due to their ability to compactly store detailed and smooth shapes and easily undergo topological changes. Compared to classic geometry representations, however, neural representations do not allow the user to exert intuitive control over the shape. Motivated by this, we leverage boundary sensitivity to express how perturbations in parameters move the shape boundary. This allows us to interpret the effect of each learnable parameter and study achievable deformations. With this, we perform geometric editing: finding a parameter update that best approximates a globally prescribed deformation. Prescribing the deformation only locally allows the rest of the shape to change according to some prior, such as semantics or deformation rigidity. Our method is agnostic to the model and its training and updates the NN in-place. Furthermore, we show how boundary sensitivity helps to optimize and constrain objectives (such as surface area and volume), which are difficult to compute without first converting to another representation, such as a mesh.

Ensembling has proven to be a powerful technique for boosting model performance, uncertainty estimation, and robustness in supervised learning. Advances in self-supervised learning (SSL) enable leveraging large unlabeled corpora for state-of-the-art few-shot and supervised learning performance. In this paper, we explore how ensemble methods can improve recent SSL techniques by developing a framework that permits data-dependent weighted cross-entropy losses. We refrain from ensembling the representation backbone; this choice yields an efficient ensemble method that incurs a small training cost and requires no architectural changes or computational overhead to downstream evaluation. The effectiveness of our method is demonstrated with two state-of-the-art SSL methods, DINO (Caron et al., 2021) and MSN (Assran et al., 2022). Our method outperforms both in multiple evaluation metrics on ImageNet-1K, particularly in the few-shot setting. We explore several weighting schemes and find that those which increase the diversity of ensemble heads lead to better downstream evaluation results. Thorough experiments yield improved prior art baselines which our method still surpasses; e.g., our overall improvement with MSN ViT-B/16 is 3.9 p.p. for 1-shot learning.

Contrastive learning is a powerful framework for learning self-supervised representations that generalize well to downstream supervised tasks. We show that multiple existing contrastive learning methods can be reinterpeted as learning kernel functions that approximate a fixed *positive-pair kernel*. We then prove that a simple representation obtained by combining this kernel with PCA provably minimizes the worst-case approximation error of linear predictors, under a straightforward assumption that positive pairs have similar labels. Our analysis is based on a decomposition of the target function in terms of the eigenfunctions of a positive-pair Markov chain, and a surprising equivalence between these eigenfunctions and the output of Kernel PCA. We give generalization bounds for downstream linear prediction using our kernel PCA representation, and show empirically on a set of synthetic tasks that applying kernel PCA to contrastive learning models can indeed approximately recover the Markov chain eigenfunctions, although the accuracy depends on the kernel parameterization as well as on the augmentation strength.

The ideally disentangled latent space in GAN involves the global representation of latent space using semantic attribute coordinates. In other words, in this disentangled space, there exists the global semantic basis as a vector space where each basis component describes one attribute of generated images. In this paper, we propose an unsupervised method for finding this global semantic basis in the intermediate latent space in GANs. This semantic basis represents sample-independent meaningful perturbations that change the same semantic attribute of an image on the entire latent space. The proposed global basis, called Fréchet basis, is derived by introducing Fréchet mean to the local semantic perturbations in a latent space. Fréchet basis is discovered in two stages. First, the global semantic subspace is discovered by the Fréchet mean in the Grassmannian manifold of the local semantic subspaces. Second, Fréchet basis is found by optimizing a basis of the semantic subspace via the Fréchet mean in the Special Orthogonal Group. Experimental results demonstrate that Fréchet basis provides better semantic factorization and robustness compared to the previous methods. Moreover, we suggest the basis refinement scheme for the previous methods. The quantitative experiments show that the refined basis achieves better semantic factorization while constrained on the same semantic subspace given by the previous method.

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.

Comparing learned neural representations in neural networks is a challenging but important problem, which has been approached in different ways. The Centered Kernel Alignment (CKA) similarity metric, particularly its linear variant, has recently become a popular approach and has been widely used to compare representations of a network's different layers, of architecturally similar networks trained differently, or of models with different architectures trained on the same data. A wide variety of claims about similarity and dissimilarity of these various representations have been made using CKA results. In this work we present analysis that formally characterizes CKA sensitivity to a large class of simple transformations, which can naturally occur in the context of modern machine learning. This provides a concrete explanation to CKA sensitivity to outliers, which has been observed in past works, and to transformations that preserve the linear separability of the data, an important generalization attribute. We empirically investigate several weaknesses of the CKA similarity metric, demonstrating situations in which it gives unexpected or counterintuitive results. Finally we study approaches for modifying representations to maintain functional behaviour while changing the CKA value. Our results illustrate that, in many cases, the CKA value can be easily manipulated without substantial changes to the functional behaviour of the models, and call for caution when leveraging activation alignment metrics.

Existing models for learning representations in supervised classification problems are permutation invariant with respect to class labels. However, structured knowledge about the classes, such as hierarchical label structures, widely exists in many real-world datasets, e.g., the ImageNet and CIFAR benchmarks. How to learn representations that can preserve such structures among the classes remains an open problem. To approach this problem, given a tree of class hierarchy, we first define a tree metric between any pair of nodes in the tree to be the length of the shortest path connecting them. We then provide a method to learn the hierarchical relationship of class labels by approximately embedding the tree metric in the Euclidean space of features. More concretely, during supervised training, we propose to use the Cophenetic Correlation Coefficient (CPCC) as a regularizer for the cross-entropy loss to correlate the tree metric of classes and the Euclidean distance in the class-conditioned representations. Our proposed regularizer is computationally lightweight and easy to implement. Empirically, we demonstrate that this approach can help to learn more interpretable representations due to the preservation of the tree metric, and leads to better in-distribution generalization as well as under sub-population shifts over six real-world datasets.

Designing robust loss functions is popular in learning with noisy labels while existing designs did not explicitly consider the overfitting property of deep neural networks (DNNs). As a result, applying these losses may still suffer from overfitting/memorizing noisy labels as training proceeds. In this paper, we first theoretically analyze the memorization effect and show that a lower-capacity model may perform better on noisy datasets. However, it is non-trivial to design a neural network with the best capacity given an arbitrary task. To circumvent this dilemma, instead of changing the model architecture, we decouple DNNs into an encoder followed by a linear classifier and propose to restrict the function space of a DNN by a representation regularizer. Particularly, we require the distance between two self-supervised features to be positively related to the distance between the corresponding two supervised model outputs. Our proposed framework is easily extendable and can incorporate many other robust loss functions to further improve performance. Extensive experiments and theoretical analyses support our claims. Code is available at https://github.com/UCSC-REAL/SelfSup_NoisyLabel.

Molecular representation learning (MRL) has gained tremendous attention due to its critical role in learning from limited supervised data for applications like drug design. In most MRL methods, molecules are treated as 1D sequential tokens or 2D topology graphs, limiting their ability to incorporate 3D information for downstream tasks and, in particular, making it almost impossible for 3D geometry prediction/generation. In this paper, we propose a universal 3D MRL framework, called Uni-Mol, that significantly enlarges the representation ability and application scope of MRL schemes. Uni-Mol contains two pretrained models with the same SE(3) Transformer architecture: a molecular model pretrained by 209M molecular conformations; a pocket model pretrained by 3M candidate protein pocket data. Besides, Uni-Mol contains several finetuning strategies to apply the pretrained models to various downstream tasks. By properly incorporating 3D information, Uni-Mol outperforms SOTA in 14/15 molecular property prediction tasks. Moreover, Uni-Mol achieves superior performance in 3D spatial tasks, including protein-ligand binding pose prediction, molecular conformation generation, etc. The code, model, and data are made publicly available at https://github.com/dptech-corp/Uni-Mol.

In representation learning, a common approach is to seek representations which disentangle the underlying factors of variation. Eastwood & Williams (2018) proposed three metrics for quantifying the quality of such disentangled representations: disentanglement (D), completeness (C) and informativeness (I). In this work, we first connect this DCI framework to two common notions of linear and nonlinear identifiability, thereby establishing a formal link between disentanglement and the closely-related field of independent component analysis. We then propose an extended DCI-ES framework with two new measures of representation quality—explicitness (E) and size (S)—and point out how D and C can be computed for black-box predictors. Our main idea is that the functional capacity required to use a representation is an important but thus-far neglected aspect of representation quality, which we quantify using explicitness or ease-of-use (E). We illustrate the relevance of our extensions on the MPI3D and Cars3D datasets.

Edges in many real-world social/information networks are associated with rich text information (e.g., user-user communications or user-product reviews). However, mainstream network representation learning models focus on propagating and aggregating node attributes, lacking specific designs to utilize text semantics on edges. While there exist edge-aware graph neural networks, they directly initialize edge attributes as a feature vector, which cannot fully capture the contextualized text semantics of edges. In this paper, we propose Edgeformers, a framework built upon graph-enhanced Transformers, to perform edge and node representation learning by modeling texts on edges in a contextualized way. Specifically, in edge representation learning, we inject network information into each Transformer layer when encoding edge texts; in node representation learning, we aggregate edge representations through an attention mechanism within each node’s ego-graph. On five public datasets from three different domains, Edgeformers consistently outperform state-of-the-art baselines in edge classification and link prediction, demonstrating the efficacy in learning edge and node representations, respectively.

We propose topology-aware feature partitioning into $k$ disjoint partitions for given scene features as a method for object-centric representation learning. To this end, we propose to use minimum $s$-$t$ graph cuts as a partitioning method which is represented as a linear program. The method is topologically aware since it explicitly encodes neighborhood relationships in the image graph. To solve the graph cuts our solution relies on an efficient, scalable, and differentiable quadratic programming approximation. Optimizations specific to cut problems allow us to solve the quadratic programs and compute their gradients significantly more efficiently compared with the general quadratic programming approach. Our results show that our approach is scalable and outperforms existing methods on object discovery tasks with textured scenes and objects.

We propose a generic module named Indirection Layer (InLay), which leverages indirection and data internal relationships to effectively construct symbolic indirect representations to improve out-of-distribution generalization capabilities of various neural architectures. InLay receives data input in the form of a sequence of objects, treats it as a complete weighted graph whose vertices are the objects and edge weights are scalars representing relationships between vertices. The input is first mapped via indirection to a symbolic graph with data-independent and trainable vertices. This symbolic graph is then propagated, resulting in new vertex features whose indirection will be used for prediction steps afterward. Theoretically, we show that the distances between indirection representations are bounded by the distances between corresponding graphs, implying that unseen samples with very different surface statistics can still be close in the representation space to the seen samples if they share similar internal relationships. We demonstrate that InLay is consistently effective in improving out-of-distribution generalization throughout a comprehensive suite of experiments, including IQ problems, distorted image classification, and few-shot domain adaptation NLP classification. We also conduct ablation studies to verify different design choices of InLay.

Intelligent systems deployed in the real world suffer from catastrophic forgetting when exposed to a sequence of tasks. Humans, on the other hand, acquire, consolidate, and transfer knowledge between tasks that rarely interfere with the consolidated knowledge. Accompanied by self-regulated neurogenesis, continual learning in the brain is governed by the rich set of neurophysiological processes that harbor different types of knowledge which are then integrated by the conscious processing. Thus, inspired by Global Workspace Theory of conscious information access in the brain, we propose TAMiL, a continual learning method that entails task-attention modules to capture task-specific information from the common representation space. We employ simple, undercomplete autoencoders to create a communication bottleneck between the common representation space and the global workspace, allowing only the task-relevant information to the global workspace, thereby greatly reducing task interference. Experimental results show that our method outperforms state-of-the-art rehearsal-based and dynamic sparse approaches and bridges the gap between fixed capacity and parameter isolation approaches while being scalable. We also show that our method effectively mitigates catastrophic forgetting while being well-calibrated with reduced task-recency bias.

In clinical environments, image-based diagnosis is desired to achieve robustness on multi-center samples. Toward this goal, a natural way is to capture only clinically disease-related features. However, such disease-related features are often entangled with center-effect, disabling robust transferring to unseen centers/domains. To disentangle disease-related features, we first leverage structural causal modeling to explicitly model disease-related and center-effects that are provable to be disentangled from each other. Guided by this, we propose a novel Domain Agnostic Representation Model (DarMo) based on variational Auto-Encoder. To facilitate disentanglement, we design domain-agnostic and domain-aware encoders to respectively capture disease-related features and varied center-effects by incorporating a domain-aware batch normalization layer. Besides, we constrain the disease-related features to well predict the disease label as well as clinical attributes, by leveraging Graph Convolutional Network (GCN) into our decoder. The effectiveness and utility of our method are demonstrated by the superior performance over others on both public datasets and inhouse datasets.