ICLR 2023 Papers

This work establishes low test error of gradient flow (GF) and stochastic gradient descent (SGD) on two-layer ReLU networks with standard initialization scale, in three regimes where key sets of weights rotate little (either naturally due to GF and SGD, or due to an artificial constraint), and making use of margins as the core analysis technique. The first regime is near initialization, specifically until the weights have moved by $\mathcal{O}(\sqrt m)$, where $m$ denotes the network width, which is in sharp contrast to the $\mathcal{O}(1)$ weight motion allowed by the Neural Tangent Kernel (NTK); here it is shown that GF and SGD only need a network width and number of samples inversely proportional to the NTK margin, and moreover that GF attains at least the NTK margin itself and in particular escapes bad KKT points of the margin objective, whereas prior work could only establish nondecreasing but arbitrarily small margins. The second regime is the Neural Collapse (NC) setting, where data lies in well-separated groups, and the sample complexity scales with the number of groups; here the contribution over prior work is an analysis of the entire GF trajectory from initialization. Lastly, if the inner layer weights are constrained to change in norm only and can not rotate, then GF with large widths achieves globally maximal margins, and its sample complexity scales with their inverse; this is in contrast to prior work, which required infinite width and a tricky dual convergence assumption.

Feature distillation makes the student mimic the intermediate features of the teacher. Nearly all existing feature-distillation methods use L2 distance or its slight variants as the distance metric between teacher and student features. However, while L2 distance is isotropic w.r.t. all dimensions, the neural network’s operation on different dimensions is usually anisotropic, i.e., perturbations with the same 2-norm but in different dimensions of intermediate features lead to changes in the final output with largely different magnitude. Considering this, we argue that the similarity between teacher and student features should \textit{not} be measured merely based on their appearance (i.e., L2 distance), but should, more importantly, be measured by their difference in function, namely how later layers of the network will read, decode, and process them. Therefore, we propose Function-Consistent Feature Distillation (FCFD), which explicitly optimizes the functional similarity between teacher and student features. The core idea of FCFD is to make teacher and student features not only numerically similar, but more importantly produce similar outputs when fed to the later part of the same network. With FCFD, the student mimics the teacher more faithfully and learns more from the teacher. Extensive experiments on image classification and object detection demonstrate the superiority of FCFD to existing methods. Furthermore, we can combine FCFD with many existing methods to obtain even higher accuracy. Our codes are available at https://github.com/LiuDongyang6/FCFD.

Unlearnable examples (ULEs) aim to protect data from unauthorized usage for training DNNs. Existing work adds $\ell_\infty$-bounded perturbations to the original sample so that the trained model generalizes poorly. Such perturbations, however, are easy to eliminate by adversarial training and data augmentations. In this paper, we resolve this problem from a novel perspective by perturbing only one pixel in each image. Interestingly, such a small modification could effectively degrade model accuracy to almost an untrained counterpart. Moreover, our produced \emph{One-Pixel Shortcut (OPS)} could not be erased by adversarial training and strong augmentations. To generate OPS, we perturb in-class images at the same position to the same target value that could mostly and stably deviate from all the original images. Since such generation is only based on images, OPS needs significantly less computation cost than the previous methods using DNN generators. Based on OPS, we introduce an unlearnable dataset called CIFAR-10-S, which is indistinguishable from CIFAR-10 by humans but induces the trained model to extremely low accuracy. Even under adversarial training, a ResNet-18 trained on CIFAR-10-S has only 10.61% accuracy, compared to 83.02% by the existing error-minimizing method.

Recent advances in the understanding of Generative Adversarial Networks (GANs) have led to remarkable progress in visual editing and synthesis tasks, capitalizing on the rich semantics that are embedded in the latent spaces of pre-trained GANs. However, existing methods are often tailored to specific GAN architectures and are limited to either discovering global semantic directions that do not facilitate localized control, or require some form of supervision through manually provided regions or segmentation masks. In this light, we present an architecture-agnostic approach that jointly discovers factors representing spatial parts and their appearances in an entirely unsupervised fashion. These factors are obtained by applying a semi-nonnegative tensor factorization on the feature maps, which in turn enables context-aware local image editing with pixel-level control. In addition, we show that the discovered appearance factors correspond to saliency maps that localize concepts of interest, without using any labels. Experiments on a wide range of GAN architectures and datasets show that, in comparison to the state of the art, our method is far more efficient in terms of training time and, most importantly, provides much more accurate localized control. Our code is available at: https://github.com/james-oldfield/PandA.

Despite the widespread use of unsupervised models, very few methods are designed to explain them. Most explanation methods explain a scalar model output. However, unsupervised models output representation vectors, the elements of which are not good candidates to explain because they lack semantic meaning. To bridge this gap, recent works defined a scalar explanation output: a dot product-based similarity in the representation space to the sample being explained (i.e., an explicand). Although this enabled explanations of unsupervised models, the interpretation of this approach can still be opaque because similarity to the explicand's representation may not be meaningful to humans. To address this, we propose contrastive corpus similarity, a novel and semantically meaningful scalar explanation output based on a reference corpus and a contrasting foil set of samples. We demonstrate that contrastive corpus similarity is compatible with many post-hoc feature attribution methods to generate COntrastive COrpus Attributions (COCOA) and quantitatively verify that features important to the corpus are identified. We showcase the utility of COCOA in two ways: (i) we draw insights by explaining augmentations of the same image in a contrastive learning setting (SimCLR); and (ii) we perform zero-shot object localization by explaining the similarity of image representations to jointly learned text representations (CLIP).

Identifying nonlinear causal relationships in high-dimensional biological data is an important task. However, current neural network based causality detection approaches for such data suffer from poor interpretability and cannot scale well to the high dimensional regime. Here we present GEASS (Granger fEAture Selection of Spatiotemporal data), which identifies sparse Granger causality mechanisms of high dimensional spatiotemporal data by a single neural network. GEASS maximizes sparsity-regularized modified transfer entropy with a theoretical guarantee of recovering features with spatial/temporal Granger causal relationships. The sparsity regularization is achieved by a novel combinatorial stochastic gate layer to select sparse non-overlapping feature subsets. We demonstrate the efficacy of GEASS in several synthetic datasets and real biological data from single-cell RNA sequencing and spatial transcriptomics.

Masking some input variables of a deep neural network (DNN) and computing output changes on the masked input sample represent a typical way to compute attributions of input variables in the sample. People usually mask an input variable using its baseline value. However, there is no theory to examine whether baseline value faithfully represents the absence of an input variable, i.e., removing all signals from the input variable. Fortunately, recent studies (Ren et al., 2023a; Deng et al., 2022a) show that the inference score of a DNN can be strictly disentangled into a set of causal patterns (or concepts) encoded by the DNN. Therefore, we propose to use causal patterns to examine the faithfulness of baseline values. More crucially, it is proven that causal patterns can be explained as the elementary rationale of the Shapley value. Furthermore, we propose a method to learn optimal baseline values, and experimental results have demonstrated its effectiveness.

This paper investigates a missing feature imputation problem for graph learning tasks. Several methods have previously addressed learning tasks on graphs with missing features. However, in cases of high rates of missing features, they were unable to avoid significant performance degradation. To overcome this limitation, we introduce a novel concept of channel-wise confidence in a node feature, which is assigned to each imputed channel feature of a node for reflecting the certainty of the imputation. We then design pseudo-confidence using the channel-wise shortest path distance between a missing-feature node and its nearest known-feature node to replace unavailable true confidence in an actual learning process. Based on the pseudo-confidence, we propose a novel feature imputation scheme that performs channel-wise inter-node diffusion and node-wise inter-channel propagation. The scheme can endure even at an exceedingly high missing rate (e.g., 99.5\%) and it achieves state-of-the-art accuracy for both semi-supervised node classification and link prediction on various datasets containing a high rate of missing features. Codes are available at https://github.com/daehoum1/pcfi.

The ability to interpret machine learning models is critical for high-stakes applications. Due to its desirable theoretical properties, path integration is a widely used scheme for feature attribution to interpret model predictions. However, the methods implementing this scheme currently rely on absolute attribution scores to eventually provide sensible interpretations. This not only contradicts the premise that the features with larger attribution scores are more relevant to the model prediction, but also conflicts with the theoretical settings for which the desirable properties of the attributions are proven. We address this by devising a method to first compute an appropriate reference for the path integration scheme. This reference further helps in identifying valid interpolation points on a desired integration path. The reference is computed in a gradient ascending direction on the model's loss surface, while the interpolations are performed by analyzing the model gradients and variations between the reference and the input. The eventual integration is effectively performed along a non-linear path. Our scheme can be incorporated into the existing integral-based attribution methods. We also devise an effective sampling and integration procedure that enables employing our scheme with multi-reference path integration efficiently. We achieve a marked performance boost for a range of integral-based attribution methods on both local and global evaluation metrics by enhancing them with our scheme. Our extensive results also show improved sensitivity, sanity preservation and model robustness with the proposed re-calibration of the attribution techniques with our method.

Federated learning is a distributed paradigm that allows multiple parties to collaboratively train deep models without exchanging the raw data. However, the data distribution among clients is naturally non-i.i.d., which leads to severe degradation of the learnt model. The primary goal of this paper is to develop a robust federated learning algorithm to address feature shift in clients’ samples, which can be caused by various factors, e.g., acquisition differences in medical imaging. To reach this goal, we propose FedFA to tackle federated learning from a dis- tinct perspective of federated feature augmentation. FedFA is based on a major insight that each client’s data distribution can be characterized by statistics (i.e., mean and standard deviation) of latent features; and it is likely to manipulate these local statistics globally, i.e., based on information in the entire federation, to let clients have a better sense of the underlying distribution and therefore alleviate local data bias. Based on this insight, we propose to augment each local feature statistic probabilistically based on a normal distribution, whose mean is the original statistic and variance quantifies the augmentation scope. Key to our approach is the determination of a meaningful Gaussian variance, which is accomplished by taking into account not only biased data of each individual client, but also underlying feature statistics characterized by all participating clients. We offer both theoretical and empirical justifications to verify the effectiveness of FedFA. Our code is available at https://github.com/tfzhou/FedFA.

Feature selection is the problem of selecting a subset of features for a machine learning model that maximizes model quality subject to a budget constraint. For neural networks, prior methods, including those based on $\ell_1$ regularization, attention, and other techniques, typically select the entire feature subset in one evaluation round, ignoring the residual value of features during selection, i.e., the marginal contribution of a feature given that other features have already been selected. We propose a feature selection algorithm called Sequential Attention that achieves state-of-the-art empirical results for neural networks. This algorithm is based on an efficient one-pass implementation of greedy forward selection and uses attention weights at each step as a proxy for feature importance. We give theoretical insights into our algorithm for linear regression by showing that an adaptation to this setting is equivalent to the classical Orthogonal Matching Pursuit (OMP) algorithm, and thus inherits all of its provable guarantees. Our theoretical and empirical analyses offer new explanations towards the effectiveness of attention and its connections to overparameterization, which may be of independent interest.

Data heterogeneity is one of the most challenging issues in federated learning, which motivates a variety of approaches to learn personalized models for participating clients. One such approach in deep neural networks based tasks is employing a shared feature representation and learning a customized classifier head for each client. However, previous works do not utilize the global knowledge during local representation learning and also neglect the fine-grained collaboration between local classifier heads, which limits the model generalization ability. In this work, we conduct explicit local-global feature alignment by leveraging global semantic knowledge for learning a better representation. Moreover, we quantify the benefit of classifier combination for each client as a function of the combining weights and derive an optimization problem for estimating optimal weights. Finally, extensive evaluation results on benchmark datasets with various heterogeneous data scenarios demonstrate the effectiveness of our proposed method.

We propose a simple data model inspired from natural data such as text or images, and use it to study the importance of learning features in order to achieve good generalization. Our data model follows a long-tailed distribution in the sense that some rare and uncommon subcategories have few representatives in the training set. In this context we provide evidence that a learner succeeds if and only if it identifies the correct features, and moreover derive non-asymptotic generalization error bounds that precisely quantify the penalty that one must pay for not learning features.

Interpretability is a pressing issue for machine learning. Common approaches to interpretable machine learning constrain interactions between features of the input, sacrificing model complexity in order to render more comprehensible the effects of those features on the model's output. We approach interpretability from a new angle: constrain the information about the features without restricting the complexity of the model. We use the Distributed Information Bottleneck to optimally compress each feature so as to maximally preserve information about the output. The learned information allocation, by feature and by feature value, provides rich opportunities for interpretation, particularly in problems with many features and complex feature interactions. The central object of analysis is not a single trained model, but rather a spectrum of models serving as approximations that leverage variable amounts of information about the inputs. Information is allocated to features by their relevance to the output, thereby solving the problem of feature selection by constructing a learned continuum of feature inclusion-to-exclusion. The optimal compression of each feature---at every stage of approximation---allows fine-grained inspection of the distinctions among feature values that are most impactful for prediction. We develop a framework for extracting insight from the spectrum of approximate models and demonstrate its utility on a range of tabular datasets.

Click-through rate (CTR) prediction plays important role in the advertisement, recommendation, and retrieval applications. Given the feature set, how to fully utilize the information from the feature set is an active topic in deep CTR model designs. There are several existing deep CTR works focusing on feature interactions, feature attentions, and so on. They attempt to capture high-order feature interactions to enhance the generalization ability of deep CTR models. However, these works either suffer from poor high-order feature interaction modeling using DNN or ignore the balance between generalization and memorization during the recommendation. To mitigate these problems, we propose an adaptive feature fusion framework called MaskFusion, to additionally capture the explicit interactions between the input feature and the existing deep part structure of deep CTR models dynamically, besides the common feature interactions proposed in existing works. MaskFusion is an instance-aware feature augmentation method, which makes deep CTR models more personalized by assigning each feature with an instance-adaptive mask and fusing each feature with each hidden state vector in the deep part structure. MaskFusion can also be integrated into any existing deep CTR models flexibly. MaskFusion achieves state-of-the-art (SOTA) performance on all seven benchmarks deep CTR models with three public datasets.

Recent advancements in explainable machine learning provide effective and faithful solutions for interpreting model behaviors. However, many explanation methods encounter efficiency issues, which largely limit their deployments in practical scenarios. Real-time explainer (RTX) frameworks have thus been proposed to accelerate the model explanation process by learning an one-feed-forward explainer. Existing RTX frameworks typically build the explainer under the supervised learning paradigm, which requires large amounts of explanation labels as the ground truth. Considering that accurate explanation labels are usually hard to obtain, due to constrained computational resources and limited human efforts, effective explainer training is still challenging in practice. In this work, we propose a COntrastive Real-Time eXplanation (CoRTX) framework to learn the explanation-oriented representation and relieve the intensive dependence of explainer training on explanation labels. Specifically, we design a synthetic strategy to select positive and negative instances for explanation representation learning. Theoretical analysis show that our selection strategy can benefit the contrastive learning process on explanation tasks. Experimental results on three real-world datasets further demonstrate the efficiency and efficacy of our proposed CoRTX framework.

Mixup, which creates synthetic training instances by linearly interpolating random sample pairs, is a simple and yet effective regularization technique to boost the performance of deep models trained with SGD. In this work, we report a previously unobserved phenomenon in Mixup raining: on a number of standard datasets, the performance of Mixup-trained models starts to decay after training for a large number of epochs, giving rise to a U-shaped generalization curve. This behavior is further aggravated when the size of original dataset is reduced. To help understand such a behavior of Mixup, we show theoretically that Mixup training may introduce undesired data-dependent label noises to the synthesized data. Via analyzing a least-square regression problem with a random feature model, we explain why noisy labels may cause the U-shaped curve to occur: Mixup improves generalization through fitting the clean patterns at the early training stage, but as training progresses, Mixup becomes over-fitting to the noise in the synthetic data. Extensive experiments are performed on a variety of benchmark datasets, validating this explanation.

For small training set sizes $P$, the generalization error of wide neural networks is well-approximated by the error of an infinite width neural network (NN), either in the kernel or mean-field/feature-learning regime. However, after a critical sample size $P^*$, we empirically find the finite-width network generalization becomes worse than that of the infinite width network. In this work, we empirically study the transition from infinite-width behavior to this \textit{variance-limited} regime as a function of sample size $P$ and network width $N$. We find that finite-size effects can become relevant for very small dataset sizes on the order of $P^* \sim \sqrt{N}$ for polynomial regression with ReLU networks. We discuss the source of these effects using an argument based on the variance of the NN's final neural tangent kernel (NTK). This transition can be pushed to larger $P$ by enhancing feature learning or by ensemble averaging the networks. We find that the learning curve for regression with the final NTK is an accurate approximation of the NN learning curve. Using this, we provide a toy model which also exhibits $P^* \sim \sqrt{N}$ scaling and has $P$-dependent benefits from feature learning.

SHAP explanations aim at identifying which features contribute the most to the difference in model prediction at a specific input versus a background distribution. Recent studies have shown that they can be manipulated by malicious adversaries to produce arbitrary desired explanations. However, existing attacks focus solely on altering the black-box model itself. In this paper, we propose a complementary family of attacks that leave the model intact and manipulate SHAP explanations using stealthily biased sampling of the data points used to approximate expectations w.r.t the background distribution. In the context of fairness audit, we show that our attack can reduce the importance of a sensitive feature when explaining the difference in outcomes between groups while remaining undetected. More precisely, experiments performed on real-world datasets showed that our attack could yield up to a 90\% relative decrease in amplitude of the sensitive feature attribution. These results highlight the manipulability of SHAP explanations and encourage auditors to treat them with skepticism.