ICML 2024 论文列表

We revisit the result of Morris et al. (AAAI'19) that message-passing graphs neural networks (MPNNs) are equal in their distinguishing power to the Weisfeiler--Leman (WL) isomorphism test. Morris et al. show their result with ReLU activation function and $O(n)$-dimensional feature vectors, where $n$ is the size of the graph. Recently, by introducing randomness into the architecture, Aamand et al. (NeurIPS'22) improved this bound to $O(\log n)$-dimensional feature vectors, although at the expense of guaranteeing perfect simulation only with high probability. In all these constructions, to guarantee equivalence to the WL test, the dimension of feature vectors in the MPNN has to increase with the size of the graphs. However, architectures used in practice have feature vectors of constant dimension. Thus, there is a gap between the guarantees provided by these results and the actual characteristics of architectures used in practice. In this paper we close this gap by showing that, for *any* non-polynomial analytic (like the sigmoid) activation function, to guarantee that MPNNs are equivalent to the WL test, feature vectors of dimension $d=1$ is all we need, independently of the size of the graphs. Our main technical insight is that for simulating multi-sets in the WL-test, it is enough to use linear independence of feature vectors over rationals instead of reals. Countability of the set of rationals together with nice properties of analytic functions allow us to carry out the simulation invariant over the iterations of the WL test without increasing the dimension of the feature vectors.

Feature learning is thought to be one of the fundamental reasons for the success of deep neural networks. It is rigorously known that in two-layer fully-connected neural networks under certain conditions, one step of gradient descent on the first layer can lead to feature learning; characterized by the appearance of a separated rank-one component---spike---in the spectrum of the feature matrix. However, with a constant gradient descent step size, this spike only carries information from the linear component of the target function and therefore learning non-linear components is impossible. We show that with a learning rate that grows with the sample size, such training in fact introduces multiple rank-one components, each corresponding to a specific polynomial feature. We further prove that the limiting large-dimensional and large sample training and test errors of the updated neural networks are fully characterized by these spikes. By precisely analyzing the improvement in the training and test errors, we demonstrate that these non-linear features can enhance learning.

Insertion/deletion metrics and their variants have been extensively applied to evaluate attribution-based explanation methods. Such metrics measure the significance of features by observing changes in model predictions as features are incrementally inserted or deleted. Given the direct connection between the attribution values and model predictions that insertion/deletion metrics enable, they are commonly used as the decisive metrics for novel attribution methods. Such influential metrics for explanation methods should be handled with great scrutiny. However, contemporary research on insertion/deletion metrics falls short of a comprehensive analysis. To address this, we propose the TRAjectory importanCE (TRACE) framework, which achieves the best scores of the insertion/deletion metric. Our contribution includes two aspects: 1) TRACE stands as the principled explanation for explaining the influence of feature deletion on model predictions. We demonstrate that TRACE is guaranteed to achieve almost optimal results both theoretically and empirically. 2) Using TRACE, we benchmark insertion/deletion metrics across all possible settings and study critical problems such as the out-of-distribution (OOD) issue, and provide practical guidance on applying these metrics in practice.

For a large class of feature maps we provide a tight asymptotic characterisation of the test error associated with learning the readout layer, in the high-dimensional limit where the input dimension, hidden layer widths, and number of training samples are proportionally large. This characterization is formulated in terms of the population covariance of the features. Our work is partially motivated by the problem of learning with Gaussian rainbow neural networks, namely deep non-linear fully-connected networks with random but structured weights, whose row-wise covariances are further allowed to depend on the weights of previous layers. For such networks we also derive a closed-form formula for the feature covariance in terms of the weight matrices. We further find that in some cases our results can capture feature maps learned by deep, finite-width neural networks trained under gradient descent.

1-bit detectors show impressive performance comparable to their real-valued counterparts when detecting commonly sized objects while exhibiting significant performance degradation on tiny objects. The challenge stems from the fact that high-level features extracted by 1-bit convolutions seem less compelling to reveal the discriminative foreground features. To address these issues, we introduce a Discriminative Feature Refinement method for 1-bit Detectors (DFR-Det), aiming to enhance the discriminative ability of foreground representation for tiny objects in aerial images. This is accomplished by refining the feature representation using an information bottleneck (IB) to achieve a distinctive representation of tiny objects. Specifically, we introduce a new decoder with a foreground mask, aiming to enhance the discriminative ability of high-level features for the target but suppress the background impact. Additionally, our decoder is simple but effective and can be easily mounted on existing detectors without extra burden added to the inference procedure. Extensive experiments on various tiny object detection (TOD) tasks demonstrate DFR-Det's superiority over state-of-the-art 1-bit detectors. For example, 1-bit FCOS achieved by DFR-Det achieves the 12.8% AP on AI-TOD dataset, approaching the performance of the real-valued counterpart.

Feature attribution, the ability to localize regions of the input data that are relevant for classification, is an important capability for ML models in scientific and biomedical domains. Current methods for feature attribution, which rely on "explaining" the predictions of end-to-end classifiers, suffer from imprecise feature localization and are inadequate for use with small sample sizes and high-dimensional datasets due to computational challenges. We introduce prospector heads, an efficient and interpretable alternative to explanation-based attribution methods that can be applied to any encoder and any data modality. Prospector heads generalize across modalities through experiments on sequences (text), images (pathology), and graphs (protein structures), outperforming baseline attribution methods by up to 26.3 points in mean localization AUPRC. We also demonstrate how prospector heads enable improved interpretation and discovery of class-specific patterns in input data. Through their high performance, flexibility, and generalizability, prospectors provide a framework for improving trust and transparency for ML models in complex domains.

Humans use context to specify preferences over behaviors, i.e. their reward functions. Yet, algorithms for inferring reward models from preference data do not take this social learning view into account. Inspired by pragmatic human communication, we study how to extract fine-grained data regarding why an example is preferred that is useful for learning an accurate reward model. We propose to enrich preference queries to ask both (1) which features of a given example are preferable in addition to (2) comparisons between objects. We derive an approach for learning from these feature-level preferences, both for cases where users specify which features are reward-relevant, and when users do not. We evaluate our approach on linear bandit settings in both visual and language-based domains. Results support the efficiency of our approach in quickly converging to accurate rewards with less comparisons vs. example-only labels. Finally, we validate the real-world applicability with a behavioral experiment on a mushroom foraging task. Our findings suggest that incorporating pragmatic feature preferences is a promising approach for more efficient user-aligned reward learning.

Subset or core-set selection offers a data-efficient way for training deep learning models. One-shot subset selection poses additional challenges as subset selection is only performed once and full set data become unavailable after the selection. However, most existing methods tend to choose either diverse or difficult data samples, which fail to faithfully represent the joint data distribution that is comprised of both feature and label information. The selection is also performed independently from the subset size, which plays an essential role in choosing what types of samples. To address this critical gap, we propose to conduct Feature similarity and Label variability Balanced One-shot Subset Selection (BOSS), aiming to construct an optimal size-aware subset for data-efficient deep learning. We show that a novel balanced core-set loss bound theoretically justifies the need to simultaneously consider both diversity and difficulty to form an optimal subset. It also reveals how the subset size influences the bound. We further connect the inaccessible bound to a practical surrogate target which is tailored to subset sizes and varying levels of overall difficulty. We design a novel Beta-scoring importance function to delicately control the optimal balance of diversity and difficulty. Comprehensive experiments conducted on both synthetic and real data justify the important theoretical properties and demonstrate the superior performance of BOSS as compared with the competitive baselines.

Local feature selection in machine learning provides instance-specific explanations by focusing on the most relevant features for each prediction, enhancing the interpretability of complex models. However, such methods tend to produce misleading explanations by encoding additional information in their selections. In this work, we attribute the problem of misleading selections by formalizing the concepts of label and feature leakage. We rigorously derive the necessary and sufficient conditions under which we can guarantee no leakage, and show existing methods do not meet these conditions. Furthermore, we propose the first local feature selection method that is proven to have no leakage called SUWR. Our experimental results indicate that SUWR is less prone to overfitting and combines state-of-the-art predictive performance with high feature-selection sparsity. Our generic and easily extendable formal approach provides a strong theoretical basis for future work on interpretability with reliable explanations.

Learning representations that generalize under distribution shifts is critical for building robust machine learning models. However, despite significant efforts in recent years, algorithmic advances in this direction have been limited. In this work, we seek to understand the fundamental difficulty of out-of-distribution generalization with deep neural networks. We first empirically show that perhaps surprisingly, even allowing a neural network to explicitly fit the representations obtained from a teacher network that can generalize out-of-distribution is insufficient for the generalization of the student network. Then, by a theoretical study of two-layer ReLU networks optimized by stochastic gradient descent (SGD) under a structured feature model, we identify a fundamental yet unexplored feature learning proclivity of neural networks, feature contamination: neural networks can learn uncorrelated features together with predictive features, resulting in generalization failure under distribution shifts. Notably, this mechanism essentially differs from the prevailing narrative in the literature that attributes the generalization failure to spurious correlations. Overall, our results offer new insights into the non-linear feature learning dynamics of neural networks and highlight the necessity of considering inductive biases in out-of-distribution generalization.

Knowledge distillation is a widely adopted model compression technique that has been successfully applied to object detection. In feature distillation, it is common practice for the student model to imitate the feature responses of the teacher model, with the underlying objective of improving its own abilities by reducing the disparity with the teacher. However, it is crucial to recognize that the disparities between the student and teacher are inconsistent, highlighting their varying abilities. In this paper, we explore the inconsistency in the disparity between teacher and student feature maps and analyze their impact on the efficiency of the distillation. We find that regions with varying degrees of difference should be treated separately, with different distillation constraints applied accordingly. We introduce our distillation method called Disparity Feature Distillation(DFD). The core idea behind DFD is to apply different treatments to regions with varying learning difficulties, simultaneously incorporating leniency and strictness. It enables the student to better assimilate the teacher’s knowledge. Through extensive experiments, we demonstrate the effectiveness of our proposed DFD in achieving significant improvements. For instance, when applied to detectors based on ResNet50 such as RetinaNet, FasterRCNN, and RepPoints, our method enhances their mAP from 37.4%, 38.4%, 38.6% to 41.7%, 42.4%, 42.7%, respectively. Our approach also demonstrates substantial improvements on YOLO and ViT-based models. The code is available at https://github.com/luckin99/DFD.

Hyperspectral image super-resolution (HISR) aims to fuse a low-resolution hyperspectral image (LR-HSI) with a high-resolution multispectral image (HR-MSI) to obtain a high-resolution hyperspectral image (HR-HSI). Due to some existing HISR methods ignoring the significant feature difference between LR-HSI and HR-MSI, the reconstructed HR-HSI typically exhibits spectral distortion and blurring of spatial texture. To solve this issue, we propose a multi-scale feature transfer network (MFTN) for HISR. Firstly, three multi-scale feature extractors are constructed to extract features of different scales from the input images. Then, a multi-scale feature transfer module (MFTM) consisting of three improved feature matching Transformers (IMatchFormers) is designed to learn the detail features of different scales from HR-MSI by establishing the cross-model feature correlation between LR-HSI and degraded HR-MSI. Finally, a multiscale dynamic aggregation module (MDAM) containing three spectral aware aggregation modules (SAAMs) is constructed to reconstruct the final HR-HSI by gradually aggregating features of different scales. Extensive experimental results on three commonly used datasets demonstrate that the proposed model achieves better performance compared to state- of-the-art (SOTA) methods.

We investigate the problem of explainability for machine learning models, focusing on Feature Attribution Methods (FAMs) that evaluate feature importance through perturbation tests. Despite their utility, FAMs struggle to distinguish the contributions of different features, when their prediction changes are similar after perturbation. To enhance FAMs' discriminative power, we introduce Feature Attribution with Necessity and Sufficiency (FANS), which find a neighborhood of the input such that perturbing samples within this neighborhood have a high Probability of being Necessity and Sufficiency (PNS) cause for the change in predictions, and use this PNS as the importance of the feature. Specifically, FANS compute this PNS via a heuristic strategy for estimating the neighborhood and a perturbation test involving two stages (factual and interventional) for counterfactual reasoning. To generate counterfactual samples, we use a resampling-based approach on the observed samples to approximate the required conditional distribution. We demonstrate that FANS outperforms existing attribution methods on six benchmarks. Please refer to the source code via https://github.com/DMIRLAB-Group/FANS.

Recent breakthroughs in generative modeling have sparked interest in practical single-model attribution. Such methods predict whether a sample was generated by a specific generator or not, for instance, to prove intellectual property theft. However, previous works are either limited to the closed-world setting or require undesirable changes to the generative model. We address these shortcomings by, first, viewing single-model attribution through the lens of anomaly detection. Arising from this change of perspective, we propose FLIPAD, a new approach for single-model attribution in the open-world setting based on final-layer inversion and anomaly detection. We show that the utilized final-layer inversion can be reduced to a convex lasso optimization problem, making our approach theoretically sound and computationally efficient. The theoretical findings are accompanied by an experimental study demonstrating the effectiveness of our approach and its flexibility to various domains.

In contrast with Mercer kernel-based approaches as used e.g. in Kernel Principal Component Analysis (KPCA), it was previously shown that Singular Value Decomposition (SVD) inherently relates to asymmetric kernels and Asymmetric Kernel Singular Value Decomposition (KSVD) has been proposed. However, the existing formulation to KSVD cannot work with infinite-dimensional feature mappings, the variational objective can be unbounded, and needs further numerical evaluation and exploration towards machine learning. In this work, i) we introduce a new asymmetric learning paradigm based on coupled covariance eigenproblem (CCE) through covariance operators, allowing infinite-dimensional feature maps. The solution to CCE is ultimately obtained from the SVD of the induced asymmetric kernel matrix, providing links to KSVD. ii) Starting from the integral equations corresponding to a pair of coupled adjoint eigenfunctions, we formalize the asymmetric Nyström method through a finite sample approximation to speed up training. iii) We provide the first empirical evaluations verifying the practical utility and benefits of KSVD and compare with methods resorting to symmetrization or linear SVD across multiple tasks.

Deep Neural Networks (DNN) have shown great promise in many classification applications, yet are widely known to have poorly calibrated predictions when they are over-parametrized. Improving DNN calibration without comprising on model accuracy is of extreme importance and interest in safety critical applications such as in the health-care sector. In this work, we show that decoupling the training of feature extraction layers and classification layers in over-parametrized DNN architectures such as Wide Residual Networks (WRN) and Vision Transformers (ViT) significantly improves model calibration whilst retaining accuracy, and at a low training cost. In addition, we show that placing a Gaussian prior on the last hidden layer outputs of a DNN, and training the model variationally in the classification training stage, even further improves calibration. We illustrate these methods improve calibration across ViT and WRN architectures for several image classification benchmark datasets.

In this manuscript, we investigate the problem of how two-layer neural networks learn features from data, and improve over the kernel regime, after being trained with a single gradient descent step. Leveraging the insight from (Ba et al., 2022), we model the trained network by a spiked Random Features (sRF) model. Further building on recent progress on Gaussian universality (Dandi et al., 2023), we provide an exact asymptotic description of the generalization error of the sRF in the high-dimensional limit where the number of samples, the width, and the input dimension grow at a proportional rate. The resulting characterization for sRFs also captures closely the learning curves of the original network model. This enables us to understand how adapting to the data is crucial for the network to efficiently learn non-linear functions in the direction of the gradient - where at initialization it can only express linear functions in this regime.

A key goal of current mechanistic interpretability research in NLP is to find *linear features* (also called "feature vectors") for transformers: directions in activation space corresponding to concepts that are used by a given model in its computation. Present state-of-the-art methods for finding linear features require large amounts of labelled data -- both laborious to acquire and computationally expensive to utilize. In this work, we introduce a novel method, called "observable propagation" (in short: ObProp), for finding linear features used by transformer language models in computing a given task -- *using almost no data*. Our paradigm centers on the concept of "observables", linear functionals corresponding to given tasks. We then introduce a mathematical theory for the analysis of feature vectors, including a similarity metric between feature vectors called the *coupling coefficient* which estimates the degree to which one feature's output correlates with another's. We use ObProp to perform extensive qualitative investigations into several tasks, including gendered occupational bias, political party prediction, and programming language detection. Our results suggest that ObProp surpasses traditional approaches for finding feature vectors in the low-data regime, and that ObProp can be used to better understand the mechanisms responsible for bias in large language models.

Feature selection is a crucial tool in machine learning and is widely applied across various scientific disciplines. Traditional supervised methods generally identify a universal set of informative features for the entire population. However, feature relevance often varies with context, while the context itself may not directly affect the outcome variable. Here, we propose a novel architecture for contextual feature selection where the subset of selected features is conditioned on the value of *context variables*. Our new approach, Conditional Stochastic Gates (c-STG), models the importance of features using conditional Bernoulli variables whose parameters are predicted based on contextual variables. We introduce a hypernetwork that maps context variables to feature selection parameters to learn the context-dependent gates along with a prediction model. We further present a theoretical analysis of our model, indicating that it can improve performance and flexibility over population-level methods in complex feature selection settings. Finally, we conduct an extensive benchmark using simulated and real-world datasets across multiple domains demonstrating that c-STG can lead to improved feature selection capabilities while enhancing prediction accuracy and interpretability.