论文检索

Applications/Computer Vision

Shi Chen, Ming Jiang, Qi Zhao

In recent years, deep saliency models have made significant progress in predicting human visual attention. However, the mechanisms behind their success remain largely unexplained due to the opaque nature of deep neural networks. In this paper, we present a novel analytic framework that sheds light on the implicit features learned by saliency models and provides principled interpretation and quantification of their contributions to saliency prediction. Our approach decomposes these implicit features into interpretable bases that are explicitly aligned with semantic attributes and reformulates saliency prediction as a weighted combination of probability maps connecting the bases and saliency. By applying our framework, we conduct extensive analyses from various perspectives, including the positive and negative weights of semantics, the impact of training data and architectural designs, the progressive influences of fine-tuning, and common error patterns of state-of-the-art deep saliency models. Additionally, we demonstrate the effectiveness of our framework by exploring visual attention characteristics in various application scenarios, such as the atypical attention of people with autism spectrum disorder, attention to emotion-eliciting stimuli, and attention evolution over time. Our code is publicly available at \url{https://github.com/szzexpoi/saliency_analysis}.

Miscellaneous Aspects of Machine Learning/Everything Else

Robin Sasse, Aniek Eijpe, Jona Ruthardt, Elias Dubbeldam

Scope of Reproducibility — CartoonX [1] is a novel explanation method for image classifiers. In this reproducibility study, we examine the claims of the original authors of CartoonX that it: (i) extracts relevant piece‐wise smooth parts of the image, resulting in explanations which are more straightforward to interpret for humans; (ii) achieves lower distortion in the model output, using fewer coefficients than other state-of‐the‐art methods; (iii) is model‐agnostic. Finally, we examine how to reduce the runtime.Methodology — The original authors’ open‐sourced implementation has been used to examine (i). We implemented the code to examine (ii), as there was no public code available for this. We tested claim (iii) by performing the same experiments with a Vision Transformer instead of a CNN. To reduce the runtime, we extended the existing implementation with multiple enhanced initialization techniques. All experiments took approximately 38.4 hours on a single NVIDIA Titan RTX.Results — Our results support the claims made by the original authors. (i) We observe that CartoonX produces piece‐wise smooth explanations. Most of the explanations give valuable insights. (ii) Most experiments, that show how CartoonX achieves lower distortion outputs compared to other methods, have been reproduced. In the cases where exact reproducibility has not been achieved, claim (ii) of the author still holds. (iii) The model‐agnosticism claim still holds as the overall quality of the ViT‐based explanations almost matches that of the CNN‐based explanations. Finally, simple heuristical initializations did not improve the runtime.What was easy — The mathematical background and intuition of CartoonX were clearly explained by the original authors. Moreover, the original author’s code was well structured and documented, which made it straightforward to run and extend.What was difficult — Some hyperparameter settings and implementation details needed to reproduce the experiments were not clear or transparent from the original paper or code. This made it difficult to implement and reproduce these experiments.Communication with original authors — We have been in brief communication with the original authors. They were able to address most of our points, providing us with some additional clarifications about the exact implementation and hyperparameter settings.

Applications/Neuroscience, Cognitive Science

Christos Sourmpis, Carl Petersen, Wulfram Gerstner, Guillaume Bellec

Simultaneous behavioral and electrophysiological recordings call for new methods to reveal the interactions between neural activity and behavior. A milestone would be an interpretable model of the co-variability of spiking activity and behavior across trials. Here, we model a mouse cortical sensory-motor pathway in a tactile detection task reported by licking with a large recurrent spiking neural network (RSNN), fitted to the recordings via gradient-based optimization. We focus specifically on the difficulty to match the trial-to-trial variability in the data. Our solution relies on optimal transport to define a distance between the distributions of generated and recorded trials. The technique is applied to artificial data and neural recordings covering six cortical areas. We find that the resulting RSNN can generate realistic cortical activity and predict jaw movements across the main modes of trial-to-trial variability. Our analysis also identifies an unexpected mode of variability in the data corresponding to task-irrelevant movements of the mouse.

Social Aspects/Trustworthy Machine Learning

Lu Yan, Zhuo Zhang, Guanhong Tao, Kaiyuan Zhang, Xuan Chen, Guangyu Shen, Xiangyu Zhang

Backdoor attacks have emerged as a prominent threat to natural language processing (NLP) models, where the presence of specific triggers in the input can lead poisoned models to misclassify these inputs to predetermined target classes. Current detection mechanisms are limited by their inability to address more covert backdoor strategies, such as style-based attacks. In this work, we propose an innovative test-time poisoned sample detection framework that hinges on the interpretability of model predictions, grounded in the semantic meaning of inputs.We contend that triggers (e.g., infrequent words) are not supposed to fundamentally alter the underlying semantic meanings of poisoned samples as they want to stay stealthy. Based on this observation, we hypothesize that while the model's predictions for paraphrased clean samples should remain stable, predictions for poisoned samples should revert to their true labels upon the mutations applied to triggers during the paraphrasing process.We employ ChatGPT, a state-of-the-art large language model, as our paraphraser and formulate the trigger-removal task as a prompt engineering problem. We adopt fuzzing, a technique commonly used for unearthing software vulnerabilities, to discover optimal paraphrase prompts that can effectively eliminate triggers while concurrently maintaining input semantics.Experiments on 4 types of backdoor attacks, including the subtle style backdoors, and 4 distinct datasets demonstrate that our approach surpasses baseline methods, including STRIP, RAP, and ONION, in precision and recall.

Optimization/First-order

Guillaume Wang, Lénaïc Chizat

We study the convergence to local Nash equilibria of gradient methods for two-player zero-sum differentiable games.It is well-known that, in the continuous-time setting, such dynamics converge locally when $S \succ 0$ and may diverge when $S=0$, where $S\succeq 0$ is the symmetric part of the Jacobian at equilibrium that accounts for the "potential" component of the game. We show that these dynamics also converge as soon as $S$ is nonzero (*partial curvature*) and the eigenvectors of the antisymmetric part $A$ are in general position with respect to the kernel of $S$.We then study the convergence rate when $S \ll A$ and prove that it typically depends on the *average* of the eigenvalues of $S$, instead of the minimum as an analogy with minimization problems would suggest.To illustrate our results, we consider the problem of computing mixed Nash equilibria of continuous games. We show that, thanks to partial curvature, conic particle methods -- which optimize over both weights and supports of the mixed strategies -- generically converge faster than fixed-support methods.For min-max games, it is thus beneficial to add degrees of freedom "with curvature": this can be interpreted as yet another benefit of over-parameterization.

Applications/Health

Jungwoo Oh, Gyubok Lee, Seongsu Bae, Joon-myoung Kwon, Edward Choi

Question answering (QA) in the field of healthcare has received much attention due to significant advancements in natural language processing. However, existing healthcare QA datasets primarily focus on medical images, clinical notes, or structured electronic health record tables. This leaves the vast potential of combining electrocardiogram (ECG) data with these systems largely untapped. To address this gap, we present ECG-QA, the first QA dataset specifically designed for ECG analysis. The dataset comprises a total of 70 question templates that cover a wide range of clinically relevant ECG topics, each validated by an ECG expert to ensure their clinical utility. As a result, our dataset includes diverse ECG interpretation questions, including those that require a comparative analysis of two different ECGs. In addition, we have conducted numerous experiments to provide valuable insights for future research directions. We believe that ECG-QA will serve as a valuable resource for the development of intelligent QA systems capable of assisting clinicians in ECG interpretations.

Miscellaneous Aspects of Machine Learning/Causality

Jann Spiess, guido imbens, Amar Venugopal

Motivated by a recent literature on the double-descent phenomenon in machine learning, we consider highly over-parameterized models in causal inference, including synthetic control with many control units. In such models, there may be so many free parameters that the model fits the training data perfectly. We first investigate high-dimensional linear regression for imputing wage data and estimating average treatment effects, where we find that models with many more covariates than sample size can outperform simple ones. We then document the performance of high-dimensional synthetic control estimators with many control units. We find that adding control units can help improve imputation performance even beyond the point where the pre-treatment fit is perfect. We provide a unified theoretical perspective on the performance of these high-dimensional models. Specifically, we show that more complex models can be interpreted as model-averaging estimators over simpler ones, which we link to an improvement in average performance. This perspective yields concrete insights into the use of synthetic control when control units are many relative to the number of pre-treatment periods.

Miscellaneous Aspects of Machine Learning/Kernel methods

Yuka Hashimoto, Masahiro Ikeda, Hachem Kadri

Reproducing kernel Hilbert $C^*$-module (RKHM) is a generalization of reproducing kernel Hilbert space (RKHS) by means of $C^*$-algebra, and the Perron-Frobenius operator is a linear operator related to the composition of functions. Combining these two concepts, we present deep RKHM, a deep learning framework for kernel methods. We derive a new Rademacher generalization bound in this setting and provide a theoretical interpretation of benign overfitting by means of Perron-Frobenius operators. By virtue of $C^*$-algebra, the dependency of the bound on output dimension is milder than existing bounds. We show that $C^*$-algebra is a suitable tool for deep learning with kernels, enabling us to take advantage of the product structure of operators and to provide a clear connection with convolutional neural networks. Our theoretical analysis provides a new lens through which one can design and analyze deep kernel methods.

Farrukh Baratov, Goksenin Yuksel, Darie Petcu, Jan Bakker

Scope of reproducibility - We study the reproducibility of the paper "Quantifying Societal Bias Amplification in Image Captioning" by Hirota et al. In this paper, the authors propose a new metric to measure bias amplification, called LIC, and evaluate it on multiple image captioning models. Based on this evaluation, they make the following main claims which we aim to verify: (1) all models amplify gender bias, (2) all models amplify racial bias, (3) LIC is robust against encoders, and (4) the NIC+Equalizer model increases gender bias with respect to the baseline. We also extend upon the original work by evaluating LIC for age bias.Methodology - For our reproduction, we were able to run the code provided by the authors without any modifications. For our extension, we automatically labelled the images in the dataset with age annotations and adjusted the code to work with this dataset. In total, 38 GPU hours were needed to perform all experiments.Results - The reproduced results are close to the original results and support all four main claims. Furthermore, our additional results show that only a subset of the models amplifies age bias, while they strengthen the claim that LIC is robust against encoders. However, we acknowledge that our extension to age bias has its limitations.What was easy - The author's code and the data needed to run it are publicly available. The code required no modification to run and the scripts were provided with an extensive argument parser, allowing us to quickly set up our experiments. Moreover, the details of the original experiments were clearly stated in the appendix.What was difficult - We found that it was difficult to interpret the author's code as the provided documentation contained room for improvement. Also, the scripts contained repetitive code. While the authors retrained all image captioning models, they did not share the model weights, making it difficult to extend upon their work.Communication with original authors - No (attempt at) communication with the original authors was performed.

Social Aspects/Accountability, Transparency and Interpretability

Chiyu Ma, Brandon Zhao, Chaofan Chen, Cynthia Rudin

We present ProtoConcepts, a method for interpretable image classification combining deep learning and case-based reasoning using prototypical parts. Existing work in prototype-based image classification uses a "this looks like that'' reasoning process, which dissects a test image by finding prototypical parts and combining evidence from these prototypes to make a final classification. However, all of the existing prototypical part-based image classifiers provide only one-to-one comparisons, where a single training image patch serves as a prototype to compare with a part of our test image. With these single-image comparisons, it can often be difficult to identify the underlying concept being compared (e.g., "is it comparing the color or the shape?''). Our proposed method modifies the architecture of prototype-based networks to instead learn prototypical concepts which are visualized using multiple image patches. Having multiple visualizations of the same prototype allows us to more easily identify the concept captured by that prototype (e.g., "the test image and the related training patches are all the same shade of blue''), and allows our model to create richer, more interpretable visual explanations. Our experiments show that our ``this looks like those'' reasoning process can be applied as a modification to a wide range of existing prototypical image classification networks while achieving comparable accuracy on benchmark datasets.

Social Aspects/Accountability, Transparency and Interpretability

Shibal Ibrahim, Gabriel Afriat, Kayhan Behdin, Rahul Mazumder

Generalized Additive Models (GAMs) are a family of flexible and interpretable models with old roots in statistics. GAMs are often used with pairwise interactions to improve model accuracy while still retaining flexibility and interpretability but lead to computational challenges as we are dealing with order of $p^2$ terms. It is desirable to restrict the number of components (i.e., encourage sparsity) for easier interpretability, and better computational and statistical properties. Earlier approaches, considering sparse pairwise interactions, have limited scalability, especially when imposing additional structural interpretability constraints. We propose a flexible GRAND-SLAMIN framework that can learn GAMs with interactions under sparsity and additional structural constraints in a differentiable end-to-end fashion. We customize first-order gradient-based optimization to perform sparse backpropagation to exploit sparsity in additive effects for any differentiable loss function in a GPU-compatible manner. Additionally, we establish novel non-asymptotic prediction bounds for our estimators with tree-based shape functions. Numerical experiments on real-world datasets show that our toolkit performs favorably in terms of performance, variable selection and scalability when compared with popular toolkits to fit GAMs with interactions. Our work expands the landscape of interpretable modeling while maintaining prediction accuracy competitive with non-interpretable black-box models. Our code is available at https://github.com/mazumder-lab/grandslamin.

Miscellaneous Aspects of Machine Learning/Causality

Naoki Egami, Musashi Hinck, Brandon Stewart, Hanying Wei

In computational social science (CSS), researchers analyze documents to explain social and political phenomena. In most scenarios, CSS researchers first obtain labels for documents and then explain labels using interpretable regression analyses in the second step. One increasingly common way to annotate documents cheaply at scale is through large language models (LLMs). However, like other scalable ways of producing annotations, such surrogate labels are often imperfect and biased. We present a new algorithm for using imperfect annotation surrogates for downstream statistical analyses while guaranteeing statistical properties—like asymptotic unbiasedness and proper uncertainty quantification—which are fundamental to CSS research. We show that direct use of surrogate labels in downstream statistical analyses leads to substantial bias and invalid confidence intervals, even with high surrogate accuracy of 80-90\%. To address this, we build on debiased machine learning to propose the design-based supervised learning (DSL) estimator. DSL employs a doubly-robust procedure to combine surrogate labels with a smaller number of high-quality, gold-standard labels. Our approach guarantees valid inference for downstream statistical analyses, even when surrogates are arbitrarily biased and without requiring stringent assumptions, by controlling the probability of sampling documents for gold-standard labeling. Both our theoretical analysis and experimental results show that DSL provides valid statistical inference while achieving root mean squared errors comparable to existing alternatives that focus only on prediction without inferential guarantees.

Applications/Chemistry and Drug Discovery

Chaoran Cheng, Jian Peng

We propose a general architecture that combines the coefficient learning scheme with a residual operator layer for learning mappings between continuous functions in the 3D Euclidean space. Our proposed model is guaranteed to achieve SE(3)-equivariance by design. From the graph spectrum view, our method can be interpreted as convolution on graphons (dense graphs with infinitely many nodes), which we term InfGCN. By leveraging both the continuous graphon structure and the discrete graph structure of the input data, our model can effectively capture the geometric information while preserving equivariance. Through extensive experiments on large-scale electron density datasets, we observed that our model significantly outperformed the current state-of-the-art architectures. Multiple ablation studies were also carried out to demonstrate the effectiveness of the proposed architecture.

Social Aspects/Accountability, Transparency and Interpretability

Max Torop, Aria Masoomi, Davin Hill, Kivanc Kose, Stratis Ioannidis, Jennifer Dy

Several recent methods for interpretability model feature interactions by looking at the Hessian of a neural network. This poses a challenge for ReLU networks, which are piecewise-linear and thus have a zero Hessian almost everywhere. We propose SmoothHess, a method of estimating second-order interactions through Stein's Lemma. In particular, we estimate the Hessian of the network convolved with a Gaussian through an efficient sampling algorithm, requiring only network gradient calls. SmoothHess is applied post-hoc, requires no modifications to the ReLU network architecture, and the extent of smoothing can be controlled explicitly. We provide a non-asymptotic bound on the sample complexity of our estimation procedure. We validate the superior ability of SmoothHess to capture interactions on benchmark datasets and a real-world medical spirometry dataset.

Applications/Time Series

Xu Liu, Yutong Xia, Yuxuan Liang, Junfeng Hu, Yiwei Wang, LEI BAI, Chao Huang, Zhenguang Liu, Bryan Hooi, Roger Zimmermann

Road traffic forecasting plays a critical role in smart city initiatives and has experienced significant advancements thanks to the power of deep learning in capturing non-linear patterns of traffic data. However, the promising results achieved on current public datasets may not be applicable to practical scenarios due to limitations within these datasets. First, the limited sizes of them may not reflect the real-world scale of traffic networks. Second, the temporal coverage of these datasets is typically short, posing hurdles in studying long-term patterns and acquiring sufficient samples for training deep models. Third, these datasets often lack adequate metadata for sensors, which compromises the reliability and interpretability of the data. To mitigate these limitations, we introduce the LargeST benchmark dataset. It encompasses a total number of 8,600 sensors in California with a 5-year time coverage and includes comprehensive metadata. Using LargeST, we perform in-depth data analysis to extract data insights, benchmark well-known baselines in terms of their performance and efficiency, and identify challenges as well as opportunities for future research. We release the datasets and baseline implementations at: https://github.com/liuxu77/LargeST.

Social Aspects/Accountability, Transparency and Interpretability

Biagio La Rosa, Leilani Gilpin, Roberto Capobianco

Compositional Explanations is a method for identifying logical formulas of concepts that approximate the neurons' behavior. However, these explanations are linked to the small spectrum of neuron activations (i.e., the highest ones) used to check the alignment, thus lacking completeness. In this paper, we propose a generalization, called Clustered Compositional Explanations, that combines Compositional Explanations with clustering and a novel search heuristic to approximate a broader spectrum of the neuron behavior. We define and address the problems connected to the application of these methods to multiple ranges of activations, analyze the insights retrievable by using our algorithm, and propose desiderata qualities that can be used to study the explanations returned by different algorithms.

Miscellaneous Aspects of Machine Learning/Representation Learning

Xuanjie Liu, Daniel Chin, Yichen Huang, Gus Xia

We have recently seen great progress in learning interpretable music representations, ranging from basic factors, such as pitch and timbre, to high-level concepts, such as chord and texture. However, most methods rely heavily on music domain knowledge. It remains an open question what general computational principles *give rise to* interpretable representations, especially low-dim factors that agree with human perception. In this study, we take inspiration from modern physics and use *physical symmetry* as a self-consistency constraint for the latent space. Specifically, it requires the prior model that characterises the dynamics of the latent states to be *equivariant* with respect to certain group transformations. We show that physical symmetry leads the model to learn a *linear* pitch factor from unlabelled monophonic music audio in a self-supervised fashion. In addition, the same methodology can be applied to computer vision, learning a 3D Cartesian space from videos of a simple moving object without labels. Furthermore, physical symmetry naturally leads to *counterfactual representation augmentation*, a new technique which improves sample efficiency.

Domen Mohorčič, David Ocepek

Scope of Reproducibility: The paper presents a novel post-hoc regularization technique for tree-based models, called Hierarchical Shrinkage (Agarwal 2022). Our main goal is to confirm the claims that it substantially increases the predictive performance of both decision trees and random forests, that it is faster than other regularization techniques, and that it makes the interpretation of random forests simpler. Methodology: In our reproduction, we used the Hierarchical Shrinkage, provided by the authors in the Python package imodels. We also used their function for obtaining pre-cleaned data sets. While the algorithm code and clean datasets were provided we re-implemented the experiments as well as added additional experiments to further test the validity of the claims. The results were tested by applying Hierarchical Shrinkage to different tree models and comparing them to the authors' results. Results: We managed to reproduce most of the results the authors get. The method works well and most of the claims are supported. The method does increase the predictive performance of tree-based models most of the time, but not always. When compared to other regularization techniques the Hierarchical Shrinkage outperforms them when used on decision trees, but not when used on random forests. Since the method is applied after learning, it is extremely fast. And it does simplify decision boundaries for random forests, making them easier to interpret. What was easy: The use of the official code for Hierarchical Shrinkage was straightforward and used the same function naming convention as other machine learning libraries. The function for acquiring already clean data sets saved a lot of time. What was difficult: The authors also provided the code for their experiments in a separate library, but the code did not run out of the box and we had no success reproducing the results with it. The code was inconsistent with the paper methodology. We had the most problems with hyperparameter tuning. The authors did not specify how they tuned the hyperparameters for the used RF regularizers. Communication with original authors: We did not contact the authors of the original paper

Social Aspects/Accountability, Transparency and Interpretability

Amit Dhurandhar, Karthikeyan Natesan Ramamurthy, Kartik Ahuja, Vijay Arya

Locally interpretable model agnostic explanations (LIME) method is one of the most popular methods used to explain black-box models at a per example level. Although many variants have been proposed, few provide a simple way to produce high fidelity explanations that are also stable and intuitive. In this work, we provide a novel perspective by proposing a model agnostic local explanation method inspired by the invariant risk minimization (IRM) principle -- originally proposed for (global) out-of-distribution generalization -- to provide such high fidelity explanations that are also stable and unidirectional across nearby examples. Our method is based on a game theoretic formulation where we theoretically show that our approach has a strong tendency to eliminate features where the gradient of the black-box function abruptly changes sign in the locality of the example we want to explain, while in other cases it is more careful and will choose a more conservative (feature) attribution, a behavior which can be highly desirable for recourse. Empirically, we show on tabular, image and text data that the quality of our explanations with neighborhoods formed using random perturbations are much better than LIME and in some cases even comparable to other methods that use realistic neighbors sampled from the data manifold. This is desirable given that learning a manifold to either create realistic neighbors or to project explanations is typically expensive or may even be impossible. Moreover, our algorithm is simple and efficient to train, and can ascertain stable input features for local decisions of a black-box without access to side information such as a (partial) causal graph as has been seen in some recent works.

Gerson de Kleuver, David Bikker, Wenhua Hu, Bram Veenman

Scope of Reproducibility — This paper analyses the reproducibility of the study Proto2Proto: Can you recognize the car, the way I do? The main contributions and claims of the study are: 1) Using Proto2Proto, a shallower student model is more faithful to the teacher in terms of interpretability than a baseline student model while also showing the same or better accuracy; 2) Global Explanation loss forces student prototypes to be close to teacher prototypes; 3) Patch‐Prototype Correspondence loss enforces the local representations of the student to be similar to those of the teacher; 4) The proposed evaluation metrics determine the faithfulness of the student to the teacher in terms of interpretability.Methodology — A public code repository was available for the paper, which provided a working but incomplete and minimally documented codebase. With some modifications we were able to carry out the experiments that were best supported by the codebase. We spent a total of 60 computational GPU hours on reproduction.Results — The results we were able to produce support claim 1, albeit weakly. Further results are in line with the paper, but we found them to go against claim 3. In addition, we carried out a theoretical analysis which provides support for claim 4. Finally, we were unable to carry out our intended experiment to verify claim 2. What was easy — The original paper was clearly structured and understandable. The experiments for which configurations were provided were simple to conduct.What was difficult — The public codebase contained minimal documentation. Moreover, the use of variable names did not correspond between the code and the paper. Furthermore, the codebase lacked elements vital to reproducing some experiments. Another significant constraint were the computational requirements needed to reproduce the original experiments. Finally, the code required to reproduce one of the visualizations was not provided.Communication with original authors — We contacted the authors to ask for trained model weights and missing hyperparameters for several experiments. We did not receive aresponse.