arXiv subjects

Pascal Frossard

Publications and source records attributed to Pascal Frossard.

At least 91 records · Page 5Linked to original sources

arXiv2022

On the Granularity of Explanations in Model Agnostic NLP Interpretability

Current methods for Black-Box NLP interpretability, like LIME or SHAP, are based on altering the text to interpret by removing words and modeling the Black-Box response. In this paper, we outline limitations of this approach when using complex BERT-based classifiers: The word-based sampling produces texts that are out-of-distribution for the classifier and further gives rise to a high-dimensional search space, which can't be sufficiently explored when time or computation power is limited. Both of these challenges can be addressed by using segments as elementary building blocks for NLP interpretability. As illustration, we show that the simple choice of sentences greatly improves on both of these challenges. As a consequence, the resulting explainer attains much better fidelity on a benchmark classification task.

cs.CL↗

arXiv2022

Fooling Explanations in Text Classifiers

State-of-the-art text classification models are becoming increasingly reliant on deep neural networks (DNNs). Due to their black-box nature, faithful and robust explanation methods need to accompany classifiers for deployment in real-life scenarios. However, it has been shown in vision applications that explanation methods are susceptible to local, imperceptible perturbations that can significantly alter the explanations without changing the predicted classes. We show here that the existence of such perturbations extends to text classifiers as well. Specifically, we introduceTextExplanationFooler (TEF), a novel explanation attack algorithm that alters text input samples imperceptibly so that the outcome of widely-used explanation methods changes considerably while leaving classifier predictions unchanged. We evaluate the performance of the attribution robustness estimation performance in TEF on five sequence classification datasets, utilizing three DNN architectures and three transformer architectures for each dataset. TEF can significantly decrease the correlation between unchanged and perturbed input attributions, which shows that all models and explanation methods are susceptible to TEF perturbations. Moreover, we evaluate how the perturbations transfer to other model architectures and attribution methods, and show that TEF perturbations are also effective in scenarios where the target model and explanation method are unknown. Finally, we introduce a semi-universal attack that is able to compute fast, computationally light perturbations with no knowledge of the attacked classifier nor explanation method. Overall, our work shows that explanations in text classifiers are very fragile and users need to carefully address their robustness before relying on them in critical applications.

cs.LG↗

arXiv2022

Hierarchical Spherical CNNs with Lifting-based Adaptive Wavelets for Pooling and Unpooling

Pooling and unpooling are two essential operations in constructing hierarchical spherical convolutional neural networks (HS-CNNs) for comprehensive feature learning in the spherical domain. Most existing models employ downsampling-based pooling, which will inevitably incur information loss and cannot adapt to different spherical signals and tasks. Besides, the preserved information after pooling cannot be well restored by the subsequent unpooling to characterize the desirable features for a task. In this paper, we propose a novel framework of HS-CNNs with a lifting structure to learn adaptive spherical wavelets for pooling and unpooling, dubbed LiftHS-CNN, which ensures a more efficient hierarchical feature learning for both image- and pixel-level tasks. Specifically, adaptive spherical wavelets are learned with a lifting structure that consists of trainable lifting operators (i.e., update and predict operators). With this learnable lifting structure, we can adaptively partition a signal into two sub-bands containing low- and high-frequency components, respectively, and thus generate a better down-scaled representation for pooling by preserving more information in the low-frequency sub-band. The update and predict operators are parameterized with graph-based attention to jointly consider the signal's characteristics and the underlying geometries. We further show that particular properties are promised by the learned wavelets, ensuring the spatial-frequency localization for better exploiting the signal's correlation in both spatial and frequency domains. We then propose an unpooling operation that is invertible to the lifting-based pooling, where an inverse wavelet transform is performed by using the learned lifting operators to restore an up-scaled representation. Extensive empirical evaluations on various spherical domain tasks validate the superiority of the proposed LiftHS-CNN.

cs.CV↗

arXiv2022

Empirical Advocacy of Bio-inspired Models for Robust Image Recognition

Deep convolutional neural networks (DCNNs) have revolutionized computer vision and are often advocated as good models of the human visual system. However, there are currently many shortcomings of DCNNs, which preclude them as a model of human vision. There are continuous attempts to use features of the human visual system to improve the robustness of neural networks to data perturbations. We provide a detailed analysis of such bio-inspired models and their properties. To this end, we benchmark the robustness of several bio-inspired models against their most comparable baseline DCNN models. We find that bio-inspired models tend to be adversarially robust without requiring any special data augmentation. Additionally, we find that bio-inspired models beat adversarially trained models in the presence of more real-world common corruptions. Interestingly, we also find that bio-inspired models tend to use both low and mid-frequency information, in contrast to other DCNN models. We find that this mix of frequency information makes them robust to both adversarial perturbations and common corruptions.

cs.CV↗

arXiv2022

Top-N: Equivariant set and graph generation without exchangeability

This work addresses one-shot set and graph generation, and, more specifically, the parametrization of probabilistic decoders that map a vector-shaped prior to a distribution over sets or graphs. Sets and graphs are most commonly generated by first sampling points i.i.d. from a normal distribution, and then processing these points along with the prior vector using Transformer layers or Graph Neural Networks. This architecture is designed to generate exchangeable distributions, i.e., all permutations of the generated outputs are equally likely. We however show that it only optimizes a proxy to the evidence lower bound, which makes it hard to train. We then study equivariance in generative settings and show that non-exchangeable methods can still achieve permutation equivariance. Using this result, we introduce Top-n creation, a differentiable generation mechanism that uses the latent vector to select the most relevant points from a trainable reference set. Top-n can replace i.i.d. generation in any Variational Autoencoder or Generative Adversarial Network. Experimentally, our method outperforms i.i.d. generation by 15% at SetMNIST reconstruction, by 33% at object detection on CLEVR, generates sets that are 74% closer to the true distribution on a synthetic molecule-like dataset, and generates more valid molecules on QM9.

cs.LG↗

arXiv2022

A Structured Dictionary Perspective on Implicit Neural Representations

Implicit neural representations (INRs) have recently emerged as a promising alternative to classical discretized representations of signals. Nevertheless, despite their practical success, we still do not understand how INRs represent signals. We propose a novel unified perspective to theoretically analyse INRs. Leveraging results from harmonic analysis and deep learning theory, we show that most INR families are analogous to structured signal dictionaries whose atoms are integer harmonics of the set of initial mapping frequencies. This structure allows INRs to express signals with an exponentially increasing frequency support using a number of parameters that only grows linearly with depth. We also explore the inductive bias of INRs exploiting recent results about the empirical neural tangent kernel (NTK). Specifically, we show that the eigenfunctions of the NTK can be seen as dictionary atoms whose inner product with the target signal determines the final performance of their reconstruction. In this regard, we reveal that meta-learning has a reshaping effect on the NTK analogous to dictionary learning, building dictionary atoms as a combination of the examples seen during meta-training. Our results permit to design and tune novel INR architectures, but can also be of interest for the wider deep learning theory community.

cs.LG↗

arXiv2022

U-Boost NAS: Utilization-Boosted Differentiable Neural Architecture Search

Optimizing resource utilization in target platforms is key to achieving high performance during DNN inference. While optimizations have been proposed for inference latency, memory footprint, and energy consumption, prior hardware-aware neural architecture search (NAS) methods have omitted resource utilization, preventing DNNs to take full advantage of the target inference platforms. Modeling resource utilization efficiently and accurately is challenging, especially for widely-used array-based inference accelerators such as Google TPU. In this work, we propose a novel hardware-aware NAS framework that does not only optimize for task accuracy and inference latency, but also for resource utilization. We also propose and validate a new computational model for resource utilization in inference accelerators. By using the proposed NAS framework and the proposed resource utilization model, we achieve 2.8 - 4x speedup for DNN inference compared to prior hardware-aware NAS methods while attaining similar or improved accuracy in image classification on CIFAR-10 and Imagenet-100 datasets.

cs.LG↗

arXiv2022

Scale-out Systolic Arrays

Multi-pod systolic arrays are emerging as the architecture of choice in DNN inference accelerators. Despite their potential, designing multi-pod systolic arrays to maximize effective throughput/Watt (i.e., throughput/Watt adjusted when accounting for array utilization) poses a unique set of challenges. In this work, we study three key pillars in multi-pod systolic array designs, namely array granularity, interconnect, and tiling. We identify optimal array granularity across workloads and show that state-of-the-art commercial accelerators use suboptimal array sizes for single-tenancy workloads. We, then evaluate the bandwidth/latency trade-offs in interconnects and show that Butterfly networks offer a scalable topology for accelerators with a large number of pods. Finally, we introduce a novel data tiling scheme with custom partition size to maximize utilization in optimally sized pods. We propose Scale-out Systolic Arrays, a multi-pod inference accelerator for both single- and multi-tenancy based on these three pillars. We show that SOSA exhibits scaling of up to 600 TeraOps/s in effective throughput for state-of-the-art DNN inference workloads, and outperforms state-of-the-art multi-pod accelerators by a factor of 1.5x.

cs.AR↗

arXiv2022

On the benefits of knowledge distillation for adversarial robustness

Knowledge distillation is normally used to compress a big network, or teacher, onto a smaller one, the student, by training it to match its outputs. Recently, some works have shown that robustness against adversarial attacks can also be distilled effectively to achieve good rates of robustness on mobile-friendly models. In this work, however, we take a different point of view, and show that knowledge distillation can be used directly to boost the performance of state-of-the-art models in adversarial robustness. In this sense, we present a thorough analysis and provide general guidelines to distill knowledge from a robust teacher and boost the clean and adversarial performance of a student model even further. To that end, we present Adversarial Knowledge Distillation (AKD), a new framework to improve a model's robust performance, consisting on adversarially training a student on a mixture of the original labels and the teacher outputs. Through carefully controlled ablation studies, we show that using early-stopping, model ensembles and weak adversarial training are key techniques to maximize performance of the student, and show that these insights generalize across different robust distillation techniques. Finally, we provide insights on the effect of robust knowledge distillation on the dynamics of the student network, and show that AKD mostly improves the calibration of the network and modify its training dynamics on samples that the model finds difficult to learn, or even memorize.

cs.LG↗

arXiv2022

PRIME: A few primitives can boost robustness to common corruptions

Despite their impressive performance on image classification tasks, deep networks have a hard time generalizing to unforeseen corruptions of their data. To fix this vulnerability, prior works have built complex data augmentation strategies, combining multiple methods to enrich the training data. However, introducing intricate design choices or heuristics makes it hard to understand which elements of these methods are indeed crucial for improving robustness. In this work, we take a step back and follow a principled approach to achieve robustness to common corruptions. We propose PRIME, a general data augmentation scheme that relies on simple yet rich families of max-entropy image transformations. PRIME outperforms the prior art in terms of corruption robustness, while its simplicity and plug-and-play nature enable combination with other methods to further boost their robustness. We analyze PRIME to shed light on the importance of the mixing strategy on synthesizing corrupted images, and to reveal the robustness-accuracy trade-offs arising in the context of common corruptions. Finally, we show that the computational efficiency of our method allows it to be easily used in both on-line and off-line data augmentation schemes.

cs.CV↗

arXiv2022

Block-Sparse Adversarial Attack to Fool Transformer-Based Text Classifiers

Recently, it has been shown that, in spite of the significant performance of deep neural networks in different fields, those are vulnerable to adversarial examples. In this paper, we propose a gradient-based adversarial attack against transformer-based text classifiers. The adversarial perturbation in our method is imposed to be block-sparse so that the resultant adversarial example differs from the original sentence in only a few words. Due to the discrete nature of textual data, we perform gradient projection to find the minimizer of our proposed optimization problem. Experimental results demonstrate that, while our adversarial attack maintains the semantics of the sentence, it can reduce the accuracy of GPT-2 to less than 5% on different datasets (AG News, MNLI, and Yelp Reviews). Furthermore, the block-sparsity constraint of the proposed optimization problem results in small perturbations in the adversarial example.

cs.CL↗

arXiv2022

FAR: A General Framework for Attributional Robustness

Attribution maps are popular tools for explaining neural networks predictions. By assigning an importance value to each input dimension that represents its impact towards the outcome, they give an intuitive explanation of the decision process. However, recent work has discovered vulnerability of these maps to imperceptible adversarial changes, which can prove critical in safety-relevant domains such as healthcare. Therefore, we define a novel generic framework for attributional robustness (FAR) as general problem formulation for training models with robust attributions. This framework consist of a generic regularization term and training objective that minimize the maximal dissimilarity of attribution maps in a local neighbourhood of the input. We show that FAR is a generalized, less constrained formulation of currently existing training methods. We then propose two new instantiations of this framework, AAT and AdvAAT, that directly optimize for both robust attributions and predictions. Experiments performed on widely used vision datasets show that our methods perform better or comparably to current ones in terms of attributional robustness while being more generally applicable. We finally show that our methods mitigate undesired dependencies between attributional robustness and some training and estimation parameters, which seem to critically affect other competitor methods.

cs.LG↗

arXiv2022

Data augmentation with mixtures of max-entropy transformations for filling-level classification

We address the problem of distribution shifts in test-time data with a principled data augmentation scheme for the task of content-level classification. In such a task, properties such as shape or transparency of test-time containers (cup or drinking glass) may differ from those represented in the training data. Dealing with such distribution shifts using standard augmentation schemes is challenging and transforming the training images to cover the properties of the test-time instances requires sophisticated image manipulations. We therefore generate diverse augmentations using a family of max-entropy transformations that create samples with new shapes, colors and spectral characteristics. We show that such a principled augmentation scheme, alone, can replace current approaches that use transfer learning or can be used in combination with transfer learning to improve its performance.

cs.LG↗

arXiv2022

Landmarking for Navigational Streaming of Stored High-Dimensional Media

Modern media data such as 360 videos and light field (LF) images are typically captured in much higher dimensions than the observers' visual displays. To efficiently browse high-dimensional media over bandwidth-constrained networks, a navigational streaming model is considered: a client navigates the large media space by dictating a navigation path to a server, who in response transmits the corresponding pre-encoded media data units (MDU) to the client one-by-one in sequence. Intra-coding an MDU (I-MDU) would result in a large bitrate but I-MDU can be randomly accessed, while inter-coding an MDU (P-MDU) using another MDU as a predictor incurs a small coding cost but imposes an order where the predictor must be first transmitted and decoded. From a compression perspective, the technical challenge is: how to achieve coding gain via inter-coding of MDUs, while enabling adequate random access for satisfactory user navigation. To address this problem, we propose landmarks, a selection of key MDUs from the high-dimensional media. Using landmarks as predictors, nearby MDUs in local neighborhoods are intercoded, resulting in a predictive MDU structure with controlled coding cost. It means that any requested MDU can be decoded by at most transmitting a landmark and an inter-coded MDU, enabling navigational random access. To build a landmarked MDU structure, we employ tree-structured vector quantizer (TSVQ) to first optimize landmark locations, then iteratively add/remove inter-coded MDUs as refinements using a fast branch-and-bound technique. Taking interactive LF images and viewport adaptive 360 images as illustrative applications, and I-, P- and previously proposed merge frames to intra- and inter-code MDUs, we show experimentally that landmarked MDU structures can noticeably reduce the expected transmission cost compared with MDU structures without landmarks.

cs.MM↗

arXiv2022

Graph Neural Networks With Lifting-based Adaptive Graph Wavelets

Spectral-based graph neural networks (SGNNs) have been attracting increasing attention in graph representation learning. However, existing SGNNs are limited in implementing graph filters with rigid transforms (e.g., graph Fourier or predefined graph wavelet transforms) and cannot adapt to signals residing on graphs and tasks at hand. In this paper, we propose a novel class of graph neural networks that realizes graph filters with adaptive graph wavelets. Specifically, the adaptive graph wavelets are learned with neural network-parameterized lifting structures, where structure-aware attention-based lifting operations (i.e., prediction and update operations) are developed to jointly consider graph structures and node features. We propose to lift based on diffusion wavelets to alleviate the structural information loss induced by partitioning non-bipartite graphs. By design, the locality and sparsity of the resulting wavelet transform as well as the scalability of the lifting structure are guaranteed. We further derive a soft-thresholding filtering operation by learning sparse graph representations in terms of the learned wavelets, yielding a localized, efficient, and scalable wavelet-based graph filters. To ensure that the learned graph representations are invariant to node permutations, a layer is employed at the input of the networks to reorder the nodes according to their local topology information. We evaluate the proposed networks in both node-level and graph-level representation learning tasks on benchmark citation and bioinformatics graph datasets. Extensive experiments demonstrate the superiority of the proposed networks over existing SGNNs in terms of accuracy, efficiency, and scalability.

cs.LG↗

arXiv2021

Distributed Graph Learning with Smooth Data Priors

Graph learning is often a necessary step in processing or representing structured data, when the underlying graph is not given explicitly. Graph learning is generally performed centrally with a full knowledge of the graph signals, namely the data that lives on the graph nodes. However, there are settings where data cannot be collected easily or only with a non-negligible communication cost. In such cases, distributed processing appears as a natural solution, where the data stays mostly local and all processing is performed among neighbours nodes on the communication graph. We propose here a novel distributed graph learning algorithm, which permits to infer a graph from signal observations on the nodes under the assumption that the data is smooth on the target graph. We solve a distributed optimization problem with local projection constraints to infer a valid graph while limiting the communication costs. Our results show that the distributed approach has a lower communication cost than a centralised algorithm without compromising the accuracy in the inferred graph. It also scales better in communication costs with the increase of the network size, especially for sparse networks.

eess.SP↗

arXiv2021

FGOT: Graph Distances based on Filters and Optimal Transport

Graph comparison deals with identifying similarities and dissimilarities between graphs. A major obstacle is the unknown alignment of graphs, as well as the lack of accurate and inexpensive comparison metrics. In this work we introduce the filter graph distance. It is an optimal transport based distance which drives graph comparison through the probability distribution of filtered graph signals. This creates a highly flexible distance, capable of prioritising different spectral information in observed graphs, offering a wide range of choices for a comparison metric. We tackle the problem of graph alignment by computing graph permutations that minimise our new filter distances, which implicitly solves the graph comparison problem. We then propose a new approximate cost function that circumvents many computational difficulties inherent to graph comparison and permits the exploitation of fast algorithms such as mirror gradient descent, without grossly sacrificing the performance. We finally propose a novel algorithm derived from a stochastic version of mirror gradient descent, which accommodates the non-convexity of the alignment problem, offering a good trade-off between performance accuracy and speed. The experiments on graph alignment and classification show that the flexibility gained through filter graph distances can have a significant impact on performance, while the difference in speed offered by the approximation cost makes the framework applicable in practical settings.

cs.LG↗

arXiv2021

What can linearized neural networks actually say about generalization?

For certain infinitely-wide neural networks, the neural tangent kernel (NTK) theory fully characterizes generalization, but for the networks used in practice, the empirical NTK only provides a rough first-order approximation. Still, a growing body of work keeps leveraging this approximation to successfully analyze important deep learning phenomena and design algorithms for new applications. In our work, we provide strong empirical evidence to determine the practical validity of such approximation by conducting a systematic comparison of the behavior of different neural networks and their linear approximations on different tasks. We show that the linear approximations can indeed rank the learning complexity of certain tasks for neural networks, even when they achieve very different performances. However, in contrast to what was previously reported, we discover that neural networks do not always perform better than their kernel approximations, and reveal that the performance gap heavily depends on architecture, dataset size and training task. We discover that networks overfit to these tasks mostly due to the evolution of their kernel during training, thus, revealing a new type of implicit bias.

cs.LG↗