An official website of the United States government
Vision transformers (ViTs) have recently set off a new wave in neural architecture design thanks to their record-breaking performance in various vision tasks. In parallel, to fulfill the goal of deploying ViTs into real-world vision applications, their robustness against potential malicious attacks has gained increasing attention. In particular, recent works show that ViTs are more robust against adversarial attacks as compared with convolutional neural networks (CNNs), and conjecture that this is because ViTs focus more on capturing global interactions among different input/feature patches, leading to their improved robustness to local perturbations imposed by adversarial attacks. In this work, we ask an intriguing question: “Under what kinds of perturbations do ViTs become more vulnerable learners compared to CNNs?” Driven by this question, we first conduct a comprehensive experiment regarding the robustness of both ViTs and CNNs under various existing adversarial attacks to understand the underlying reason favoring their robustness. Based on the drawn insights, we then propose a dedicated attack framework, dubbed Patch-Fool, that fools the self-attention mechanism by attacking its basic component (i.e., a single patch) with a series of attention-aware optimization techniques. Interestingly, our Patch-Fool framework shows for the first time that ViTs are not necessarily more robust than CNNs against adversarial perturbations. In particular, we find that ViTs are more vulnerable learners compared with CNNs against our Patch-Fool attack which is consistent across extensive experiments, and the observations from Sparse/Mild Patch-Fool, two variants of Patch-Fool, indicate an intriguing insight that the perturbation density and strength on each patch seem to be the key factors that influence the robustness ranking between ViTs and CNNs. It can be expected that our Patch-Fool framework will shed light on both future architecture designs and training schemes for robustifying ViTs towards their real-world deployment. Our codes are available at https://github.com/RICE-EIC/Patch-Fool.
more »
« less
This content will become publicly available on July 1, 2026
Symbolic Rule Extraction From Attention-Guided Sparse Representations in Vision Transformers
Abstract Recentneuro-symbolic approaches have successfully extracted symbolic rule-sets from Convolutional Neural Network-based models to enhance interpretability. However, applying similar techniques to Vision Transformers (ViTs) remains challenging due to their lack of modular concept detectors and reliance on global self-attention mechanisms. We propose a framework for symbolic rule extraction from ViTs by introducing a sparse concept layer inspired by Sparse Autoencoders (SAEs). This linear layer operates on attention-weighted patch representations and learns a disentangled, binarized representation in which individual neurons activate for high-level visual concepts. To encourage interpretability, we apply a combination of L1 sparsity, entropy minimization, and supervised contrastive loss. These binarized concept activations are used as input to the FOLD-SE-M algorithm, which generates a rule-set in the form of a logic program. Our method achieves abetter classification accuracy than the standard ViT while enabling symbolic reasoning. Crucially, the extracted rule-set is not merely post-hoc but acts as a logic-based decision layer that operates directly on the sparse concept representations. The resulting programs are concise and semantically meaningful. This work is the first to extract executable logic programs from ViTs using sparse symbolic representations, providing a step forward in interpretable and verifiable neuro-symbolic AI.
more »
« less
- Award ID(s):
- 1910131
- PAR ID:
- 10649080
- Publisher / Repository:
- Cambridge University Press
- Date Published:
- Journal Name:
- Theory and Practice of Logic Programming
- Volume:
- 25
- Issue:
- 4
- ISSN:
- 1471-0684
- Page Range / eLocation ID:
- 722 to 738
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Motivated by the increasing importance of general-purpose Graphic Processing Units (GPGPU) programming, exemplified by NVIDIA’s CUDA framework, as well as the difficulty, especially for novice programmers, of reasoning about performance in GPGPU kernels, we introduce a novel quantitative program logic for CUDA kernels. The logic allows programmers to reason about both functional correctness and resource usage of CUDA kernels, paying particular attention to a set of common but CUDA-specific performance bottlenecks: warp divergences, uncoalesced memory accesses, and bank conflicts. The logic is proved sound with respect to a novel operational cost semantics for CUDA kernels. The semantics, logic, and soundness proofs are formalized in Coq. An inference algorithm based on LP solving automatically synthesizes symbolic resource bounds by generating derivations in the logic. This algorithm is the basis of RaCUDA, an end-to-end resource-analysis tool for kernels, which has been implemented using an existing resource-analysis tool for imperative programs. An experimental evaluation on a suite of benchmarks shows that the analysis is effective in aiding the detection of performance bugs in CUDA kernels.more » « less
-
The majority of computer algebra systems (CAS) support symbolic integration using a combination of heuristic algebraic and rule-based (integration table) methods. In this paper, we present a hybrid (symbolic-numeric) method to calculate the indefinite integrals of univariate expressions. Our method is broadly similar to the Risch-Norman algorithm. The primary motivation for this work is to add symbolic integration functionality to a modern CAS (the symbolic manipulation packages of SciML, the Scientific Machine Learning ecosystem of the Julia programming language), which is designed for numerical and machine learning applications. The symbolic part of our method is based on the combination of candidate terms generation (ansatz generation using a methodology borrowed from the Homotopy operators theory) combined with rule-based expression transformations provided by the underlying CAS. The numeric part uses sparse regression, a component of the Sparse Identification of Nonlinear Dynamics (SINDy) technique, to find the coefficients of the candidate terms. We show that this system can solve a large variety of common integration problems using only a few dozen basic integration rules.more » « less
-
General-purpose programming on GPUs (GPGPU) is becoming increasingly in vogue as applications such as machine learning and scientific computing demand high throughput in vector-parallel applications. NVIDIA's CUDA toolkit seeks to make GPGPU programming accessible by allowing programmers to write GPU functions, called kernels, in a small extension of C/C++. However, due to CUDA's complex execution model, the performance characteristics of CUDA kernels are difficult to predict, especially for novice programmers. This paper introduces a novel quantitative program logic for CUDA kernels, which allows programmers to reason about both functional correctness and resource usage of CUDA kernels, paying particular attention to a set of common but CUDA-specific performance bottlenecks. The logic is proved sound with respect to a novel operational cost semantics for CUDA kernels. The semantics, logic and soundness proofs are formalized in Coq. An inference algorithm based on LP solving automatically synthesizes symbolic resource bounds by generating derivations in the logic. This algorithm is the basis of RaCuda, an end-to-end resource-analysis tool for kernels, which has been implemented using an existing resource-analysis tool for imperative programs. An experimental evaluation on a suite of CUDA benchmarks shows that the analysis is effective in aiding the detection of performance bugs in CUDA kernels.more » « less
-
Vision Transformers (ViTs) have achieved state-of-the-art performance on various vision tasks. However, ViTs’ self-attention module is still arguably a major bottleneck, limiting their achievable hardware efficiency and more extensive applications to resource constrained platforms. Meanwhile, existing accelerators dedicated to NLP Transformers are not optimal for ViTs. This is because there is a large difference between ViTs and Transformers for natural language processing (NLP) tasks: ViTs have a relatively fixed number of input tokens, whose attention maps can be pruned by up to 90% even with fixed sparse patterns, without severely hurting the model accuracy (e.g., <=1.5% under 90% pruning ratio); while NLP Transformers need to handle input sequences of varying numbers of tokens and rely on on-the-fly predictions of dynamic sparse attention patterns for each input to achieve a decent sparsity (e.g., >=50%). To this end, we propose a dedicated algorithm and accelerator co-design framework dubbed ViTCoD for accelerating ViTs. Specifically, on the algorithm level, ViTCoD prunes and polarizes the attention maps to have either denser or sparser fixed patterns for regularizing two levels of workloads without hurting the accuracy, largely reducing the attention computations while leaving room for alleviating the remaining dominant data movements; on top of that, we further integrate a lightweight and learnable auto-encoder module to enable trading the dominant high-cost data movements for lower-cost computations. On the hardware level, we develop a dedicated accelerator to simultaneously coordinate the aforementioned enforced denser and sparser workloads for boosted hardware utilization, while integrating on-chip encoder and decoder engines to leverage ViTCoD’s algorithm pipeline for much reduced data movements. Extensive experiments and ablation studies validate that ViTCoD largely reduces the dominant data movement costs, achieving speedups of up to 235.3×, 142.9×, 86.0×, 10.1×, and 6.8× over general computing platforms CPUs, EdgeGPUs, GPUs, and prior-art Transformer accelerators SpAtten and Sanger under an attention sparsity of 90%, respectively. Our code implementation is available at https://github.com/GATECH-EIC/ViTCoD.more » « less
