ATMAN: Understanding Transformer Predictions Through Memory Efficient Attention Manipulation

Authors: Björn Deiseroth, Mayukh Deb, Samuel Weinbach, Manuel Brack, Patrick Schramowski, Kristian Kersting

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our exhaustive experiments on text and image-text benchmarks demonstrate that ATMAN outperforms current state-of-the-art gradient-based methods on several metrics and models while being computationally efficient.
Researcher Affiliation Collaboration Björn Deiseroth1,2,3 Mayukh Deb1 Samuel Weinbach1 Manuel Brack2,4 Patrick Schramowski2,3,4,5 Kristian Kersting2,3,4 1Aleph Alpha 2Technical University Darmstadt 3Hessian Center for Artificial Intelligence (hessian.AI) 4German Center for Artificial Intelligence (DFKI) 5LAION
Pseudocode No The paper describes the proposed method through text and diagrams (e.g., Fig. 2a), but it does not include a formal pseudocode or algorithm block.
Open Source Code Yes Source code: https://github.com/Aleph-Alpha/At Man
Open Datasets Yes For evaluation, we used the Stanford Question Answering (QA) Dataset (SQu AD) [23].
Dataset Splits No The paper describes sampling strategies for its evaluation (e.g., 'randomly sample 200 images per class on the filtered set') but does not specify explicit train/validation/test dataset splits with percentages, counts, or predefined splits.
Hardware Specification Yes Fig. 5 illustrates the runtime and memory consumption on a single NVIDIA A100 80GB GPU.
Software Dependencies No The paper mentions using 'Captum' as a library for integrating some methods but does not provide specific version numbers for Captum or other software dependencies.
Experiment Setup Yes We fixed the parameter κ = 0.7 of Eq. 6 and f = 0.9 of Eq. 4 throughout this work. They were empirically concluded by running a line sweep on a randomly sampled subset of the Open Images dataset once, c.f. Fig. 11.