Summary NeuroSurgeon A Toolkit for Subnetwork Analysis arxiv.org
1,687 words - PDF document - View PDF document
One Line
The NeuroSurgeon python library enables subnetwork analysis in neural networks, focusing on Huggingface Transformers, and introduces a visualization of two subnetworks in GPT2.
Slides
Slide Presentation (9 slides)
Key Points
- NeuroSurgeon is a python library developed for subnetwork analysis in neural networks.
- The library supports popular models like ViT, ResNet, GPT2, and BERT.
- It uses optimization-based techniques like Hard-Concrete Masking and Continuous Sparsification to discover functional subnetworks.
- Subnetwork analysis helps uncover the internal structure and high-level functions of trained models.
- NeuroSurgeon includes a visualizer to understand how subnetworks are distributed throughout the layers of a model.
Summaries
36 word summary
NeuroSurgeon is a python library for subnetwork analysis in neural networks, specifically those in the Huggingface Transformers library. The document discusses the goal of subnetwork analysis and introduces a visualization of two subnetworks within a GPT2
76 word summary
NeuroSurgeon is a python library developed to facilitate subnetwork analysis in neural networks. It allows researchers to discover and manipulate subnetworks within models in the Huggingface Transformers library. The goal of subnetwork analysis is to understand the internal
The document discusses subnetwork analysis for mechanistic interpretability. It references several studies and introduces a visualization of two subnetworks within a GPT2-style transformer. The transformer was trained on addition and multiplication tasks, and one subnetwork was optimized for