deeplocalizer 0.0.1

Package to localize torch deep learning models

DeepLocalizer: Quickly find functional specialization in deep neural networks.

Extends The LLM Language Network: A Neuroscientific Approach for Identifying Causally Task-Relevant Units and Brain-like functional specialization emerges spontaneously in deep neural networks to other models and data.

Examples:

• Face Localizer Resnet Report
• Face Localizer Resnet Notebook

Roadmap

• Replicate some parts of original paper (https://github.com/xnought/paper-implement/tree/main/language_network)
• Write code from scratch to do analysis on face with resnet
  • Set up face localizer example w/ goal of applying to a resnet model
    • 5k positive (faces) from CelebA
    • 5k negative (objects) from COCO
  • Extract activations from the resnet model
    • test track the activations
    • store activations on disk
  • Contrast positive vs negative activations
  • Ablation w/ statistical tests on resnet
    • write code to ablate torch models easily
    • ablate given the top percent face activations
    • Compare performance after ablation
• Write general API from most helpful functions so others can easily use the library
  • Activation computation
  • Analysis computation
    • Top percent global
    • Visualizations
    • Ablate model with the top percent
    • Compute statistics on ablated model
• Write report on the resnet example and if localization seems to work and what evidence (here -> https://www.donnybertucci.com/project/deeplocalizer)

Usage

API

See resnet34_example.ipynb for doing localization on a torch model with a custom dataset/task.

uv add deeplocalizer

Or if you are old school

Install

pip install deeplocalizer

Development

cd deeplocalizer # this git repo

Make sure to have https://docs.astral.sh/uv/ installed.

Install and Run

uv sync
uv run deeplocalizer/deeplocalizer.py

or run an example python notebook within the .env generated.

References

papers

• https://arxiv.org/abs/2411.02280
• https://web.mit.edu/bcs/nklab/media/pdfs/Dobs_2022.pdf

code/datasets

• https://huggingface.co/datasets/tpremoli/CelebA-attrs
• https://huggingface.co/datasets/phiyodr/coco2017
• https://huggingface.co/microsoft/resnet-50