NDN4-tf2-pytorch-exploration

Exploration into replacing/reimplementing tools provided by NeuroTheoryUMD/NDN3 with/in TF2 and/or PyTorch. Starting to contain reimplementation of bunch of msc-neuro models.

So far only a personal playground without any goals.

Ideas to explore (no particular order):

• Laplacian2D regularization
• Pearson's R metric
• Non-DoG models (rLN, rLNLN, conv) from msc-neuro
• Rudimentary performance evaluation
  • Seems to be ~fast/slow for msc-neuro-like models
• DoG layer
• Reimplementation of baseline 4 from msc-neuro
  • Match loss computation 100 %
• NDN3's data filters
• Consider cleaning up tools ^^ and packaging them

Setup:

• Create & activate virtual environment python3 -m venv env (or conda, or ...)
• Install required packages pip install -r ./requirements.txt
• Hydrate ./Data

Data:

Scripts assume freely available data from Model Constrained by Visual Hierarchy Improves Prediction of Neural Responses to Natural Scenes paper.

• Navigate to Supporting Information
• Download the first supplement
• Unzip it to ./ (to have folder ./Data in repo root with three subdirectories region1 to region3)