Door: NeuroAI¶

Start here¶

• The ResNet-18 running example on one image.
• Extract features for a stimulus set: per-layer activations in stimulus order, written with the model's identity and the preprocessing.
• See the graph and inspect any operation.

Going further¶

• The NeuroAI feature-extraction journey: load, verify the input side, address the layer, extract at scale, hand the features on, prove it.
• Task RNN circuits: train a recurrent network on a memory task, edit windows of timesteps, find its fixed points with torch.autograd.
• A protein language model, traced: ESM-2 attention maps, a per-residue readout and a contact map.
• Loading models for feature extraction and bring your own alignment.
• torchlens.neuro and the brain-benchmarking seam: rsatoolbox and Brain-Score are where the analysis runs.
• Receptive and projective fields.

Coming from another tool¶

thingsvision, torchvision feature extraction, plain forward hooks. Where another tool is the better choice: where to go instead.