qianglisinoeusa / basicv1 Goto Github PK

This code implements super-simple, super-fast Hebbian learning of nice V1-like simple-cell receptive fields.

25 lines of Python, runs in a couple seconds (though the RF look even smoother if you run it for longer with a smaller learning rate).

This code uses the "instar" Hebbian rule (see e.g. Vasilkoski et al. IJCNN 2011), SoftMax competition between output cells, and adaptive thresholds to drive firing rates towards 1/NBCELLS (though the actual firing rates vary quite a lot).

The resulting firing is sparse, and most often only one cell fires. But sometimes several cells have significant firing (rarely more than 3 though).

The main program is hebb.py, which can be run as-is. makepatches.py generates the input patches (it requires a directory with many natural images in it - see code). plotRFs.py plots the learned receptive fields.

This repository also includes pre-computed patches of z-scored natural images as inputs, stored as linearized vectors in patches.npy (see makepatches.py).

To plot the resulting RFs, after running hebb.py, run plotRFs.py.