Representational dissimilarity matrix (RDM) - /tutorials/W1D5_DeepLearning/student/W1D5_Tutorial3.html - add map from mouse data to model layer

[pas-calc]

In Section 2.1: Representational dissimilarity matrix (RDM) the RDM is computed.
This gives a SxS matrix for S stimuli.

If this matrix is the same for the mouse v1 data and a model layer, that means for every pair of stimuli s1,s2 the correlation corr(s1,s2) is the same in both representations. The data is already z-scored (along the correct axis?) and corr is just a dot product.

Given that both the RDM are equal, then it means there is a dot-product preserving map from mouse data to model layer. This is an isometry.

mousedata_stim_representations = resp_dict["V1 data"]
model_stim_representations = resp_dict["model 'pool' layer"]

map_repr_model_to_mouse = np.linalg.pinv(model_stim_representations) @ mousedata_stim_representations

This is a non-diagonal matrix (N x n, with N = number of model units, n = number of v1 neurons).

I suggest to add this point in the discussion of the data analysis.