Code for "Nishal P. Shah, N. Brackbill, C. Rhoades, A. Tikidji-Hamburyan, G. Goetz, A. Litke, A. Sher, E. P. Simoncelli, E.J. Chichilnisky. Inference of Nonlinear Receptive Field Subunits with Spike-Triggered Clustering. eLife, 2019. https://doi.org/10.7554/eLife.45743"
The code requires Python 2, Tensorflow 1.x, and other standard libraries such as numpy, scipy and pickle.
The data associated with the paper can be downloaded from "Shah, Nishal et al. (2020), Inference of nonlinear receptive field subunits with spike-triggered clustering, v4, Dryad, Dataset, https://doi.org/10.5061/dryad.dncjsxkvk"
The data for a given figure are loaded using pickle.
import pickle
data = pickle.load(open('Figure_2.pkl', 'rb'))
stim_use, resp_use, stim_dim1, stim_dim2 = (data['A']['stimulus'], data['A']['response'],
data['A']['stim_dim1'], data['A']['stim_dim2'])Verify the loaded data.
sta = stim_use.T.dot(resp_use) / np.sum(resp_use)
plt.figure()
plt.imshow(np.reshape(sta[:, 0], [stim_dim1, stim_dim2]),
cmap='gray', interpolation='nearest')
Set fitting parameters and partition the data.
nsub = 5 # number of subunits.
# Take last 10% as testing, and randomly partition the remaining samples for training and validation.
frac_test = 0.1
tms_test = np.arange(np.floor(stim_use.shape[0]*(1 - frac_test)),
1*np.floor(stim_use.shape[0])).astype(np.int)
frac_validate = 0.1
frac_train = 1 - frac_test - frac_validate
perm = np.random.permutation(tms_train_validate)
tms_train_validate = np.arange(0, np.floor(
stim_use.shape[0]*(1 - frac_test))).astype(np.int)
tms_train = perm[0: np.int(np.floor(frac_train * perm.shape[0]))]
tms_validate = perm[np.int(np.floor((1 - frac_validate) * perm.shape[0])): np.int(perm.shape[0])]op = su_model.spike_triggered_clustering(stim_use, resp_use, nsub,
tms_train,
tms_validate,
save_filename_partial='test',
fitting_phases=[1, 2, 3])
k, b, nl_params, lam_log_train, lam_log_validation, fitting_phase, fit_params = opa) Visualize filters :
for isub in range(nsub):
plt.subplot(1, nsub, isub + 1)
plt.imshow(np.reshape(k[:, isub], [stim_dim1, stim_dim2]), cmap='gray', interpolation='nearest')
plt.xticks([])
plt.yticks([])
b) Predict firing rate and loss on test data :
fitting_phase = 1
k, b, nl_params = fit_params[fitting_phase]
pred_test, loss_test = su_model.compute_fr_loss(k, b, stim_use[tms_test, :], resp_use[tms_test, :],
nl_params=nl_params)For locally normalized L1 regularization (promotes spatial locality, see paper) :
# L1 regularization
lam_proj = 0.1
projection_type = 'lnl1'
# Locally normalized L1 regularization
mask = np.ones((stim_dim1, stim_dim2)).astype(np.bool)
neighbor_mat = su_model.get_neighbormat(mask, nbd=1)
lam_proj = 0.1
projection_type = 'lnl1'
# Fit the model for either regularization.
op = su_model.spike_triggered_clustering(stim_use, resp_use, nsub,
tms_train,
tms_validate,
steps_max=10000, eps=1e-9,
projection_type=projection_type,
neighbor_mat=neighbor_mat,
lam_proj=lam_proj, eps_proj=0.01,
save_filename_partial='test',
fitting_phases=[1, 2, 3])
k, b, nl_params, lam_log_train, lam_log_validation, fitting_phase, fit_params = opThe subunits are cleaner (less background noise) for LNL1 regularization.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent