Skip to main content
. 2015 Sep 14;10(9):e0137915. doi: 10.1371/journal.pone.0137915

Fig 5. STRF estimated from STA and from GLM.

Fig 5

(A) STRF from STA. Stimulus center frequencies ranged from 97 to 11,234 Hz. Spike-times were binned at 1 ms resolution. (B) Evolution of Spike-count STRF from GLM as a function of increasing λ for L1 norm LASSO: (a) very low values of λ lead to noisy estimates, (c) at very high values of λ all covariates are zero valued, (d) at optimal value of λ chosen from discovery-based selection. (C) Evolution of Spike-count STRF from GLM as a function of increasing λ for L1/L2 norm group LASSO: (e) very low, (g) very high, and (h) optimal values of λ. Optimization selects or removes, simultaneously, all the covariates forming a group. Groups are composed of 4x4, adjacent and non-overlapping covariates. (D) Predicted discharge λ CIF(t|H t) from representative segment of gammatone stimulus, with and without sparse-group regularization. Group-sparse regularized GLM consistently improved the prediction of validation data over non-regularized GLM prediction of expected spike-counts. Correlation coefficients are 0.133 with regularization (red) and 0.066 without regularization (gray). Neural responses from S178, electrode contact #4.