Figure 8.

The DI was computed with shuffled network connectivity to assess the importance of the learned connectivity for pattern discrimination. (A) The results from the six control experiments (B: layer 1 → 2 synapses shuffled shown in red; C: layer 2 (E or I) to layer 2 (E or I) shown in blue; D—layer 2 (I) to layer 2 (E or I) shown in cyan; E—layer 2 (E) to layer 2 (E or I) shown in brown; F—layer 2 (E) to layer 3 (E) shown in magenta; G—all these connections shuffled simultaneously) are compared against the original network with learned connectivity (A) shown in green. The DI scores represent an average obtained after shuffling 10 times for each control experiment. The plot shows that altering the connectivity by shuffling the connections between layer 1 to E neurons in layer 2 affects discriminability severely while shuffling connections within layer 2 does not (see text for more details). (B) A control experiment was also performed to assess the importance of the specific locations of the firing activity within the readout code. This was achieved by first randomly shuffling the components of the readout code 10 different times. The DI scores were then computed for each case and then averaged. The results shows that the DI score is much lower (red trace) compared to the original network (green trace) suggesting that the locations of the firing activity within the readout code caused due to learning is also very important. (C) The average firing rate of the network for the various control experiments in (A) is shown here. (D) The average firing rate for the control experiment in (B) is shown here. The two firing rates overlap completely (red and green overlap completely). This is because there is no change in the connections between layer 2 and layer 3 neurons after STDP potentiates all of them over time early within the first hour. Furthermore, all connections go to its max value (i.e., unimodal distribution). So, swapping the outputs does not make a difference to the firing rates.