Skip to main content
. 2013 Nov 13;111(3):544–551. doi: 10.1152/jn.00678.2013

Fig. 4.

Fig. 4.

Functional significance of silent synapses. A: histograms of synaptic weights before (black) and after (blue) training with 15 granule cell patterns generated by the Ramp network in Fig. 2A. Bin size is 0.05. B: the fraction of synapses suppressed is almost linear in the number of patterns presented. C: the patterns to be stored have different L1 norms, the summed activity over all granule cells. The “conformity” of a cell is the correlation across the entire set of patterns between activity of the cell and the L1 norm of the pattern. The conformity of each granule cell is calculated for a training set of 15 patterns. Top: the distribution of conformity indices for all granule cells (dashed line) and for those cells having synapses that survive the training process (solid line). Bin size is 0.05. Bottom: the relationship between the conformity of a cell and the likelihood that its synapse will be suppressed during training (prob. of synaptic suppression). D: the L1 norms of the patterns in a training set are broadly distributed (dashed line). The cells for which synapses are suppressed during training form a subpopulation over which L1 norms are even more broadly distributed (gray). In contrast, the distribution of L1 norms over the remaining cells is narrow (solid line). Each distribution is normalized to its own mean. Bin size is 0.05. E, top: cells for which synapses are suppressed during training show correlated activity (gray), whereas the cells with the largest synaptic weights after training show slightly anticorrelated activity (green). Each sample contains 20 cells. Bin size is 0.05. Bottom: as above but for a simulation in which patterns are modified to have uniform L1 norms before training (see methods). To ensure that all results are based on similar distributions of final synaptic weights (including the same proportion of zeroes), training sets for the top and bottom contain 15 unmodified and 40 normalized patterns, respectively. Both simulations resulted in the distribution of final weights shown in A.