Figure 4.

Simulation results for each distribution case. (A) The synaptic matrix—color-coded density of potentiated synapses from neuron j to neuron i—across “sessions,” reflecting changes in the synaptic configuration during learning. Different matrices columns correspond to different sampling distribution. Each synaptic matrix is plotted at four stages of simulation. Connectivity patterns evolve gradually. The diagonal pattern reflects the high potentiation probability between neighboring neurons. In the multipeak cases potentiated connections are clustered in clouds that correspond to the distribution peaks both in number and scope. Note slower evolution for four peaks than for three peaks, resulting from simulation with a large overlap between neurons. (B) Activity map across “sessions” (rows) and stimulus distributions (columns). Each map reflects the activity level (color-coded) of each neuron (y axis) as a function of the stimuli (x axis; activity is averaged across 20 simulation runs at the 1,000th iteration, i.e., near activity asymptote). The simulation of each activity map was run using its corresponding synaptic matrix (A). After 500 stimulus presentations, there is only weak activity along the diagonal reflecting original stimulus-evoked excitation. In the multipeak cases, further stimulation leads to clustered activity that is correlated with the statistical structure of the stimulation, while in the uniform case unstructured activity is seen. Note the delayed evolution of clustered activity in the four-peak case relative to the three-peak case.