Additive Synaptic Noise.

<p>Figure shows the effect of additive uniformly distributed synaptic noise on the network performance by setting (see Equation 9). Panels A and B show the network performance without and with lateral connections (as in previous figures) respectively. The plots of average reward (left column, solid line) are calculated as in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0018539#pone-0018539-g003" target="_blank">Figures 3</a> and <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0018539#pone-0018539-g005" target="_blank">5</a> showing learning curves over 9 blocks of 512 trials. The red dashed line shows the values without noise from <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0018539#pone-0018539-g003" target="_blank">Figure 3</a> (systems A and B correspond) for direct comparison. Similarly, the plots of average error (right column, solid line) are calculated as in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0018539#pone-0018539-g004" target="_blank">Figure 4</a>. The red dashed line shows the values without noise from <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0018539#pone-0018539-g004" target="_blank">Figure 4</a> (again, systems A and B respectively). We can observe that both the average reward and average error performance measures show that the system without lateral connections is far more robust to noise applied directly to the synaptic weight.