Figure 3:

Comparing k-tree and FCNN models With or Without Synapse nonlinearity. Architecture ranges from low input multi-synaptic input repetitions (k=1) to high input repetitions (k=32) and mean accuracy of each model is displayed with standard error bars. Number of trials ranges from 1–10, and trial accuracy was omitted if the training was deemed to fail (accuracy threshold 0.55). Table 3 lists trial counts. Red: FCNN with Synapses, Pink: FCNN without Synapses, Black: k-tree with Synapses, Grey: k-tree without Synapses.