Fig. 6: Clustering analysis on the spiking activity of the networks for static and MEMSORN architectures on the counting task with a sequence length of 10.
From: Self-organization of an inhomogeneous memristive hardware for sequence learning
a Principal Component Analysis (PCA) was applied to the firing rate of the two networks in response to 50 sequences of length 10 (600 letters). Each color is assigned to a different position of the letter in the sequence with similar colors encoding the temporal adjacency of the letters in the sequence. In the Principal Component (PC) space, the different input conditions form random clusters in the static network that are not well separated. On the other hand, in MEMSORN network compact clusters are formed which are well separated for different input conditions. b Histogram of the captured variance by the first 20 PCs in the static and MEMSORN networks. The explained variance amounts to 79% for the static network, compared to 87% in the MEMSORN, suggesting more orderly dynamics in MEMSORN network. c Dendrogram of static and MEMSORN networks showing the hierarchical relationship between clusters of neurons. The normalized height of the dendrogram indicates the distance between the clusters and the links indicate the order in which the clusters are joined. For any given distance, the number of branch numbers for MEMSORN are larger than those for the static network, indicating that the clusters in MEMSORN are better structured.
