Information Technology Reference
In-Depth Information
th
= 30 msec to reflect the average spike precision in bursts, as was shown in
Harris
et al.
(2003) and Bonifazi
et al.
(2005). We also tested other values of
th
,
proving the robustness of our classification method to this parameter where the
results are shown in Raichman and Ben-Jacob (2008).
The computed similarity matrix contains the complete data on the SBE-
to-SBE relations. However, in order to detect and define the number
M
of
well-defined sets of SBEs with similar activation pattern, we need to develop a
measure that evaluates the goodness of each set of SBE motif. Such a proposed
method is described in length in Raichman and Ben-Jacon (2008), introducing a
simple procedure to detect centers of SBE clusters with high similarity, by
applying a two-stage method that uses a hierarchical clustering algorithm
followed by an iterative search for independent cluster centers. By re-ordering
the similarity matrix, one can visualize the clustered data, where similar SBEs are
grouped into square areas along the matrix diagonal with high similarity values.
12.9 Network Repertoire
In Fig. 12.7 we show a clustering process of one culture. As is shown in the
figure, the similarity matrix of the examined culture breaks into five distinctive
clusters of SBEs with high similarity of activation profile (Fig. 12.7A and B).
The different motif sets are marked on the right of the reordered similarity matrix
(Fig. 12.7B). In this example we have chosen a set size of 20 SBEs, from a
sample of 300 consecutive SBEs. In Fig. 12.7C plotted are the similarity value
between every SBE that belongs to one of the five sets and the mean similarity
matrix that represents each of the sets. The values show that inside the clusters
the SBEs are highly similar to their representative mean activation matrices, and
are not similar to any of the other sets. Figure12.7D shows the raster of the
activation pattern (first spikes only of each neuron) for each of the identified
motifs, where each motif has a different propagation profile across the culture.
Figure12.7E plots the appearance of each motif along time. The order of
appearance of the different motifs is rather random and almost uniform. Thus,
there is no dominant motif that takes over the network in this specific culture.
We have applied our motif detecting method on different cultures with
different spatial morphologies (Raichman and Ben-Jacob, 2008). In large
uniform cultures, 50% showed to have at least two distinguished motifs. Motifs
were also detected in smaller cultures and in clustered ones, where the clustered
cultures showed an increased number of motifs, proportional to the number of
clusters.
