This title appears in the Scientific Report :
2013
Spike Pattern Detection by Frequent Itemset Mining
Spike Pattern Detection by Frequent Itemset Mining
Information processing in the cortex was suggested to be organized in cell assemblies (Hebb, 1949), i.e. groups of correlated neurons (cell assemblies) forming building blocks of information processing. Modern massively parallel extracellular recordings (Buzsaki 2004; see Symposium 24) make such net...
Saved in:
Personal Name(s): | Torre, Emiliano (Corresponding author) |
---|---|
Denker, Michael / Borgelt, Christian / Picado-Muino, David / Gerstein, George / Grün, Sonja | |
Contributing Institute: |
Computational and Systems Neuroscience; IAS-6 Computational and Systems Neuroscience; INM-6 |
Imprint: |
2013
|
Conference: | Tenth Goettigen Meeting of the German Neuroscience Society, Goettingen (Germany), 2013-03-13 - 2013-09-16 |
Document Type: |
Poster |
Research Program: |
Brain-inspired multiscale computation in neuromorphic hybrid systems Supercomputing and Modelling for the Human Brain Helmholtz Alliance on Systems Biology Signalling Pathways and Mechanisms in the Nervous System |
Publikationsportal JuSER |
Information processing in the cortex was suggested to be organized in cell assemblies (Hebb, 1949), i.e. groups of correlated neurons (cell assemblies) forming building blocks of information processing. Modern massively parallel extracellular recordings (Buzsaki 2004; see Symposium 24) make such network organization increasingly accessible. However, the analysis of such massively parallel spike trains (MPSTs) for the presence of active assemblies cannot be achieved by straightforward extension of existing methods (e.g. the Unitary Events (UE) analysis, Grün et al, 2002; Grün, 2009), since the evaluation of individual spike patterns would lead to combinatorial explosion. In addition, testing the significance of all patterns leads to a multiple testing problem, with the consequence of false positives. On the other hand, approaching such data by only pairwise analysis with subsequent clustering led to the finding of distinct clusters of mutually correlated neurons (Berger et al, 2007). However, by this approach it is not possible to conclude on higher-order correlated groups of neurons.
Therefore we make here use of frequent itemset mining (FIM; Goethals, 2010) to efficiently count coincident spike patterns in MPSTs (Borgelt et al, in prep). This approach accretes more and more neurons thereby building an efficient search tree that avoids redundant searches, in contrast to the accretion algorithm by Gerstein et al (1978). By defining a minimum occurrence count (support) frequent pattern are detected. Further, we consider only ”closed patterns”, i.e. patterns that cannot be trivially explained by the occurrences of supersets containing them. The number of patterns considered by this procedure is considerably lower -though generally still large- than all possible combinations of patterns for N neurons (2Nfor N=100 ). The multiple testing problem is further downsized by pooling closed frequent patterns into groups of same complexity (number of spikes in the pattern) and number of occurrences (“pattern spectrum”, Gerstein et al, 2012). The significance of each entry in the spectrum is assessed through surrogates obtained by spike dithering (Gerstein, 2004). Only patterns corresponding to significant entries are considered and further evaluated for overlapping patterns. We calibrate the algorithm on simulated spike trains containing assemblies of synchronous spiking neurons (SIP correlations, Kuhn et al, 2003). Finally, we illustrate the application to massively parallel spike trains from monkey motor cortex (see also Poster by Zehl et al.), and compare our findings to results obtained by pairwise (Berger et al,2007) and UE (Grün et al, 2002) analysis. |