Post on 15-Apr-2017
Non-parametric change point detection for spike trains
Thiago S MosqueiroBioCircuits Institute
University of California San Diego thmosqueiro.vandroiy.com
Conference on Information Sciences and SystemsPrinceton (NJ), 03/15/2016
Take-home messageReaction times of neural populations:
multivariate change-point detection
Electric fish communication:change-point as a time-series segmentation
Complexity of Odorant Time SeriesVergara et al. ‘2013,
Sensors Actuators B 185 462
M. Trincavelli et al. ’2009, Sensors Actuators B 139 165
Picture by Kim S. Mosqueiro (Apr 2015)
Rodriguez-Lujan & J. Fonollosa et al. '2014, Chem and Intell Lab Systems 30 123
Courtesy of M Trincavelli
Change point technique
The (single) change point problem can be stated as the hypothesis testing below:
We are interested in two aspects:How likely is H0 vs H1? Estimate the transition point τ
Change point technique
Divergence:
Solution for the transition time:
Matteson and James ‘2014, J American Statistical Association 109, 334–345.
Mosqueiro & Maia ‘2012, Phys Rev E 88 012712
Neural systemsWe know some coding mechanisms
In insects, anatomy is well documented
Mosqueiro & Huerta ‘2014, Current opinion in insect science
Using all spike trains• To use all spike trains, we
get the first 5 components from PCA
• We then find the change point jointly
Neural reaction times
• No need for proxies and a single general concept
• Use the information of the whole spike train
• Yield much more precise results
• Could be applied to fMRI or EEGs, to jointly find change points within brain regions
• Can be performed on the fly
Fast time scale
• Change points are very close (most of time <2s apart) • Average of 1.6 symbols / sec • To turn it into a symbolic dynamic, we construct features:
(variance, avg slope, area under curve, interval duration)
Clustering of the segments
• Both fish showed similar symbols — cue on vocabulary • Mutual Information drops after bootstrapping/surrogating
Segments showed 3 clusters:
Clustering of the segments
• Both fish showed similar symbols — cue on vocabulary • Mutual Information drops after bootstrapping/surrogating
Segments showed 3 clusters:
Cues to Time-series segmentation
• No need for bins with fixed size
• Coarser time scale may link to behavior
• Clustering symbols seems the same for three different fish — is there a general vocabulary?
• Symbolic dynamics — is there a grammar?
• Current methods are VERY slow for such number of change points
we have a new strategy coming soon…
Free implementation
github.com/VandroiyLabs/chapolins
Parallel, multiple change points implementation in C for efficient of several algorithms
with an API for Python
Logo courtesy of Andre MR Santos
Change Point Library for Non-parametric Statistics