Imperial College London
Alternate

DrKonstantinNikolic

Faculty of Engineering, Department of Electrical and Electronic Engineering

Visiting Professor
 
 
 
//

Contact

 

k.nikolic

 
 
//

Location

 

Bessemer 420CBessemer BuildingSouth Kensington Campus

//

Summary

 

Publications

Citation

BibTex format

@inproceedings{Nikolic:2016,
author = {Nikolic, K and Evans, B},
title = {Identifying optimal feature transforms for classification and prediction in biological systems: recovering receptive field vectors from sparse recordings},
url = {http://hdl.handle.net/10044/1/44218},
year = {2016}
}
Download

RIS format (EndNote, RefMan)

TY  - CPAPER
AB  - With biological systems it is often hard to adequately sample the entire input space. With sensoryneural systems this can be a particularly acute problem, with very high dimensional natural inputs andtypically sparse spiking outputs. Here we present an information theory based approach to analysespiking data of an early sensory pathway, demonstrated on retinal ganglion cells (RGC) responding tonatural visual scene stimuli (Katz et al., 2016). We used a non-parametric technique based on theconcept of mutual information (MI), in particular, Quadratic Mutual Information (QMI). The QMIallowed us to very efficiently search the high dimensional space formed by the visual input for a muchsmaller dimensional subspace of Receptive Field Vectors (RFV). RFVs give the most informationabout the response of the cell to natural stimuli. This approach allows us to identify the RFVs far moreefficiently using limited data as we can search the complete stimulus space for multiple vectorssimultaneously. The RFVs were also used to predict the RGCs’ responses to any natural stimuli.Another suitable area of application of this algorithm is in diagnostic inference. Currently we areadapting the method to be used for identifying the cancer markers in the volatile organic compoundspresent in exhaled breath. Once the maximally informative features are established they can be usedfor diagnostic predictions on new breath samples. Preliminary results of the breathomics analysis willbe discussed at the conference.There are several other potential applications such as multiclass categorisation for bacterial strainsusing ISFET arrays for DNA sequencing. This algorithm can be part of a rapid point-of-care device foridentifying the specific infectious agents and recommending appropriate antibiotics.Here we will focus on presenting the algorithm using the example of RFVs of RGCs.
AU  - Nikolic,K
AU  - Evans,B
PY  - 2016///
TI  - Identifying optimal feature transforms for classification and prediction in biological systems: recovering receptive field vectors from sparse recordings
UR  - http://hdl.handle.net/10044/1/44218
ER  -
Download