Neural decoding with SVM and feature selection in a rat active tactile discrimination task

Understanding, mapping, and repairing the central nervous system would benefit from a better comprehension of how sensory stimuli information is conveyed by the neural activity underlying animal behavior. In this way, a promising approach, neural decoding, include machine-learning methods that assoc...

Full description

Saved in:
Bibliographic Details
Main Authors: Gajadhar, Andy Anand, Moioli, Renan Cipriano, de Melo, Bianca Karla Amorim Sousa, Kunicki, Ana Carolina Bione, Peres, Andre Salles Cunha, Rego, Thais Gaudencio do
Format: Conference Proceeding
Language:English
Subjects:
Decoding
Feature extraction
Feature Selection
Machine Learning
Neural Decoding
Neurons
Rats
Support Vector Machine
Support vector machines
Task analysis
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Understanding, mapping, and repairing the central nervous system would benefit from a better comprehension of how sensory stimuli information is conveyed by the neural activity underlying animal behavior. In this way, a promising approach, neural decoding, include machine-learning methods that associate patterns of brain activity with sensory stimuli. By solving a regression or classification problem, decoding contributes to highlighting the features from neural activity that best relate to a target variable. In this work, we built a support vector machine decoder coupled with a recursive feature selection algorithm to investigate the relevance of three distinct aspects of neural activity in an active tactile discrimination experiment involving rodents. Considering simultaneous recordings from a population of single-neuron spikes and local field potentials from the prefrontal, posterior parietal, primary sensory and visual cortices of nine Long-Evans rats, we were able to decode animal behavioral choices with approximately 95% of accuracy. When only the 30 most informative features from neural activity were used, accuracy values marginally decreased, indicating that a reduced set of features convey most of stimuli information, at least from the decoder perspective. Importantly, all regions contributed with features, and no single feature was associated to high classification accuracy. In addition to further unveiling the distributed nature of information processing in the brain, our results suggest that efficient brain-machine interfaces may be built with a reduced set of features.
ISSN:2161-4407
DOI:10.1109/IJCNN.2018.8489474