Discriminating visual stimuli from local field potentials recorded with a multi-electrode array in the monkey's visual cortex

• Published in: 2008 IEEE Workshop on Machine Learning for Signal Processing pp. 157 - 162
• Main Authors: Manyakov, N.V., Van Hulle, M.M.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.10.2008
• Subjects: Brain; Delay; Electrodes; Electroencephalography; Electronic mail; Enterprise resource planning; Frequency synchronization; Laboratories; Psychology; System testing
• ISBN: 9781424423750; 1424423759
• ISSN: 1551-2541; 2378-928X
• DOI: 10.1109/MLSP.2008.4685472

We report on the development and testing of a system for classifying two types of visual stimuli from local field potentials (LFPs) recorded with a multi-electrode array chronically implanted in the visual cortical area V4 of a rhesus monkey. The monkey was trained during consecutive training sessions in a classical conditioning paradigm in which one stimulus was consistently paired with a fluid reward and another stimulus not. We first look at the flow of activation in the array by examining the non-linear Granger causality between pairs of electrodes. We observe that, as a function of training, the connectivity increases dramatically, for both stimulus types, making it unsuited for discriminating them. We also looked at the LFP amplitudes for both cases and discovered that the electrodes appear in two groups, depending on the similarity of their LFPs. Changes in synchrony between the two stimuli also mostly occur in the connections between the electrodes from the two groups. This provides us with a first set of discriminative features. The spectra of the LFPs showed a difference in the low and high frequency ranges for the two stimuli, which led us to consider specific coefficients of the wavelet decomposition as a second set of discriminative features. Finally, a classifier is constructed based on the feature scores in the 300 ms interval after stimulus onset, and prior to the possible fluid reward (at 400 ms), so as to avoid any influence of the reward. We obtained 80% classification performance from single trials, thus without any averaging.