Fig. 5: Decoding within and generalization across recording sessions.
Displayed is the classification accuracy, a, computed for individual datasets and b, generalization accuracy across recording sessions. This analysis requires a correspondence of features across them. The units were identified after spike sorting, and can therefore be different across individual sessions. Thus, for this analysis, the spiking activity recorded from each of the 96 sites recorded with the array served as features, on which we trained and tested the classifier across the recording sessions. Results are presented for both the flash suppression (Perceptual dominance - BFS) and the rivalry phase (Perceptual dominance - BR) of the paradigm computed with neural activity elicited during an 800 ms window starting 200 ms after a stimulus or a perceptual change. The main diagonal (upper left to lower right) represents classification accuracy within a session and lateral diagonals represent generalization accuracy across sessions. While we find on average, high accuracy within individual recording sessions, the generalization across datasets is limited. Moreover, the classification accuracy within datasets was significantly more compared to generalization accuracy across datasets (two-sample t-test, two-sided, Physical Alternationwithin vs. Physical Alternationacross, p = 2.42*10−9, Perceptual Dominance (BFS)within vs. Perceptual Dominance (BFS)across, p = 1.94*10−5, Physical Dominancewithin vs. Physical Dominanceacross, p = 7.45*10−10, Perceptual Dominance (BR)within vs. Perceptual Dominance (BR)across, p = 1.07*10−7). Potential reasons for such limited generalization could be that units either change or lose their preference, or it is likely, that we sampled from different units across recording sessions.
