Reconstructing the perceptual organization of sound from neural responses
No Thumbnail Available
Background: Sets of stimuli can span different stimulus spaces. Examples include linear, circular or planar. Since the neural system does not know the geometry of this stimulus space, it needs to have a way of estimating it from information contained in the neural population responses. Recently, a set of techniques was proposed that can achieve this estimation, known as representational similarity analysis (RSA). Methods: We expand the current framework of RSA by establishing a criterion for taking into account the local geometry of the neural response manifold. We refer to this expansion as gRSA (global RSA). To do so, we compute distances between stimuli within the response manifold (Local Distance Matrix, LDM). Once pairwise distances have been identified, we reconstruct the global geometry from the local geometry by recreating the neighborhoods of the manifold (Global Distance Matrix, GDM). The GDM is constructed by stochastic exploration of the LDM. Once a certain value of cross correlation is established two neighbors are identified based on a local decoder. That way, the path between two stimuli in the response manifold can be thought as the shortest distance between two responses within that manifold. Results: We applied gRSA to simulations and real data (neural responses from the auditory cortex of the ferret). We successfully reconstructed the stimulus geometry of the simulated data. The analysis led to a satisfactory reconstruction of the stimulus space geometry for the real responses. Conclusion: The perseverance of similarity from the external to the internal space (2nd order isomorphism) is only achieved when the local geometry is taken into account. Our results showed that when this local aspect is not taken into account, the 2nd order isomorphism is sometimes violated and the stimulus space reconstruction can fail.
Faculteit der Sociale Wetenschappen