The response fields of higher cortical neurons are usually approximated with smooth mathematical functions for the purpose of population parameterization or theoretical modeling. We used instead two nonparametric methods (principal component analysis and independent component analysis), which provided a basis for the response field clustering. Although both methods performed satisfactorily, the principal component analysis space is more straightforward to calculate. It also gave a clear preference toward the smallest number of functional response field classes. Clustering was performed with both K-means and superparamagnetic clustering algorithms with similar results. We also show that the shapes of the eigenvectors remain consistent regardless of the response field data sets size. This finding reflects the fact that the response fields were generated by the same neural network and encode the same underlying process.
Division of Biology, California Institute of Technology, Pasadena, California, USA
Correspondence and request for reprints to Dr Marina Brozović, California Institute of Technology, Division of Biology, 1200 E. California Blvd., Mail Code 216-76, Pasadena, CA 91125, USA
Tel: +1 626 395 8337; fax: +1 626 795 2397; e-mail: firstname.lastname@example.org
Sponsorship: This work was supported by the James G. Boswell Foundation, the Sloan-Swartz Center for Theoretical Neurobiology and the National Eye Institute.
Received 30 March 2006; accepted 3 April 2006