Terrence J. Sejnowski, Kechen Zhang, Iris Ginzburg, and Bruce McNaughton
Salk Institute and
University of Arizona, Tucson
Reconstruction or decoding is an inverse problem in which physical variables are estimated from observed neural activity. The accuracy of two methods for reconstruction, probabilistic and basis function methods are compared on the reconstruction of the full trajectory of freely moving rats using spike timing data of simultaneously recorded hippocampal place cells. The theoretical limits on reconstruction accuracy estimated from Fisher information are compared with the actual performances of the reconstruction methods. The lowest reconstruction errors for Bayesian methods were within a factor of two of the information-theoretic limit, comparable with the intrinsic experimental errors in tracking. Both reconstruction methods can be implemented by feedforward neural networks.