Connectionist models as neural abstractions

DIMENSION CHANGE, COARSE GRAINED CODING AND PATTERN RECOGNITION IN SPATIO-TEMPORAL NONLINEAR SYSTEMS

Several research programs employing spatio-temporal recurrent dynamics and changes in dimensionality have extended the dialog on neural computation and coding beyond classical frameworks such as feed forward and attractor neural networks and feature detectors. Some have emphasized spiking networks, while others emphasize oscillations and synchronization as the locus of computation and coding. In this paper, the formalism of locally connected homogeneous coupled map lattices is described. Its deployment in an extended version of the dynamical recognizer framework is described, and is compared with density coding, computational mechanics, and liquid state machine frameworks for neural computation. A population coding strategy based on coarse graining the continuous valued distribution of all sites in the lattice is developed and examined as a form of dimension reduction. Results on recognition of 3-D objects are reported. In order to better understand the dynamics supporting recognition, measures suggested by these other research programs and computational frameworks were examined. Dynamics trajectories from object recognition trials were examined for correlation with recognition rates and measures of the distance of the representation space statistics between the target objects and noise initial conditions, and the intrinsic separation between different objects in the set to be classified were performed. These results raise questions about the efficacy of density coding as an explanation for the results, and on the validity of recent criticisms that chaotic systems cannot satisfy separation requirements required for real time computation.