Spatial/temporal coding for involuntary concept abstraction using linear dynamic systems

The main purpose of biological memory is to robustly recognize objects and concepts, not to reconstruct their copies. For this the memory should mostly record the relations among the constituents of the incoming signals, rather than the constituents themselves. Relational memories can be stored in the form of network connectivity between neurons. However, similarity testing for such spatial representations is difficult. We suggest that the spatial representations are converted to temporal representations for similarity testing, so as to enable a large scale approximate content addressable memory. This not only proposes a new way for automatic concept abstraction in data analysis, but also explains some of the mysteries in the behaviors of human and animal minds.