A Temporal Sequence Processor Based on the Biological Reaction-diffusion Process

Temporal sequences are a fundamental form of information and communication both in natural and engineered systems. The biological control process which directs the generation of iterative structures from undifferentiated tissue is a type of temporal sequential process. A quantitative explanation of this temporal process is reaction-diffusion, initially proposed by Turing [1952] and later widely studied and elaborated. We have adapted the reaction-diffusion mechanism to create a novel network and algorithm based on a chemical ``neuron'''' model, which performs storage, associative retrieval and prediction for temporal sequences. Experiments demonstrate the ability of the device to achieve any desired depth limited only by storage capacity, to remember and predict on the basis of count to any length, and to learn and recognize embedded Reber grammar strings to 98% accuracy. The network is also capable of preserving time extent of stored symbols, such as in a musical melody, and permitting retrieval of both the symbols and their temporal extent. Redundancy can be controlled during retrieval or recognition permitting flexible grouping of stored sequences into broader classes during readout.