Associative Processing: A Paradigm for Massively Parallel AI

In associative memory, recall is based on similarity to a cue. With its inherent data parallelism, associative memory naturally lends itself to implementation on massively parallel hardware; it is our thesis that associative processing can serve as the basis of AI systems. We believe that the associative paradigm can encompass both neural network and symbolic applications. Current research indicates that architectural innovations will permit massively parallel processing beyond brute force search and parallelizing vector-matrix multiplication neural network models. We are developing a hardware architecture, the MISC Machine (Misc Is Symbolic Computing), to support associative processing for AI applications. It is a massively parallel hierarchical system combining multiple SIMD (Single Instruction stream, Multiple Data streams) arrays with a MIMD (Multiple Instruction streams, Multiple Data streams) processor network. In contrast to PlZOLOG and LISP engines, MISC is based on application rather than programming language requirements, and emphasizes association as the basis for intelligent systems. Our three target application areas are neural networks, marker-passing semantic networks, and conceptual graphs. Our approach analyses these applications for fundamental high-level associative operations. These high-level operations are then translated into machine primitives, emphasizing memory content-addressing via the SIMD arrays. Software development will be based on object and collection oriented languages such as SETL [Sch92, SDDS86]. By working directly from AI application requirements, the system will be tailored to symbolic codes, with simplified programming and a significant execution time speedup.