Skip to Main Content
Powerful parallel cognitive processors can be developed by studying biologically plausible models of cognitive systems in animals and extrapolating key principles to be adapted for implementation in digital computer architectures. The network described here uses basic statistical methods such as proportion sampling on a massively parallel scale to create a general purpose pattern classifier. From these principles, we can achieve auto association and self organization that provides fundamental cognitive processing. Signal preprocessing is essential to transform the signal into a scale and rotation invariant binary pattern. The network avoids the curse of dimensionality by filtering out irrelevant inputs, allowing us to combine large sensor input vectors from multiple sources. Recent hardware designs define the network structure and state in memory, and then use accelerator processor cores to modify these memory structures in parallel.