Piriform model execution on a neurocomputer

The Lynch-Granger model of olfactory/piriform cortex (the piriform model) has been shown to perform a useful signal processing operation known as vector quantization. The piriform model utilizes a layered architecture to achieve this hierarchically, with successive layers (subnets) producing increasingly refined input categorizations. Researchers have developed alternate forms of the basic piriform model and have implemented these on a commercial neurocomputer system. The piriform model was implemented on the connected network of adaptive processors (CNAPS) architecture. This architecture is based on an SIMD model, and it features a linear array of homogeneous processors. The example networks studied were scaled to the dimension of networks that will be required for real-time speech processing applications. The model performance is discussed, CNAPS 256 parallel processors, high local memory bandwidth, and hardware-assisted maximization techniques significantly reduce the execution time required for the piriform algorithm. However, those portions of the piriform algorithm that require serial execution or global communication of locally stored data remain execution bottlenecks. This is reflected in overall network performance