Cognimem Technologies, previously known as Recognetics, is now delivering their Cognimem PM1K neural network chip in the CogniMem PM1K module (Fig. 1) and the V1KU camera module (Fig. 2). These are based on the PM1K chip (Fig. 3) that implements a 3-layer, 1024 neuron system that links the K nearest vectors in priority order. This provides a high-speed, non-linear classifier. The system can be expanded using multiple chips that operate in parallel. Trained system information can be easily downloaded and uploaded making cloning easy.

Cognimem Technologies was highlighting the V1KU camera module connected a Freescale microcontroller at the Freescale Technology Forum. It allowed a microcontroller to do almost instantaneous image recognition. The PM1K chip implements fully parallel pattern recognition including K-Nearest Neighbor (KNN) and Radial Basis Function (RBF). KNN is useful for finding the closest match for fingerprints. RBF addresses nonlinear mapping and classification chores that can reduce the set of results. Learning mode classification speed and recognition is only 10 µsec. Classification can be controlled during learning to minimize false positives.

The PM1K can be used with any data source not just image information. It uses a 256 byte input vector.

CogniMem PM1K module is priced at $190. It uses an I2 interface with only a dozen registers. There is an 8-bit digital input bus compatatible with some cameras. Data can also be supplied via the serial interfac.e

The V1KU incorporates a Micron/Aptina monochrome CMOS sensor that operates at 60 frames/s. In addition to the CogniMem CM1K chip, the module has a 600K gate Actel IGLOO FPGA. This contains the standard CogniSight Engine for region monitoring. The module also has 16 Mbytes of SRAM and 4 Mbytes of flash memory. There is a high-speed USB interface and a 921.6 Kbaud RS485 interface.