Real-time hybrid learning and recognition system with software-hardware cooperation

The reported hybrid algorithm with software-hardware cooperation for prototype-based learning uses dynamic K-means to intelligently produce new cluster centers as reference patterns. The single nearest neighbor matching (K-nearest neighbor (KNN), K=3D1) with full-hardware architecture is used for efficient recognition. Large numbers of training instances are dynamically partitioned into k clusters by K-means according to the desired accuracy rate for recognizing the original training instances using the generated cluster centers. A major deficiency in conventional implementations for both K-means and KNN is the processing speed of the nearest neighbor matching. This deficiency is resolved by an FPGA-implemented associative-memory for parallel nearest neighbor matching. The software part of the learning implementation is used to update the cluster centers and communicate with the hardware. We benchmarked our hybrid learning algorithm with the popular case of handwritten digit recognition where abundant previous work is available. In case of the MNIST database we obtained the most efficient reported result of 96.15% accuracy rate using only 128 cluster centers. A single nearest neighbor recognition with the developed hardware architecture among these 128 reference data needs only 10.1 μs.