Associative-memory-based prototype LSI with recognition and on-line learning capability and its application to handwritten characters

In the presented research on VLSI-system design for handwritten-character learning and recognition, an associative memory architecture for searching the most similar data among previously stored reference data is applied. The chosen associative memory achieves high speed, low power consumption and small implementation area due to a mixed digital-analog fully-parallel nearest-match search circuitry. To recognize new data, a learning capability based on the concept of short/long-term memory which tries to mimic the function of the human brain is realized. For improvement of the recognition rate, we propose a reference-data-optimization algorithm that averages the recognized input patterns for each stored reference object and updates the corresponding reference pattern periodically. We evaluated the proposed intelligent VLSI-design method for the application of hand-written character learning and recognition. Therefore, test-chip in 0.18 um CMOS technology was designed to demonstrate the proposed algorithm and design method. The simulated processing capability of this test-chip amounts to 0.3 million input-character images per second.