Publication

Abstract : Approximate Computing (AC) trades computational accuracy to optimize performance, power efficiency and area utilization. This was developed based on the observation that many applications such as image processing, and machine learning can tolerate small inaccuracies without significantly affecting the overall quality of results. Convolutional Neural Networks (CNNs) are deep learning models specifically designed for image processing and pattern recognition that outperform traditional machine learning techniques. While AC offers significant advantages, error unpredictability and its propagation has been a huge drawback. This work investigates error-tolerant methods in approximate computing to increase the efficiency and robustness of CNN to be used in pattern classification. Using stochastic computing, arithmetic blocks such as Stochastic Number Generators (SNGs), adders and multipliers are implemented. Triple Modular Redundancy (TMR) is employed with majority voting to enhance error tolerance in hardware. These arithmetic units are embedded within CNN layers and they provide acceptable model accuracy. Comparative study indicates enhanced error tolerance over conventional techniques, and reports 14% error rate reduction.