Optimizing embedded neural networks based on saturated states

With the expansion of embedded systems and the new applications of these computing systems in Internet of Things, smart homes and sensors, such systems are faced with a large amount of data. Therefore, designing an embedded system for massive sensor data analysis for detecting a situation and making a decision is becoming more complex. Machine learning algorithms are a promising method for designing such systems. Neural networks are very popular learning systems to be used in data analysis. However, huge computations of the neural networks make them unsuitable for embedded systems which have strict limitations on energy consumption and execution time as the main design concerns. In this paper we propose an optimization for multi-layer perceptron neural networks based on neuron saturation states. Experimental results show that in average 89% of the neurons are in saturated state for selected benchmark applications. Since in a saturated neuron, exact numerical values are not required, predicting the saturation state helps to perform the computation with less effort. The goal of the saturation prediction is to keep accuracy while minimizing the computations.