Aging Aware Retraining with a Sparse Update for Neuromorphic Computing

A low-power system with minimal area and high inference accuracy defines the performance of the neural network accelerators. This tradeoff between accuracy and hardware performance is challenging for software hardware co-designers to build an optimum design due to various hardware variabilities. A compressed model can reduce the area and power consumption without compromising accuracy. Pruning is a commonly used method of removing unwanted weight values. After pruning, the distribution of redundant weight values for different neural architectures shows different hardware performances. A highly pruned neural network can be easily affected by hardware variability issues. Analyzing the hardware variability with sparse updates can minimize the area, power, and computational complexity. This co-design approach can decide the level of pruning required for different neural network architectures based on their hardware implementation. Here in this paper, aging-aware retraining with sparse updates on pre-trained neural networks is analyzed. Model compression techniques are analyzed for different neural architectures with hardware variability challenges. Neural network performance analysis and hardware-optimized parameters are compared with other decomposition techniques, such as tensor decomposition and quantization.