Abstract
The computational workload involved in CNNs is typically out of reach for low-power embedded devices. The Approximate Computing paradigm can be exploited to reduce the CNN complexity since it improves performances and energy-efficiency by relaxing the need for fully accurate operations. In this work, we target weight-sharing as an approximate technique to reduce the memory footprint of a CNN. More in detail, we prove that optimizing the number of shared weights can enable significant network memory compression without noticeable accuracy loss without retraining or fine-tuning steps. However, we observe that the exploration time can easily explode in state-of-the-art CNNs. We thus propose the use of a fast accuracy estimation metric to guide the design space exploration and drastically reduce the exploration time up to 12×. Compared with state-of-the-art CNN approximation methods, we obtained more than 4× compression on GoogleNet on the ImageNet dataset with less than 1% accuracy loss in less than 5 h and without any retraining step.