Ternary Neural Networks for Gait Identification in Wearable Devices

Recently, wearable devices such as smartwatches and smart glasses have considerably risen in popularity. Typically, they are equipped with sensors that can provide a huge flow of potentially useful data for different applications. However, the available computational resources and power supply are rather limited, so complex classification and transmission architectures are not feasible. For these reasons, a growing research interest has been focused on strategies that enhance the deployability of Deep Neural Networks (DNNs) while preserving performance on a given task by reducing the number of network weights and quantizing them. In this paper, we employ Ternary Neural Networks (TNNs), a method combining quantization and parameter pruning. Specifically, we investigate the effectiveness of TNNs for gait biometric identification from wearable sensors, whose application in gait identification systems is largely unexplored. Ternary quantization sets a high number of network parameters falling into an interval to zero, effectively removing them from the network topology, hence enabling great memory and computational efficiency. We provide a lightweight deep classifier model with ternarized weights and activations and train it end-to-end to achieve competitive performance with the state-of-the-art while ensuring remarkably high sparsity rate, at times even greater than 90% (i.e., less than 10% of the parameters remain in the topology). Furthermore, the obtained ternary parameter distribution reaches entropy rates that are significantly lower than 1 bit, allowing further compressibility compared to plain binary neural networks which we also considerably outperform.