turkmath.org

Deep neural networks have achieved state-of-the-art results on a wide range of artificial intelligence tasks. However, the computation and storage requirements are usually quite high. This issue limits their deployability on ubiquitous computing devices such as smart phones, wearables and autonomous drones. In this talk, I will present ternary neural networks (TNNs) in order to make deep learning more resource-efficient. A TNN is a discretised neural network where the weights and activations are constraint to three values only (-1,0, and 1). Due to this extreme limitation, there is no standard training procedure for TNNs. I will introduce a novel teacher-student approach for training TNNs without compromising too much accuracy. Next, I will describe our purpose-built hardware architecture for TNNs and present benchmark results that demonstrate up to 3x better energy efficiency than the existing solutions.