Video Recordings

Neural networks are notorious for their high computational intensity and energy usage. However, recent advances in the field have made it possible to reduce their computational load to the point where machine translation systems can be run on a mobile phone.

In this talk, we will a take kaleidoscopic view of neural network optimisation, focusing on neural machine translation as a case study. We will cover model improvements, neural machine translation specific improvements and software improvements both on the GPU and the CPU. Combining all improvements we manage to decrease inference time by a factor of ~600 with a tiny drop in BLEU.