Architectures and algorithms for on-device user customization of CNNs

Abstract

A convolutional neural network (CNN) architecture supporting on-device user customization is proposed. The network architecture consists of a large CNN trained on a general data and a smaller augmenting network that can be re-trained on-device using a small user-specific data provided by the user. The proposed approach is applied to handwritten character recognition of the Latin and the Korean alphabet, Hangul. Experiments show a 3.5-fold reduction of the prediction error after user customization for both the Latin and the Korean character set compared to the CNN trained with general data. To minimize the energy required when retraining on-device, the use of a coarse-grained reconfigurable array processor (CGRA) in a low-power, efficient manner is presented. The CGRA achieves a speedup of 36× and a 54-fold reduced energy consumption compared to an ARMv8 processor. Compared to a 3-way VLIW processor, a speedup of 42× and a 12-fold energy reduction is observed, demonstrating the potential of general-purpose CGRAs as light-weight DNN accelerators.