Porting state of the art deep learning algorithms to resource constrainedcompute platforms (e.g. VR, AR, wearables) is extremely challenging. We proposea fast, compact, and accurate model for convolutional neural networks thatenables efficient learning and inference. We introduce LCNN, a lookup-basedconvolutional neural network that encodes convolutions by few lookups to adictionary that is trained to cover the space of weights in CNNs. Training LCNNinvolves jointly learning a dictionary and a small set of linear combinations.The size of the dictionary naturally traces a spectrum of trade-offs betweenefficiency and accuracy. Our experimental results on ImageNet challenge showthat LCNN can offer 3.2x speedup while achieving 55.1% top-1 accuracy usingAlexNet architecture. Our fastest LCNN offers 37.6x speed up over AlexNet whilemaintaining 44.3% top-1 accuracy. LCNN not only offers dramatic speed ups atinference, but it also enables efficient training. In this paper, we show thebenefits of LCNN in few-shot learning and few-iteration learning, two crucialaspects of on-device training of deep learning models.
展开▼
