Quantization-Aware In-situ Training for Reliable and Accurate Edge AI
João Pauloa and Luigi Carrob
Informatics Institute, Federal University of Rio Grande do Sul, Porto Alegre, Brazil
ajpclima@inf.ufrgs.br
bcarro@inf.ufrgs.br
ABSTRACT
In-memory analog computation based on memristor crossbars has become the most promising approach for DNN inference. Because compute and memory requirements are larger during training, memristive crossbars are also an alternative to train DNN models within a feasible energy budget for edge devices, especially in the light of trends towards security, privacy, latency, and energy reduction, by avoiding data transfer over the Internet. To enable online training and inference on the same device, however, there are still challenges related to different minimum bitwidth needed in each phase, and memristor non-idealities to be addressed. We provide an in-situ training framework that allows the network to adapt to hardware imperfections, while practically eliminating errors from weight quantization. We validate our methodology with image classifiers, namely MNIST and CIFAR10, by training NN models with 8-bit weights and quantizing to 2 bits. The training algorithm recovers up to 12% of the accuracy lost to quantization errors even under high variability, reduces training energy by up to 6×, and allows for energy-efficient inferences using a single cell per synapse, hence enhancing robustness and accuracy for a smooth training-to-inference transition.
Keywords: Synaptic Devices, Deep Neural Networks, In-Situ Training, Inference, Process Variation.
