Putting AI on Diet: TinyML and Efficient Deep Learning
Song Han · 2021
Summary form only given, as follows. A complete record of the panel discussion was not made available for publication as part of the conference proceedings. Deep leaning algorithms are resource-demanding. This talk will present techniques to reduce the computation recourse, human resource, and data resource for deep learning. First, I’ll present MCUNet, a framework that jointly designs the efficient neural architecture (TinyNAS) and the light-weight inference engine (TinyEngine), enabling ImageNet-scale inference on tiny MCUs that has only 256KB SRAM and 1MB Flash. Next, I’ll present Once-for-All Network a push-the-button solution that can automatically design specialized neural network architectures that best fit many different hardware platforms, which is the winning solution for the 3rd/4th/5th IEEE Low Power Computer Vision Challenge. Finally, I’ll present DiffAugment for data-efficient GAN training, which greatly reduces the training data and can generate high-fidelity images using only training 100 images. I’ll conclude by discussing the applications of these efficient techniques on AIoT, automotive and data center applications.