Timothée LE BERRE

Abstract: This PHD aims to develop a theory of non-uniformly quantized neural networks, for some categories of network: we do not aim at a full generality here, but typically we target Convolutional Neural Nets. This theory should provide hints for choosing, layer per layer, and at a finer granularity, a quantization factor. Based on this theory, an architectural study in terms of resources and performances will be conducted, using state-of-the-art CNNs on meaningful datasets.
The search for appropriate hardware-friendly quantizations will lead the actual design of a family of hardware accelerators, typically targeting FPGAs. Many alternative designs are possible, and quantization will be intermixed with hardware evaluation to build a codesigned hardware aware network and quantization aware network architecture search.