CNN inference acceleration on limited resources FPGA platforms_epilepsy detection case study

Abstract

The  use  of  a  convolutional  neural  network  (CNN)  to  analyze  and  classify electroencephalogram  (EEG)  signals  has  recently  attracted  the  interest  of researchers  to  identify  epileptic  seizures.  This  success  has  come  with  an enormous    increase    in    the    computational    complexity    and    memory requirements  of  CNNs.  For  the  sake  of  boosting  the  performance  of  CNN inference,  several  hardware  accelerators  have  been  proposed.  The  high performance  and  flexibility  of  the  field  programmable  gate  array  (FPGA) make it an efficient accelerator for CNNs. Nevertheless, for resource-limited platforms, the deployment of CNN models poses significant challenges. For an  ease  of  CNN  implementation  on  such  platforms,  several  tools  and frameworks  have  been  made  available  by  the  research  community  along with different optimization techniques. In this paper, we proposed an FPGA implementation for an automatic seizure detection approach using two CNN models,  namely  VGG-16  and  ResNet-50.  To  reduce  the  model  size  and computation  cost,  we  exploitedtwo  optimization  approaches:  pruning  and quantization.   Furthermore,   we   presented   the   results   and   discussed   the advantages   and   limitations   of   two   implementation   alternatives   for   the inference acceleration of quantized CNNs on Zynq-7000: an advanced RISC machine   (ARM)software   implementation-based   ARM,NN,software development kit (SDK) and a software/hardware implementation-based deep learning processor unit (DPU) accelerator and DNNDK toolkit.