Start Time: 1:30 pm
End Time: 2:00 pm
This talk explores the challenges of deploying serial computer vision tasks implemented with DNNs. Neural network accelerators have demonstrated significant gains in performance for DNN inference, especially when the net has been quantized. Quantization often brings a loss in accuracy. This loss in accuracy may be considered acceptable in itself but may cause problems if the output of the DNN is used as input for a second DNN which itself has been quantized. We present our research into this challenge in the context of a face verification CNN which consumes the output of a face detection CNN, and present approaches for reducing the impact of quantization in such scenarios.