10-11 am EST, Monday, 18 May 2020


Edgification: Deep Learning Model Inference On The Edge


Increasing popularity of convolution neural nets and its ability to perform complex computations on low power/cost CPU’s has made its adoption ubiquitous. Traditionally, Deep Learning (DL) frameworks like Tensorflow, MxNet, PyTorch were used for both training and inferencing. These frameworks provided higher level abstraction for the developers to create and debug sophisticated architecture. However, for the model inference, these libraries were in fact sub optimal which led to creation of other frameworks such as OpenVINO and TVM. OpenVINO and TVM provided frameworks specifically for model inference, utilizing both hardware acceleration via kernel intrinsic and graph optimization via operator fusing. In this talk, we shall focus on edge computing, use-cases that mandates high computation with low latency, need for DL learning models to be run on the edge and finally approaches for fine-tuning these models, a term that is coined as “edgification”.


Goutham Kamath is currently Head of AI and ML at Foghorn Systems. His work focuses on developing products that work in the intersection between AI/ML, Edge/IoT. His work has led to the development of EdgeML, a first industrial ML platform for edge computing. Optimizing, Shrinking, Quantization of deep learning models are other areas of interest. Goutham Kamath received his M.S. in Electrical Engineering from University of Wyoming in 2012 and Ph.D. in Computer Science from Georgia State University in 2016.