Senior Technical Manager, Vision Systems Laboratory, Center for Vision Technologies
David Zhang, Ph.D., is the Senior Technical Manager of the Edge Computing Group at SRI International. His group studies AI online learning and adaptation at the edge, focusing on the algorithm-hardware codesign of neuromorphic architecture for low power AI devices. The group develops disruptive solutions in real-time multi-sensor fusion and smart visions.
David has expertise in computer vision, machine learning and edge computing, with experience in computational sensing and real-time video processing, and image enhancements in degraded visual environment. His recent research is on Hyperdimensional Computing, gradient-free learning, federated learning at the edge and efficient big data ML processing. His work on quantized neural network has supported a spin-off company which won the 2020 Startup of the Year Award by IoT World. He was a PI or CO-PI on numerous DARPA/DoD programs, including DARPA Prowess, DARPA IP2, IARPA MicroE4AI and ARPA-E MicroCam. He is also leading the SRI IRAD NeuroEdge program. David received his PhD in Physics from Penn State in 2001. He has had over 30 published papers, more than 10 patents and numerous patent applications.
We present a scalable in-pixel processing architecture that can reduce the data throughput by 10X and consume less than 30 mW per megapixel at the imager frontend.
To enable learning on edge devices with fast convergence and low memory, we present a novel backpropagation-free optimization algorithm dubbed Target Projection Stochastic Gradient Descent (tpSGD).
We examine how the saccade mechanism from biological vision can be used to make deep neural networks more efficient for classification and object detection problems.
In this paper we present Hyper-Dimensional Reconfigurable Analytics at the Tactical Edge using low-SWaP embedded hardware that can perform real-time reconfiguration at the edge leveraging non-MAC deep neural nets (DNN)…
We review HyDRATE, a low-SWaP reconfigurable neural network architecture developed under the DARPA AIE HyDDENN (Hyper-Dimensional Data Enabled Neural Network) program.
In this paper, we present a comparison of model-parameter driven quantization approaches that can achieve as low as 3-bit precision without affecting accuracy.