Designed toolsets to model deep learning workload performance on GPU.Created API for workload performance tools, bridged infrastructure to internal teams.

Acceleration of embedded machine learning algorithms on FPGA.Designed a demo visualization of Xilinx ML acceleration suite / intermediate convolutional layers.

Exploring Recurrent Neural Networks (RNNs) to Recognize Human Gestures in Edge Computing EnvironmentsThe goal of the project is to explore identifying a set of general human gestures and activities in edge computing scenarios. Edge computing is any computer environment in which data processing occurs at the edge of the network generally characterized with limited connectivity to high-performance centralized or cloud systems. The research project deploys an edge camera system with BodySLAM (Body Simultaneous Localization and Mapping), a deep neural network framework comprised of both convolutional neural networks (CNNs) and fully connected networks (FCNs) trained to detect human position of body joints segments (arms, legs, head,etc.) BodySLAM generates in real-time the human joint positions as a set of 3D (x,y,z) points over short time durations. The primary project work investigates the training and building of a new recurrent neural network (RNN) to characterize simple human movements (raising a hand, reaching out, waving, or crouching) from the BodySLAM pose results. An RNN is a class of artificial neural network that can capture connections between a sequence of data to detect temporal behavior over a time sequence. The platform used is an NVIDIA TX2 with on-board graphics processing unit (GPU) that can execute 2 TFLOPS (trillion floating point operations per second) and relay simple event results of a scene. The resulting edge computing system impacts many new automated opportunities to assess worker efficiency, detect unsafe scenarios, and judge athletic performance.

I took part in a National Science Foundation undergraduate research experience at Wright State University in Ohio. The program lasted 11 weeks, and covered the topic of ‘Intelligent Autonomous Vehicles.’ Our team researched the Localization and Mapping aspect of an Autonomous Drone. My work covered the data acquisition and processing of data to map the environment. We took 2-D Lidar data from a sensor onboard, and ran it through a Rao-Blackwellized particle filter. We localized off of this data, and combined it with other onboard sensors to localize accurately. These estimates were averaged with an Extended Kalman Filter to receive a more accurate estimate of the drone’s position. This data was handed to the path planning team in the form of an occupancy grid to plan a course. We created a poster, “Realtime Quadrotor Simultaneous Localization and Mapping (SLAM) using Rao-Blackwellized Particle Filtering,” and presented our cumulative results at the end of the project.