Working with CuDNN and Deep learning library team.

My research focus has been on "Addressing the algorithmic gap on low precision hardware susbtrates".- Limited precision matrix inverse: Developed algorithm and mathematical framework for computing limited-precision matrix inverse. Goal was to deploy such algorithm to energy-efficient stochastic computing hardware substrates such as FPGAs and IBM TrueNorth. The proposed work was tested for applications related to machine vision, robotics and pattern recognition using low-power architectures.- Learning affine transforms as a separate layer of abstraction: Proposed an on-line learning framework for Map-Seeking Circuits algorithm. A biologically inspired neural circuit based algorithm that can be used for object tracking and inverse tracking. My goal was to propose an algorithm where the affine transformation matrices can be learned as a separate layer of abstraction for object recognition and tracking.

➢ Extended the prior work on object detection to overhead aerial images of cars and implemented the proposed CNN using IBM’s EEDN framework.➢ The CNN model was redesigned based on the concepts from SqueezeNet, TinyYOLO and TinySSD.➢ This implementation was compared against standard neural network structures that were implemented on NVIDIA Titan X (Maxwell) GPU using Caffe deep learning framework.- For my internship at Lawrence Livermore National Laboratory, we performed analysis of area surveillance based application using a low-power IBM TrueNorth Neurosynaptic System. For our evaluation we looked at a publicly available dataset that has overhead imagery of cars with context present in the image. The trained neural network for image analysis was deployed on the NS16e system using IBM's EEDN training framework. The results have been compared with caffe-based implementations of standard neural networks that were deployed on a Titan-X GPU. Results showed that TrueNorth can achieve better accuracy when comparable to high-precision neural networks like AlexNet, and have comparable accuracies to GoogLeNet and ResCeption, at the same time show a manifold improvement in power consumption.Mentors: Brian Van Essen, Adam Moody and Naoya Maruyama.

- Proposed a “You Only Look Once” style architecture for convolutional neural network architecture that can do object detection and classification on NS16e. The proposed CNN model was modified for ternary weight and binary activation based neural network. - The proposed algorithm was inspired from You Only Look Once and Single Shot Detector style CNNs and was designed from scratch.- Used Keras and Tensorflow for analyzing the different convolutional neural network structures using different bit precisions.Mentors: Brian Van Essen, Adam Moody and David Widemann.

Implemented Smith-Waterman algorithm for the sequencing hardware accelerator on Kintex-7 FPGA to improve the local alignment of two gene sequences in terms of speed. This accelerator can perform local alignment of gene sequences that are millions of base pair long in the order of minutes.Tasks involved- Developing a datapath for parallel computation of Smith-Waterman matrix scores. These computations were done in parallel using systolic array architecture.- Writing a multi-threaded C++ program to send multiple gene sequences and get the matrix scores for these multiple sequences.Development and verification were done using Xilinx ISE 14.7 on Kintex-7 FPGA.

➢ I was involved with the setup of EEG based control system for virtual reality space. The EEG data was collected using Emotiv headset.➢ Also contributed towards computational neuroscience based experiments that were aimed at studying effects of subliminal priming