Object Recognition Using Structured Feature Extraction With A Reconfigurable, Neurosynaptic Processor

Restricted (Penn State Only)
Author:
Gomatam, Priyanka
Graduate Program:
Computer Science and Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
April 05, 2017
Committee Members:
  • Vijaykrishnan Narayanan, Thesis Advisor
  • Jack Sampson, Committee Member
Keywords:
  • Truenorth
  • neuromorphic hardware
  • object detection
  • histogram of oriented gradients
  • neural network hardware
Abstract:
Neuromorphic hardware has culminated increased interest, with a focus on designing efficient platforms that can support neural network tasks. Many algorithms used in applications today extensively perform pre-processing of data, in which structured features of the data are extracted and used for classification. The idea is to explore the ability to implement a structured computation in a neuromorphic platform. This involves leveraging the operations that are best suited for neuromorphic hardware, and using them to achieve the same results as a traditional algorithm. In this paper, a case study of mapping the feature extraction stage of pedestrian detection using Histogram of Oriented Gradients (HoG) onto a neuromorphic platform is performed. Further, this neuromorphic feature extractor is then connected to a neural network based classifier. The performance of the feature extraction done by a 1:1 mapping of the algorithm is evaluated against other neuromorphic implementations, as well as an FPGA implementation. The neuromorphic platform chosen for this experiment is IBM’s TrueNorth, a Neurosynaptic System.