Configurable Accelerators for Visual and Text Analytics
Open Access
- Author:
- Park, Mi Sun
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- September 19, 2014
- Committee Members:
- Vijaykrishnan Narayanan, Dissertation Advisor/Co-Advisor
Mary Jane Irwin, Dissertation Advisor/Co-Advisor
C Lee Giles, Committee Member
Bradley Paul Wyble, Committee Member
Omesh Tickoo, Special Member - Keywords:
- Accelerators
Recognition
Vision
Text Processing
Personal Analytics - Abstract:
- Continuously recording a person's daily life and creating a digital backup of it is no longer just science fiction. Current lifelogging technologies allow us to automatically capture every single minute of a person's experience from wearable sensors. However, how to efficiently manage the captured personal multimedia data and provide user services using the data still remains a great challenge. This dissertation presents a proposal to leverage hardware acceleration for a personal analytics system that can continuously capture a person's daily life and respond user queries using the captured lifelog. First, two visual accelerators for biologically-inspired recognition are introduced: HMAX accelerator for object recognition and gist accelerator for scene recognition. This includes a description of the novel architecture, and an analysis of the experimental results of each accelerator by comparing its performance to contemporary solutions on different platforms. Second, a prototype system of personal analytics, which consists of front-end recognition and back-end text analytics, is presented to realize the vision of creating a lifelog system that enables total recall through total capture of a person's everyday experience. Third, a text accelerator for Naive Bayes multiclass classifier is presented to support the real-time text processing of the back-end analytics in personal analytics system. By leveraging these visual and text accelerators, the front-end system is able to extract essential information from multimedia data and transmit the lifelog data into the back-end system in real-time. Therefore, the analytics can perform content-aware lifelog analysis and support efficient query processing using the lifelog data.