Thesis Defense Announcement for Amit Das
June 4, 2019
This is Amit Das, I’m going to defend my thesis on 14th June at 3pm. My thesis detail is as follows:
Title: Investigation of the local minima of the energy function of RBM trained by different training sets
When: 14th June, 3pm
Where: Simrall hall, Room 228
Degree: Masters, Electrical and Computer Engineering
Committee:
Dr. James Fowler
(Major Professor)
Dr. Yaroslav Koshka
(Co-advisor)
Dr. Jenny Du
(Committee Member)
Abstract:
An interest in using adiabatic quantum computers for machine learning and neural networks applications is growing. In the previous work, the first commercial quantum annealer (QA), the D-wave, was used to obtain a sample from statistical model distributions, replacing the classical Monte Carlo technique during the training of Energy Based models, such as Restricted Boltzmann Machines (RBMs). It was found that the D-Wave misses many important Local Valleys (LVs) of the RBM energy landscape, which could indicate an insufficiently high quality of the D-Wave sampling. The number of LVs found or missed by the D-Wave was very different for a few different training datasets investigated in the previous work. The goal of this thesis was to understand how the properties of the training datasets influence the energy (i.e., the state probability) landscape (in particular, the number and the “width” of the LVs) formed during the RBM training. Specifically, Monte Carlo technique was applied to analyze the energy landscape statistically to determine how many of the training patterns may end up in the same LV after the training.