Please use this identifier to cite or link to this item:
Type: Thesis
Title: Decision making with reciprocal chains and binary neural network models
Author: Stamatescu, George
Issue Date: 2020
School/Discipline: School of Electrical & Electronic Engineering
Abstract: Automated decision making systems are relied on in increasingly diverse and critical settings. Human users expect such systems to improve or augment their own decision making in complex scenarios, in real time, often across distributed networks of devices. This thesis studies binary decision making systems of two forms. The rst system is built from a reciprocal chain, a statistical model able to capture the intentional behaviour of targets moving through a statespace, such as moving towards a destination state. The rst part of the thesis questions the utility of this higher level information in a tracking problem where the system must decide whether a target exists or not. The contributions of this study characterise the bene ts to be expected from reciprocal chains for tracking, using statistical tools and a novel simulation environment that provides relevant numerical experiments. Real world decision making systems often combine statistical models, such as the reciprocal chain, with the second type of system studied in this thesis, a neural network. In the tracking context, a neural network typically forms the object detection system. However, the power consumption and memory usage of state of the art neural networks makes their use on small devices infeasible. This motivates the study of binary neural networks in the second part of the thesis. Such networks use less memory and are e cient to run, compared to standard full precision networks. However, their optimisation is di cult, due to the non-di erentiable functions involved. Several algorithms elect to optimise surrogate networks that are di erentiable and correspond in some way to the original binary network. Unfortunately, the many choices involved in the algorithm design are poorly understood. The second part of the thesis questions the role of parameter initialisation in the optimisation of binary neural networks. Borrowing analytic tools from statistical physics, it is possible to characterise the typical behaviour of a range of algorithms at initialisation precisely, by studying how input signals propagate through these networks on average. This theoretical development also yields practical outcomes, providing scales that limit network depth and suggesting new initialisation methods for binary neural networks.
Advisor: Lang, To
Dissertation Note: Thesis (Ph.D.) -- University of Adelaide, School of Electrical & Electronic Engineering, 2020
Provenance: This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at:
Appears in Collections:Research Theses

Files in This Item:
File Description SizeFormat 
Stamatescu2020_PhD.pdf1.52 MBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.