Date of Award

12-31-2016

Document Type

Campus Access Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Marc Pomplun

Second Advisor

Dan A. Simovici

Third Advisor

Craig Yu

Abstract

Roughly, 50% of the human brain is devoted to visual processing. The remarkable abilities of our visual system to process and recognize complex scenes within a fraction of a second speak to the computational magnitude of this feat. The central theme of this dissertation is the way our brain achieves computational efficiency given its limited resources. Specifically, I investigated existing or learnt biases in attention and memory mechanisms in the visual system that aide in judicious budgeting of processing resources.

Attention is the mechanism by which visual input is filtered so we can rapidly compute important properties of the environment. Attention can act at the moment of input such as when light falls on the retina (in a ‘bottom-up’ fashion) or it can act in a ‘top-down’ manner, such as using previous knowledge to search for a target in a scene. The first set of experiments in this dissertation investigated bottom-up biases, specifically whether oculomotor biases affect properties of spatial attention. Prior to these experiments, it was not known whether changes to parameters of the oculomotor system affected distribution of spatial attention. These experiments demonstrate that changes occurring in one system perturb the other in a similar fashion, implying that topographic maps are possibly shared by these processes.

Next, I investigated how memory might be affected by top-down biases during object recognition. How are individual objects represented and encoded in visual working memory? Do features of individual objects or {\em whole} objects or relationships between these objects affect efficiency of encoding in memory? Results from these experiments show that core object representations are episodically organized in memory according to scene context and that facilitates rapid recognition.

Finally, I studied how top-down and bottom-up biases in attention interact during dynamic scene viewing with multiple moving objects. Specifically, what is the limit of our abilities to analyze interactions among multiple moving objects? Lastly, taking a connectionist perspective, I proposed a computational model that fills in bottom-up feature information from impoverished visual input mimicking filling-in mechanisms of our visual system at the earliest stages of visual processing.

Together, these experiments broaden understanding of the consequence of attentional biases on visual processing. The main findings are: (1) changes to the oculomotor system could bias spatial attention, (2) semantic relationships between objects in a set are inferred by a subconscious automatic mechanism that is impeded by conscious processing, (3) we can monitor up to three events and/or moving objects in a dynamic scene and (4) non-linear computations at the earliest stages of visual processing are capable of filling in low level feature information from impoverished visual input.

Comments

Free and open access to this Campus Access Dissertation is made available to the UMass Boston community by ScholarWorks at UMass Boston. Those not on campus and those without a UMass Boston campus username and password may gain access to this dissertation through resources like Proquest Dissertations & Theses Global or through Interlibrary Loan. If you have a UMass Boston campus username and password and would like to download this work from off-campus, click on the "Off-Campus UMass Boston Users" link above.

Share

COinS