Attention-based machine perception for intelligent cyber-physical systems

Liu, Shengzhong

Permalink

https://hdl.handle.net/2142/113883 Copy

Description

Title

Attention-based machine perception for intelligent cyber-physical systems

Author(s)

Liu, Shengzhong

Issue Date

2021-12-02

Director of Research (if dissertation) or Advisor (if thesis)

Abdelzaher, Tarek

Doctoral Committee Chair(s)

Abdelzaher, Tarek

Committee Member(s)

• Sha, Lui
• Gupta, Indranil
• Le, Franck

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Cyber-Physical System
• Machine Perception
• Attention

Abstract

Cyber-physical systems (CPS) fundamentally change the way of how information systems interact with the physical world. They integrate the sensing, computing, and communication capabilities on heterogeneous platforms and infrastructures. Efficient and effective perception of the environment lays the foundation of proper operations in other CPS components (e.g., planning and control). Recent advances in artificial intelligence (AI) have unprecedentedly changed the way of how cyber systems extract knowledge from the collected sensing data, and understand the physical surroundings. This novel data-to-knowledge transformation capability pushes a wide spectrum of recognition tasks (e.g., visual object detection, speech recognition, and sensor-based human activity recognition) to a higher level, and opens an new era of intelligent cyber-physical systems. However, the state-of-the-art neural perception models are typically computation-intensive and sensitive to data noises, which induce significant challenges when they are deployed on resources-limited embedded platforms. This dissertation works on optimizing both the efficiency and efficacy of deep-neural- network (DNN)-based machine perception in intelligent cyber-physical systems. We extensively exploit and apply the design philosophy of attention, originated from cognitive psychology field, from multiple perspectives of machine perception. It generally means al- locating different degrees of concentration to different perceived stimuli. Specifically, we address the following five research questions: First, can we run the computation-intensive neural perception models in real-time by only looking at (i.e., scheduling) the important parts of the perceived scenes, with the cueing from an external sensor? Second, can we eliminate the dependency on the external cueing and make the scheduling framework a self- cueing system? Third, how to distribute the workloads among cameras in a distributed (visual) perception system, where multiple cameras can observe the same parts of the environment? Fourth, how to optimize the achieved perception quality when sensing data from heterogeneous locations and sensor types are collected and utilized? Fifth, how to handle sensor failures in a distributed sensing system, when the deployed neural perception models are sensitive to missing data? We formulate the above problems, and introduce corresponding attention-based solutions for each, to construct the fundamental building blocks for envisioning an attention-based machine perception system in intelligent CPS with both efficiency and efficacy guarantees.

Graduation Semester

2021-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/113883

Copyright and License Information

Copyright 2021 Shengzhong Liu

Owning Collections