Towards Transparent Machine Perception Systems

Abstract

The past two decades have witnessed significant progress in machine perception Ð today, there are commercial systems for face detection (Face.com), speech recognition (Siri), handwriting recognition (Microsoft OneNote), product search (Google Goggles), and pedestrian detection (Mobileye). Unfortunately, when todayÕs machine perception systems fail, they fail in a spectacularly disgraceful manner, without warning or explanation, leaving the user staring at an incoherent output, wondering why the system did what it did. The root cause is lack of transparency. With a few rare exceptions, the emphasis in machine learning and computer vision communities today is on building systems with good predictive performance, not interpretability or transparency. As a result, the users of these vision systems perceive them as inscrutable black boxes that cannot be understood or trusted. In order to build operator and institutional trust in vision systems, to identify bottlenecks in existing systems, to identify the most fruitful research directions, and to ultimately move towards meaningful integration of perception systems into our everyday lives, we must address the fundamental and difficult question of Òwhy does a learning-based perception system do what it does?Ó We propose to develop theory, algorithms, and implementations for transparent vision systems that are able to concisely explain what they believe about the world around them and why. Perception systems that are able to explain their beliefs in terms of hypotheses and provide support for their predictions have the potential to fundamentally change the way we live. The main barrier to certifiably safe integration of autonomous systems in civil society is lack of transparency and trust.

Document Details

Document Type
DoD Grant Award
Publication Date
Feb 14, 2019
Source ID
W911NF1910058

Entities

People

  • Dhruv Batra

Organizations

  • Army Contracting Command
  • Georgia Tech Research Corporation
  • United States Army

Tags

Fields of Study

  • Computer science

Readers

  • Agent-Based Social Robotics and Mobile-Assisted Learning in Virtual Environments.
  • Neural Network Machine Learning.
  • Systems Analysis and Design

Technology Areas

  • AI & ML
  • AI & ML - DoD AI Strategy
  • AI & ML - Neural Networks
  • Autonomy
  • Autonomy - Human-Robot Interaction