Brain-deep AI interface for user-responsive reinforcement learning

This technology is a brain-computer interface that monitors and utilizes users’ neurophysiological responses to improve training of artificial intelligence (AI) systems.

Unmet Need: AI that understands non-verbal cues from human users

Advances in AI are enabling applications where autonomous systems have extended interactions with human users. However, these systems, such as those implemented in driverless cars, can behave in ways that cause discomfort to users. There remains an unmet need for ways to adapt AI behavior to human expectations and reactions in order to improve task performance.

The Technology: Brain-computer interface to improve AI training based on human response

This technology is a system that analyzes and uses the neurophysiological responses of human users as an input for improved deep reinforcement-based learning of AI systems. As users interact with the autonomous system, their physiological signals are continuously and non-invasively tracked using methods such as EEG, pupillometry, and heart rate monitoring. These inputs can be used to determine what types of automated behaviors are stressful or uncomfortable for users, subsequently informing the AI agent to respond in appropriate ways to improve the user’s experience. Furthermore, correlation with neurophysiological signals may help the AI solve problems of higher complexity by leveraging the input of human expertise.

This technology has been validated in a simulated autonomous car environment.

Applications:

• Improved performance of autonomous vehicles and aircraft
• Increased problem-solving capacity of deep learning algorithms
• AI systems for personalized training or coaching
• Virtual and augmented reality environments
• Smart home systems

Advantages:

• Improves AI learning
• Increases users’ comfort
• Non-invasively monitors human physiological response to the environment
• Utilizes multiple types of physiological signals as input
• Enhances personalization of AI systems

Lead Inventor:

Paul Sajda, Ph.D.

Patent Information:

Patent Status

Related Publications:

• Shih V, Jangraw DC, Sajda P, Saproo S. “Towards personalized human AI interaction - adapting the behavior of AI agents using neural signatures of subjective interest” IEEE Trans on Neural Networks and Learning Systems. 2017 Sep 14.

• Shih V, Jangraw D, Saproo, S, Sajda P. “Deep reinforcement learning using neurophysiological signatures of interest” Proceedings of the Companion of the 2017 ACM/IEEE International Conference on Human-Robot Interaction. 2017 Mar; 285-286.

Tech Ventures Reference:

• IR CU16192

• Licensing Contact: Dovina Qu