Attention Detection with Transfer Entropy

Abstract

Quals talk:

Interpreting a persons focus, or attention, plays a large role in human-robot collaboration and allows for more reactive, attentive agents. People interpret attention from their interaction partner's behavior, and accurately communicating focus improves collaboration, understanding, and learning. Unfortunately attention is not directly observable, and most prior work in this area has taken a passive approach to estimating attention, such as gaze detection and theory-of-mind models. These methods require extensive sensing or the building of context specific models that may not be implementable in the wild, in addition they do not take advantage of the implicit connection between robot action and human response.

Our research goal is to find a general method for detecting a person’s attention in multi-agent settings. Our work builds on prior active methods, first by using a passive, general, information theoretic correlation detector, and then by coordinating agent behavior to actively disambiguate which agents (source) is transferring information to the person (target).

We show the efficacy of this approach in multi-agent simulations of increasing complexity. Our method uses Transfer Entropy (TE) performs competently with much less data than would be otherwise be needed to train a classifier for the same task.

Please join meeting via Zoom: https://tufts.zoom.us/j/97183120811

Meeting ID: 971 8312 0811

Password: See colloquium email

Dial by your location: +1 646 558 8656 US (New York)

Meeting ID: 971 8312 0811

Passcode: See colloquium email