Gal A. Kaminka's Publications

Sorted by DateClassified by Publication TypeClassified by TopicGrouped by Student (current)Grouped by Former Students

Cognitive Modeling of Human Intention Recognition

Elisheva Bonchek-Dokow. Cognitive Modeling of Human Intention Recognition. Ph.D. Thesis, Bar Ilan University, 2012.

Download

[PDF]2.4MB  

Abstract

Human beings, from the very young age of 18 months, have been shown to be able to extrapolate intentions from actions. That is, upon viewing another human executing a series of actions, an observing child can guess the underlying intention, even before the goal has been achieved, and even when the performer failed at achieving the goal. In this work, we propose a cognitive model of this human ability, namely, that of intention recognition. The proposed model deals with the challenge of recognizing the intention of an observed sequence of actions, performed by some acting agent. Intention recognition is apparently one of the core components of social cognition. Such a model is therefore important both from a cognitive science point of view and from an engineering perspective. It could provide a deeper understanding of normal and pathological development of human social cognition processes, as well as allowing for artificial implementation of this ability in software agents and physical robots, towards the end of creating more socially intelligent artificial beings. Much work has already been done in all the many areas touching upon this topic, from psychology through neuroscience to artificial intelligence and engineering. In this work we aim to address those aspects of intention recognition which have not yet been treated satisfactorily. We provide a high-level overview of the process as a whole, and detail this model in a way which can explain how failed sequences of actions can be dealt with and their underlying intention extracted, and how novel objects can be dealt with, and goals regarding them predicted, although there is---seemingly---no prior knowledge about them. We elaborate on two components of our proposed model, which we believe to be at its core, namely, those of intention detection and intention prediction. By intention detection we mean the ability to discern whether or not a sequence of actions has any underlying intention at all, or whether it was performed in an arbitrary manner with no goal in mind. By intention prediction we mean the ability to extend an incomplete sequence of actions to its most likely intended goal. The overall structure of the model, i.e. its components and the connections between them, is justified by psychological theories and supported by a plethora of empirical results reviewed in the relevant literature. These theories and experiments are referred to appropriately throughout this work. As for the two core modules on which we elaborate---Intention Detection and Intention Prediction---we present results from several experiments which we have designed and implemented for this purpose. The Intention Detection module is based on a measure of intention, which captures a notion of efficiency, in keeping with the Principle of Rational Action, which states that intentions are brought about by the most rational means available to the actor. This module is validated by two experiments. The first is an artificial emulation of the original intention re-enactment procedure by Meltzoff (1995). The results show that the proposed measure of intention indeed succeeds at categorizing streams of action according to the extent to which they convey an underlying intention. The second experiment validating the Intention Detection module is closer to real life. It uses surveillance videos taken from an online database, and analyzes them according to the proposed measure of intention. This analysis is then compared to human judgment of intention on the same videos. The resulting correlation between the output of our module and that of the human subjects is high, showing once again that our measure of intention indeed captures the notion of intention present in action. Like the Intention Detection module, the Intention Prediction module is based on a measure of intention as well. This measure is also designed to be in line with the Principle of Rational Action, however, it is formalized differently, for reasons which will be discussed. The Intention Prediction module also makes use of the psychological notion of affordances, for extracting goal states from objects in the environment, which the observed actions might possibly be intending to realize. In order to test this second measure of intention as far as its usefulness for predicting intention, we designed an online experiment in which human subjects were presented with abstract objects (various geometric shapes), and were asked to predict the end-configuration of the objects, which observed sequences of movements were aiming to achieve. The predictions arrived at by our measure of intention were compared to the human results, and proved to be highly reliable. Other possible measures, such as proximity of the terminal state arrived at by the actions to the various goals, were also considered. However the success of these measures at predicting the intended goal was inferior to that of our measure of intention, and they at most play a secondary role in the process. To conclude our work, we summarize our findings and propose several directions for future research on intention modeling. We hope this work will be of interest and of use to researchers in the multidisciplinary communities dealing with intention recognition, and look forward to seeing the ideas proposed here implemented in socially cognitive artificial systems.

Additional Information

BibTeX

@PhdThesis{elisheva-phd, 
author = {Elisheva Bonchek-Dokow}, 
title = {Cognitive Modeling of Human Intention Recognition}, 
school = {{B}ar {I}lan {U}niversity}, 
year = {2012}, 
  wwwnote = {}, 
OPTannote = {} ,
  abstract = { 
Human beings, from the very young age of 18 months, have been shown to be able to extrapolate  intentions from 
actions. That is, upon viewing another human executing a series of actions,  an observing 
child can guess the underlying intention, even before the goal has been achieved, and even when the  performer 
failed at achieving the goal. In this work, we propose a cognitive model of this human ability, namely,  that 
of intention recognition. 
The proposed model deals with the challenge of recognizing the intention of an observed sequence of actions, 
performed by some acting agent. Intention recognition is apparently one of the core components of social 
cognition.  Such a model is therefore important both from a cognitive science point of view and from an 
engineering  perspective. It could provide a deeper understanding of normal and pathological development 
of  human social  cognition processes, as well as allowing for artificial implementation of this ability 
in software agents and  physical robots, towards the end of creating more socially intelligent artificial 
beings. 
Much work has already been done in all the many areas touching upon this topic, from psychology through 
 neuroscience to artificial intelligence and engineering. In this work we aim to address those aspects 
of intention  recognition which have not yet been treated satisfactorily. We provide a high-level 
overview of the process as a  whole, and detail this model in a way which can explain how \emph{failed} 
sequences of actions can be dealt with  and their underlying intention extracted, and how \emph{novel} 
objects can be dealt with, and goals regarding them  predicted, although there is---seemingly---no 
prior knowledge about them. 
We elaborate on two components of our proposed model, which we believe to be at its core, namely, 
those of  intention detection and intention prediction. By \emph{intention detection} we mean the 
ability to discern whether  or not a sequence of actions has any underlying intention at all, or 
whether it was performed in an arbitrary  manner with no goal in mind. By \emph{intention prediction} 
we mean the ability to extend an incomplete sequence of  actions to its most likely intended goal. 
The overall structure of the model, i.e. its components and the connections between them, is justified 
by  psychological theories and supported by a plethora of empirical results reviewed in the relevant 
literature. These  theories and experiments are referred to appropriately throughout this work. As 
for the two core modules on which  we elaborate---Intention Detection and Intention Prediction---we 
present results from several experiments which we  have designed and implemented for this purpose. 
The Intention Detection module is based on a measure of intention, which captures a notion of efficiency, 
in  keeping with the Principle of Rational Action, which states that intentions 
are brought  about by the most rational means available to the actor. This module is validated by two 
experiments. The first is  an artificial emulation of the original intention re-enactment procedure 
by Meltzoff (1995). The results show  that the proposed measure of intention indeed succeeds at 
categorizing streams of action according to the extent to  which they convey an underlying intention.  
The second experiment validating the Intention Detection module is closer to real life. It uses 
surveillance videos  taken from an online database, and analyzes them according to the proposed measure 
of intention. This analysis is  then compared to human judgment of intention on the same videos. The 
resulting correlation between the output of  our module and that of the human subjects is high, showing 
once again that our measure of intention indeed captures  the notion of intention present in action. 
Like the Intention Detection module, the Intention Prediction module is based on a measure of intention 
as well.  This measure is also designed to be in line with the Principle of Rational Action, however, 
it is formalized  differently, for reasons which will be discussed. The Intention Prediction module also 
makes use of the  psychological notion of \emph{affordances}, for extracting goal states from objects in 
the environment, which the  observed actions might possibly be intending to realize. 
In order to test this second measure of intention as far as its usefulness for predicting intention, we 
designed an  online experiment in which human subjects were presented with abstract objects (various 
geometric shapes), and  were asked to predict the end-configuration of the objects, which observed 
sequences of movements were aiming to  achieve. The predictions arrived at by our measure of intention 
were compared to the human results, and proved to  be highly reliable. Other possible measures, such as 
proximity of the terminal state arrived at by the actions to  the various goals, were also considered. 
However the success of these measures at predicting the intended goal was  inferior to that of our measure 
of intention, and they at most play a secondary role in the process. 
To conclude our work, we summarize our findings and propose several directions for future research on 
intention  modeling. We hope this work will be of interest and of use to researchers in the multidisciplinary 
communities  dealing with intention recognition, and look forward to seeing the ideas proposed here 
implemented in socially  cognitive artificial systems. 
},
}

Generated by bib2html.pl (written by Patrick Riley ) on Sun Jul 23, 2017 22:08:49