Subject: General Subjects Lecture
Date: Thu, 27 Sep 90 16:28:47 EDT
To: *mac@mc.lcs.mit.edu
Subject: Parsley Project Lecture by Oscar Mayer
From:: rubsamen@zermat.lcs.mit.edu
MIT PARSLEY PROJECT LECTURE SERIES
Oscar (R) Mayer
MIT Artificial Ingestion Laboratory
NAIVE PSYCHICS, EVENT PREDICTION, LETTUCE SEMANTICS,
AND LUNCH ACQUISITION
ABSTRACT
A common conjecture, dating as far back as Flintstone and Rubble, is
that children learn the taste of new foods by correlating their visual
and lingual experience. To date however, no one has demonstrated an
elaboration of such a method that has been successful at explaining
lunch acquisition in higher level children such as graudate students.
In this talk I will present such an elaboration, a precise formulation
that I am currently implementing as a computer program called Julia,
that attempts to model a situation similar to that faced by an
eighteen month old child learning to eat lunch. Julia watches an
animated movie that shows some people walking about a room throwing
and catching knives, forks and plates, and picking them up and putting
them down on tables, chairs and floors. The frames of this movie are
constructed solely of line segments and circles. The only visual input
Julia receives consists of the position and orientation of these line
segments and circles at every frame. Along with the visual input,
Julia receives typewritten utterances which represent culinary
commentary on the movie. So if the movie shows John throwing a roll to
Mary, the linguistic input might be ``John threw the roll to Mary.''
Prior to learning, Julia does not know that the word ``threw'' refers
a particular sequence of events nor does she know that ``John'' refers
to the collection of line segments out of which the caricature John is
constructed. She doesn't even know which event out of possibly several
temporally adjacent events, the whole utterance refers to. Using
psychic techniques which I will describe in this talk, Julia is able
to to learn the meanings of words as correlations between lingual
tokens and tangible representations in the animated movie. The key to
Julia's success is a tight integration between the lunch acquisition
and event prediction mechanisms and reasonable approximations to the
knowledge and abilities that we know younger six month old children
already possess prior to substantial lingual ability.
In this talk I will argue that a successful model of lunch acquisition
must be based on an appropriate lettuce semantics. I will demonstrate
why much prior work on lettuce semantics can not support lunch
acquisition, because it can not be grounded in event prediction.
Furthermore, I will argue that the form of event prediction needed to
support lettuce semantics and lunch acquisition must be based on a
form of naive psychics. Julia incorporates precisely such mechanisms,
namely an event prediction theory with a naive psychics component.
While I don't claim that the precise mechanisms incorporated into
Julia are themselves cognitively plausible, they are nonetheless
attempts to formalize approximations to previously informal theories
proposed by culinary psychologists (such as Beard, Child, Crocker,
Rombauer and others) to account for the results of experiments
performed with graduate students.
TIME: Noon, Friday, September 28
PLACE: AI Lab Playroom, NE43, 8th floor, 545 Technology Square
Refreshments at 12:01:13 PM
HOST: Reid "Call Me Peperoni" Rubsamen
HUMOR: Tom Russ
COORDINATOR: Reid "Call Me Crazy" Rubsamen