[Prev][Next][Index]

CSLI Events, 21 and 22 March 1990

		SEMINAR ON COMPUTERS, DESIGN, AND WORK
		       Design Course Discussion
		  Terry Winograd and Brad Hartfield
		     (winograd@csli.stanford.edu)
		   Wednesday, 21 March, 12:15 p.m.
			      Ventura 17

We will discuss our plans for a design-oriented course on
human-computer interaction that we will be inventing next quarter.  It
is an experiment and we are looking for suggestions and critiques that
will help us make it work.

COURSE DESCRIPTION (PRELIMINARY VERSION, WHICH MAY CHANGE AS WE GO)

CS 247: Human-Computer Interaction
Terry Winograd with Brad Hartfield (Apple)
Spring 1989-90 

Issues of human-computer action: including interface design, interface
styles, work design, communication structure, and organizational
factors.  Students in small groups will develop substantial
user-interface prototypes of systems for situations of actual use,
applying concepts from the course readings and interacting in project
reviews with faculty and other experienced system designers.

Primary audience: CS Masters students and upper division undergraduates.  
Prerequisites: CS109, experience in C, Lisp, or Hypertalk.
			     ____________
				   
			    CSLI TINLUNCH
		  Perceiving Sound Patterns in Time
			     Robert Port
	    Departments of Linguistics, Computer Science,
	      and Cognitive Science, Indiana University
		     (port@iuvax.cs.indiana.edu)
		    Thursday, 22 March, 12:00 noon
			     Cordura 100

How can we perceive patterns that are distributed over time?  The
standard view requires a "time window," in which time is mapped onto
physical distance (as in a sound spectrogram).  But it will be argued
that time windows are biologically implausible as a representational
basis for recognition of temporal patterns like words.  I will
describe connectionist simulations of auditory perception that are fed
a single spectrum-slice at a time, and are trained to recognize
melody-like patterns.  The networks have a recurrent memory module of
multiplicative (or sigma-pi) units.  Each target "tune" produces a
stable trajectory in the state space of the module.  This dynamic
memory learns temporal patterns without saving the inputs themselves,
but by representing relevant information about history an abstract
space.  This representation has many useful properties, including
allowing recognition as early in time as the information in the
stimulus allows, and a tendency to be invariant under changes in rate.