Implications for Computer Vision from Cognitive Science - PowerPoint PPT Presentation

1 / 17
About This Presentation
Title:

Implications for Computer Vision from Cognitive Science

Description:

Artificial systems should be able to mimic these primitive abilities, but, so ... unlearned principles: cohesion, boundedness, rigidity, & no action at a distance ... – PowerPoint PPT presentation

Number of Views:64
Avg rating:3.0/5.0
Slides: 18
Provided by: davidm158
Category:

less

Transcript and Presenter's Notes

Title: Implications for Computer Vision from Cognitive Science


1
Implications for Computer Vision from Cognitive
Science
  • David Minnen
  • Student CogSci Conference
  • April 13, 2001

2
Overview
  • Motivation
  • Portents of Trouble in Categorization
  • Insight from Infant Psychology
  • Implications on Computer Vision
  • Impact on Future Research

3
Motivation
  • The human conceptual systems grows from
    generalizations of sensorimotor experience
  • Infants have no knowledge base and thus must have
    a procedural means for image understanding and
    knowledge acquisition
  • Artificial systems should be able to mimic these
    primitive abilities, but, so far, efforts have
    been largely unsuccessful

4
Background Objectivist Theory
  • Categories in the world are defined by necessary
    and sufficient properties
  • Humans can recognize these natural divisions
  • Internal representations mirror the external world

5
Adult Psychology Problem
  • People dont demonstrate traditional categories
  • Rather, experiments show prototype effects
  • Analysis of problem solving reveals abstract,
    non-deductive processes such as analogy
  • Evidence for use of imprecise reasoning

6
Developmental Psych Problem
  • Adults reason over accessible, propositional
    knowledge infants dont
  • The transition from a procedural to a conceptual
    framework is not understood

7
Lakoffs Explanation
  • Understanding is not purely logical and objective
  • Rather, concepts are based on metaphors from
    embodied experience

8
Support fromDevelopmental Psychology
  • Classical theory
  • Infants are purely sensorimotor organisms that
    follow preprogrammed stages
  • Concepts arise from experience
  • Current trend
  • Humans are born with (primitive) accessible
    concepts which are then refined, not created

9
Perceptual Analysis
  • Local, spatio-temporal stimuli are analyzed
  • Comparisons are encoded in image-schema
  • Image-schema are used to understand new
    experiences

10
Relationship to Computer Vision
  • Machine Vision constrained environment
    characterized by Objectivist assumptions
  • Move toward human perceptual abilities proved
    very challenging
  • Possible explanation Bad background assumptions
    led to poor models

11
Implications
  • Fundamental Problem
  • The Necessary and sufficient conditions
    definition of categories is easily represented in
    computer science terms
  • Embodied cognition implies that an application
    Objectivist ideas will not be successful even in
    theory

12
Possible Resolution
  • Learn from our only example
  • Mimic the process of moving from sensorimotor to
    conceptual reasoning
  • Simulate the process of knowledge retrieval and
    application
  • Problem
  • These processes are exactly what is not
    understood in psychology or neuroscience

13
Implications on Future Research
  • Basis Automatic processes directed by innate
    principles jump start the human conceptual system
  • Mandler posits that perceptual analysis is an
    innate ability
  • Spelke suggests unlearned principles cohesion,
    boundedness, rigidity, no action at a distance
  • Gelman introduces skeletal principles with
    similar intent

14
Implications on Future Research
  • Areas of Focus
  • Motion Coherence Spatial separation
  • Procedural encoding of base principles
  • Transfer between (simple) experience and schema
    (i.e. abstraction reification)
  • Of course, this is easy to say

15
Current Efforts
  • Current systems augment motion analysis with
    other techniques
  • Static, intensity based segmentation - (Weiss 96)
    (Siskind 2000)
  • Principle Component Analysis (Gibson, et al
    2000)
  • This results in better performance relative to
    an adult human

16
Improvements from aCognitive Perspective
  • However, in a cognitive sense, these systems
    cheat
  • The designer is imposing external constraints to
    achieve local performance
  • A better method would be to allow the system to
    err but build processes that use later
    information to detect and correct the mistake

17
Conclusion
  • Current ideas in cogsci help explain the
    difficulties in human-level artificial perception
  • Neuroscience and developmental psychology can
    help refocus efforts
  • One plausible avenue for explanation is the
    artificial recreation of infant perception and
    development
Write a Comment
User Comments (0)
About PowerShow.com