Learning Invariance across Attention Shifts

1
Learning Invariance across Attention Shifts
Muhua Li and James J. Clark McGill University,
Montreal, CANADA
Goal of this research work
System overview
Experimental result of attention-shift invariance

• To develop a neural system that achieves
  invariant representation of objects
• In particular, to develop a system that achieves
  position- and attention-shift- invariance

What happens to the visual sensory input when we
shift attention around an object?
Dynamically calculated saliency maps with IOR
across attention shifts

• Positions of the object change (also the
  appearance changes if taken into consideration of
  distortions in the periphery!)
• Local features within the attention window change

Performance evaluation
Compared with other temporal approaches that
learn position invariance, in the case of lacking
temporal smoothness in the visual input
We might have different visual input across
attention shifts
Definition of invariance
Experimental result of position invariance

• Attention-shift invariance
• A set of neurons retain their coherence across
  attention shifts, when such attention shifts stay
  on the same object
• Position invariance
• A set of neurons can discriminate one stimulus
  from others and remain constant response across
  all positions (changes in the visual input are
  mainly induced by overt attention)

The key roles attention shifts play

• Actively select local features using a saliency
  map mechanism
• Attention shift command used to gate the learning
  process
• Temporal constraint of perceptual stability
  across attention shifts

Conclusion
Oriented lines with distortion in the periphery

• Developed a neural network that learns position-
  and attention-shift- invariance
• Incorporated visual-related self-actions such as
  attention shifts to actively select features and
  gate learning
• Solved the problems with time-varying scenery and
  retinal image distortions that plagued previous
  methods

Two levels of learning

• Extraction of position-invariant local features
  across overt attention shifts by correlating pre-
  and post-saccade visual information
• Temporal correlation of local features across
  attention shifts to represent an object as a whole

Real objects with simple shapes
For more information please visit
http//www.cim.mcgill.ca/limh/publications/WAPCV0
4.pdf