Transcript Psychophysical and Physiological Evidence for Viewer

Psychophysical and Physiological
Evidence for Viewer-centered Object
Representations in the Primate
N.K. Logothetis and J. Pauls
Cerebral Cortex (1995)
Background
Image
Input
Transformations representation
Theories of object representations
Match ?
Recognition
Stored memory
representations
3D models (Marr, Biederman)
View dependent 2D templates (Basri & Ullman, Poggio)
Face selective cells
Found in STS
Mostly view dependent
Methods
Trained three juvenile rhesus macaques on an object
recognition task
Performed psychophysical tests after training
Recorded from the upper bank of the anterior medial
temporal sulcus (AMTS)
Stimuli:
Computer generated ‘wire like’
and ‘amoeboid’ objects
Training
Began with training
monkeys to recognize a
single view of an object
presented sequentially
among distractor objects
Slowly increased rotations
up to + or – 90o before
training with a new object
Feedback with juice reward
Testing
Recordings
Recorded from 773 neurons in AMTS
Findings—psychophysical
Recognition performance fell off sharply when object
rotated more than 30-40o beyond training view
Both for wire and amoeboid objects
Findings—psychophysical
Interpolation with wire objects
Monkeys could interpolate between two training views up
to 120o apart
Three to five views
allowed monkey to
generalize to entire
‘great circle’
Findings—psychophysical
‘Pseudo-mirror symmetrical’ wire objects
Some of the wire
objects have mirror
symmetrical 0o and
180o views due to lack
of self-occlusion
Findings—psychophysical
Viewpoint invariance for ‘basic’ objects
among different class distractors
Findings—physiological
View specific,
object specific
cells (71 of 773)
Cell responses to distractor views
Cell responses to target views
Findings—physiological
View invariant,
object specific
cells (8 of 773)
Findings—physiological
Findings—physiological
Multiple cells tuned to
different views of the same
object
Author’s conclusions:
Object recognition depends on training view
A small number of stored views can be used to achieve
invariance with wire like objects
Neurons in IT found that respond selectively to learned
objects, mostly to specific views
Problems:
Highly unnatural stimuli
View selective neurons used for recognition or after
recognition?
Interpolation with self occluded (solid) objects?