Atoms of recognition in human and computer vision
- Source: http://dx.doi.org/10.1073/pnas.1513198113
- Author: Shimon Ullman, Assif, Fetaya, Harari
- Related: Convolutional Neural Network, Neural Network, Computer Vision
Claim
Introducing and using minimal recognizable images shows that the human visual system uses features and processes that are not used by current models and that are critical for recognition.
Notes
A
MIRCis an image patch that can be reliably recognized by human observers and which is minimal in that further reduction in either size or resolution makes the patch unrecognizable.
-
minimal recognizable images
- a minimal change in the image can have drastic effect on the recognition
- this identifies crucial features
- phase transition phenomenon
- a minimal change in the image can have drastic effect on the recognition
-
Minimal recognizable configurations (
MIRCs)- the task is computationally difficult as each one is non-redundant
- requires effective use of available information
- the task is computationally difficult as each one is non-redundant
-
training on full-object images
- AP 0.07 ± 0.10
-
training on
MIRC- AP 0.74 ± 0.21
-
results indicate the human visual system uses features and processes that current models do not
- the sharp drop in recognition at the
MIRClevel indicate similar visual representations in humans- the transition occurs for the same images
- feed-forward or top-down processes (currently missing from the
FFmodels)?
- the sharp drop in recognition at the