Visual Expertise Is a General Skill

1
Visual Expertise Is a General Skill

• Maki Sugimoto
• University of California, San Diego
• November 20, 2000

2
Overview

• Is the Fusiform Face Area (FFA) really a face
  specific area?
• Our results support the view that it is NOT
• Motivation Evidence for/against the face
  specific view
• Our Approach Our model and experimental design
• Results
• Conclusion

3
Motivation Evidence for the Face Specific View

• Prosopagnosia patients may have deficit in
  identifying individual faces but normal in
  detecting faces or other non-face objects, while
  visual object agnosia patients may be normal with
  face recognition but have deficit with reading or
  object recognition.
• Recognition of faces is more sensitive to
  configural changes than objects.
• Face and non-face objects have separate
  processing mechanisms

4
Motivation Evidence against the face specific
view

• Gauthier et al. points out faces and objects
  differ not only in the image geometries, but also
  in
• Level of discrimination
• Level of experience
• We are face experts.
• FFA showed high activation for a wide variety of
  non-objects when these two conditions were
  controlled.

5
Greeble Experts (Gauthier et al. 1999)

• Activation of the FFA increased when Greebles
  were presented as the training proceeded.
• When subjects met the criteria of experts, the
  activation level differences between faces and
  Greebles were insignificant.

6
Our Hypothesis

• Why does the FFA engage in expert classification
  of non-face objects as well?
• We hypothesized that the FFA responds to visual
  features that are generally useful in
  discriminating homogeneous input images.
• Expertise on one class should facilitate the
  learning of other expert tasks.

7
Model
(Experts)

• Pretrain two groups of neural networks on
  different tasks.
• Compare the abilities to learn a new individual
  Greeble classification task.

(Non-experts)
Hidden layer
8
Database

• 64x64 8bit grayscale
• 5 basic categories
• 12 individuals per category
• 5 different images per individual
• Total of 5x12x5300 images

9
Preprocessing

• Form Gabor jets using 8 orientations and 5
  scales, then subsample on a 8x8 grid.
• For each scale, apply PCA separately and reduce
  dimensionality to 8.
• 8x5x642560 5x840 dimensions

10
Experimental Setting Details
Pretraining tasks

• Fixed configurations
• 1 hidden layer with 40 units
• Learning rate .005
• Momentum .5
• Controlled condition
• 30 training patterns for each basic class

(Experts)
face10
(Non-experts)
11
Training Set Variations
Holdout
Training set
Test

• Training set
• 10 individual for each class, 3 images for each
  individual
• Hold out / test set
• Basic level 3 images of unseen individual
• Individual level 1 unseen image of each
  individual

10
(indiv.)
Holdout
(basic)
Test
12
Topics for Analysis

• Will the experts learn the new task faster?
• Is there a correlation between network plasticity
  and the speed to learn the new task?
• Network plasticity can be defined as the
  average slope of the hidden layer units.

13
Criteria to Stop Training

• Fixed RMSE threshold
• Number of training epochs
• Best holdout set performance
• We eliminated the third criterion due to
  extremely high variance in number of training
  epochs observed in preliminary studies.

14
Experiment 1 (Design)

• For each pretraining task, 20 networks were
  trained, i.e. 20 book experts, 20 face experts,
  etc.
• Training set RMSE threshold was fixed to
• .08 (pretraining)
• .158 (after new task added)
• The threholds were derived from preliminary
  cross-validation experiments on the most
  difficult task, i.e. face expert classification.

15
Experiment 1 (Results)

• Pretraining tasks were much harder for the
  experts.
• Non-experts were significantly slower in learning
  the new task than any of the experts.

Number of Epochs to Achieve RMSE Threshold
16
Experiment 2 (Design)

• For each pretraining task, 10 networks for
  trained for 5120 epochs.
• Intermediate weights were recorded at epochs
  5,10, 20, ... , 2560.
• Total of 11x10110 networks were trained on the
  new task with RMSE threshold .158.

17
Experiment 2 (Results)
Pretraining Epochs and the Speed to Learn New Task

• Non-experts were the slowest to learn the new
  task, provided that the pretraining tasks were
  fully acquired.

18
Experiment 2 (Results)
Pretraining RMSE

• If the pretraining were stopped prematurely, the
  networks must continue improving on the
  pretrained classes as well as the new task.

19
Analysis of Network Plasticity

• Network plasticity can be defined as the average
  slope across all hidden layer units and all
  patterns in a given set of patterns

where
20
Network Plasticity

• Plasticity was lower for the experts!
• It is more appropriate to interpret plasticity as
  a measurement of mismatch.

21
Conclusion

• Expert networks learned the new task faster.
• Expert networks learned faster despite the low
  plasticity, further supporting the claim that the
  hidden layer representation developed for one
  expert class are general features that are useful
  for classifying other classes as well.
• Visual expertise is a general skill that is not
  specific to any class of images including faces.