Computational Cognition Lecture 3

1
Computational Cognition Lecture 3

• Some cognitively related Classical Applications

2
Classical neural Network Applications Feed
Forward

• Last time we saw
• Perceptron (1960s)
• Feedforward Network (1987)
• Net Talk application (1989)
• Main issues Learns
• Memory in weights
• Connectivity simple
• Note Limited Brain Damage

3
Adaline versus Perceptron

• Perceptron is more natural?
• Adaline (Widrow) is going down on the gradient of
  total error function. Because of nature of
  quadratic function, no local minima.
• Adaline algorithm is generalizable to
  multi-levels so we obtain e.g. back propagation

4
Future Work on Brain Reading

• Can we improve the one-class approach?
• Can we breed second class by using GA as a form
  of co-evolution?
• Can we read (in any way) more complex cognitive
  tasks? (Later in course)

5
Brain Reading

• Clip of Just and Mitchell from 60 minutes
• Actually best two class results from SVM
• One Class methodologies are maybe called for.
• Application of Reading the Mind
• Give Hardoon, Boehm and M. talk
• How far can these ideas be taken? (next time)