The Perceptron As a Classifier
- Rosenblatt's Mark I perceptron OCR
The simple perceptron
- 20*20 input
- 512 potentiometers with motors
- Output character identification
- How to set weights?
- McCulloch Pitts neuron with N real inputs and one binary output
- MP showed that neuron can compute many useful functions
- Low pass filter
- Image noise removal
- Perceptron can be trained to properly classify a training set
Strength of perceptron as a classifier
- Perceptron computes a dot-product and then thresholds
- Perceptron defines a (N-1) dimensional hyperplane
- All +class patterns on one side, all -class patterns on the other
- Don't need the bias term if increase dimension by one
- XOR can't be calculated
- Cover's theorem for randomly chosen patterns
- Number of hypercube vertex dichotomies D(N)
- Find D(1), D(2), D*(2), D*(3).
(D(N) is the number of dichotmies realizable by a perceptron in N-space.
D* stands for the number of unbiased perceptron dichotomies).
- What is the connection between D(N) and D*(N)? Why?
Back to the CS667 Home Page