Statistical Learning
1. The Problem
quantitative hypothesis evaluation/comparison according to given data
pattern association, classification, and recognition; data mining
allowing exception and approximation, learning under uncertainty
2. Proposed Solutions
learning in a Bayesian network vs. learning a Bayesian network
model-based parametric learning vs. instance-based learning
neural networks:
- Static description: state vector and weight matrix
- Short-term changes: non-linear activation spreading
- Long-term changes: parameter (weight) adjusting
- Overall objective: function learning
Demo:
pattern classification, principal component analysis, associative memory
3. Issues
the expressive power of vector space, data set as table
availability of training data and time: one-shot, incremental, real-time
explanation power
4. Reading
Chapter 20
5. Ideas
Bayesian network:
The Limitation of Bayesianism
neural network:
Artificial General Intelligence and Classical Neural Network