1. Memory, Representation and Abstraction

Recall Elman's  Simple Recurrent Network

• Trained on sequences of symbols
• Training was simply prediction
• Symbols represented various aspects of language
1. phonemes
2. letters
3. words
• Found evidence that the network was detecting underlying patterns
1. error levels indicated expectations of word boundaries
2. cluster analysis found words that were used similarly had similar representations

Could this same, simple methodology be used in a non-symbolic world?

Yes, but the real world has some issues:

1. Sometimes there are long stretches of very similar inputs
2. Interesting events can be rare
3. This creates a situation of "catastrophic forgetting"

1.1. The Human Network Experiments

1.2. What can be done about catastrophic forgetting?

Recall our goals BringingUpRobot

2. Governor For Neural Networks

Something like a Self-Organizing Map that sits between the environment and the network that automatically "balances" the categories of training data.

2.1. A Governor for a Feedforward Network

Categories [input] + [output]

2.2. A Governor for a SRN

Categories [input] + [context] + [output]

2.3. It works!