Section 5, Chapter 9 -- Information Transformation Requirements: a summary
We have tried to make a case that information transformation requirements lead us to, and support, the concept of an "inner language", an information encoding scheme that is somewhat like the human languages used for communication.
- Perception needs to be transformed into the specification of action and action sequences -- to account for imitation
- We can see this requirement most clearly in the case of learning by imitation. An action-sequence is perceived and then action is generated that approximates the perceived action.
- "What you see is what you do", but even chicks do it in imprinting
- Novel action sequences need to be generated -- to account for curiosity and play
- We can see this in kittens and puppies playing and exploring the world
- Not all behaviour has to be based on past actions, imitation, and built-in (genetic) action patterns
- Prediction of the future is needed
- Many predators can run a path to catch a prey that are not predicted by a strict "follow your eyes" rule
- The prediction is probably encoded similarly to memories from the past.
- Compression of information is needed
- There are not enough neurons to store all the visual and action information in "raw" form
- The language encoding provides tremendous compression that 'loses' irrelevant detail. On the other hand, it allows the neurons to store a greater variety of experiences (combining vision and action), and thus supports more learning for a greater variety of situations.