Section 5, Chapter 7 -- 'Inner language', memory, and simple learning

In the previous two chapters, we developed the hypothesis of an inner language to interface between perception and action. We found that this concept would work with memory and the forms of learning we discussed in the third chapter. We also found support for the concept of planning and working with the future, even for animals lower than humans in the evolutionary tree.

We have to evaluate whether the 'inner language' can extend the limited information capacity of the brain - discussed in chapter 2 (page 4).

• The first challenge is to use memory more efficiently, thus increasing our capacity to learn. This need for compression, i.e. for requiring fewer bits to store perception and action for learning, is discussed on the next page. A very approximate estimate of compression is derived from our example of imitation above.

We now have to evaluate whether the 'inner language' can deal with the simple forms of positive-outcome learning we discussed in the third chapter

• The second challenge, matching perceptions, is made much easier because we are comparing imprecise descriptions, not images pixel by pixel.
• The third challenge, reusing action specifications, is also made much easier for the same reason. We are using commands and approximate goals rather than precise muscle movements.
• It still makes sense to include some images and some muscle-based memories of action

We then evaluate whether the 'inner language' can deal with some of the information processing problems uncovered in chapter 4. Hopefully the 'inner language' can help to explain some of the extra functionality required and .

• The fourth challenge is to find a reasonable explanation for multi-trial positive-outcome learning
• This will be addressed in later pages in this chapter.
• The fifth challenge is to find a reasonable explanation for avoidance (negative outcome) learning
• This is a much more difficult problem that will be addressed in later chapters.