|
In this chapter, Ballard reviews the most important issues covered in his book.
After reminding the importance of minimum description length as a general criterion, the following three
principal topics are addressed. Regarding memories, emphasis is placed on content-addressable memories
and it is even suggested that the cortex and cerebellum could be modelled according to this paradigm.
As far as programs are concerned, Ballard emphasizes the role of secondary rewards and stresses the
relevance of hidden Markov models and reinforcement learning. It is also observed that this latter
strategy is not yet capable of simulation in absence of input, such as more developed animals do (e.g.
rehersing actions). The third major issue, namely systems, is discussed mainly with respect to genetic
algorithms and genetic programming. The last subsection in Ballard's, though very short, concludes the
book by observing that much of the principles underlying natural computation can be traced back to Marr's
emphasis on programs in detriment of more microscopic approaches such as neural networs. After all, it was
Marr's conviction that trying to make sense of the brain in terms of neurons is like trying to understand
the flight of birds in terms of their feathers [Marr
].
|
|
|