Customer Reviews

Truth from Trash: How Learning Makes Sense (Complex Adaptive Systems)

5 star:		(1)
4 star:		(0)
3 star:		(1)
2 star:		(0)
1 star:		(1)

 

 

Sure, the book jumps around a bit, is patchy when it comes to technical details and is fairly poorly referenced, but there's some interesting and inspiring ideas here. However, if you're from a country with a lousy exchange rate with the US (such as poor old Australia) then wait for the paperback edition!

I was very disappointed by this book. He makes a valid point that most machine learning research is concerned with attribute-based (propositional) representations, and that many problems require relational (first-order) representations, but this is not a novel claim.

He calls propositional learning "fence'n'fill" algorithms, because they basically carve up the input space (e.g., a perceptron uses linear boundaries). The advantage is that they are fast and well-understood. In the final chapter, he proposes an algorithm for relational learning which is based on top of a standard fence-n-fill algorithm, but doesn't explain it well, and doesn't give any compelling evidence that it works. The papers on his web site are no better.

He intersperses what little technical material he has with some historical anecdotes about code-breaking during WWII, etc. It's not really clear what the connection is. Overall, the book just does not hang together.

If you felt inclined to buy this book, I would recommend you check out Andy Clark's excellent "Being There" instead.

This is a very good book. I'm surprised it is out of print and overlooked (0 review on Amazon so far).

The thesis is that spatial classification techniques (such as neural networks, support vector machines, principle component analysis, etc) are inadequate for classifying certain types of data that have "relational" or "logical" structure. This is related to the "propositional fixation of neural networks" as pointed out first by John McCarthy, but Thornton takes a more detailed look into the problem.

I may return here to write more about it...