on 3 November 2013
Recently an idea of computing nature started to take prominent place among philosophers and theoreticians of computing. This is a highly informative book by leading computer scientist framing the topic of computing nature into his central thesis of PAC (Probably Approximately True) learning algorithms. Learning (that is ability of adequate adaptation) is central idea and it is based on the awareness of the physical constraints learning system has - it is never infinite or equipped with perfect information.
Valiant coined the term "ecorithms" for the computational rules which help a system to learn through interactions with the environment. The book relates adaptation with learning (PAC), evolution and cognition in a common naturalistic computational framework. Recommended to anyone interested in how nature works and how its functioning relies on computational strategies.
on 11 October 2016
There's an 'endorsement' of this book that reads "This little book is hugely ambitious..." by a professor of math. That's a wonderfully polite way of saying this guy is WAY out of his depth.
We have a computer scientist trying to articulate his take on evolutionary biology. As if Giants in that field don't exist. And takes on statistics with almost embarrassing naivety. He mentions trying to find correlations in data (p160), without realising correlation is not causation. His Reductionist mindset asks why machine learning is so hard, whilst trying to "curve fit" 20,000 variables to the data. Has this guy been in jail for 30years and not heard of complexity theory, information theory even Bayes rule??? He would do well to read Shannon, Jaynes, even Chaitin in 'Proving Darwin: Making Biology Mathematical'.
He provides examples of 'average loss functions' without realising nature is non Gaussian, non linear and indeed non stationary.
Another howler is his presumption that the world is static (p114) and should be modelled as so, where upon p117 his reductionist mindset is trying to "control errors" as if errors & uncertainty are a nuisance.
One is reminded of the fairground attraction where you hammer down on the pop-up heads, only for it to reappear in another hole for you to realise your entire approach is futile...
This is a guy selling 1970's science to a 2016 audience. At first I found this book cringe-worthy I honestly became embarrassed for this guy. His entire approach is the wrong way round and as if the last 30years of developments in Bayesian inference has never existed.
Save your money...
on 18 October 2015
Some of the comments which complain about style are probably fair. This is what can happen if a book is written by a professional, rather than a professional author. On the other hand, I found this to be one of the most stimulating books I've read all year. I do some machine learning for a living, but there's a difference between understanding how some algorithms work, and understanding why it is even possible to learn in the first place. This book is great at generating "aha" moments. Books that do that are rare, hence my five stars. However, be prepared to go elsewhere if you want to follow up on those insights. This is not an exhaustive textbook on PAC theory.