A tour through the five tribes of machine learning, each with their own master algorithm:
| Tribe | Representation | Evaluation | Optimization |
|---|---|---|---|
| Symbolists | Logic | Accuracy | Inverse Deduction |
| Connectionists | Neural Networks | Squared Error | Gradient Descent |
| Evolutionaries | Genetic Programs | Fitness | Genetic Search |
| Bayesians | Graphical Models | Posterior Probability | Probabilistic Inference |
| Analogizers | Support Vectors | Margin | Constrained Optimization |
The goal of this tour is ultimately to introduce the quest for a single Master Algorithm or unified learner that combines the best parts of all these approaches. Domingos proposed solution is a unified learner that uses a Markov Logic Network (MLN) as the representation, posterior probability as the evaluation function and genetic search coupled with gradient descent as the optimizer. An implementation called Alchemy is downloadable to play with.
The book is an interesting overview of the different approaches to machine learning, but only gives a taste of each, without getting in to any mathematical detail. For someone new to machine learning, they will require to put in a reasonable amount of effort to read this book, whereas for those who understand one or two of the approaches they may feel unsatisfied with the high-level overview of each approach.