Microsoft Word Kurzweil, Ray The Singularity Is Near doc

Combining Methods.
The most powerful approach to building robust AI systems is to combine approaches, which is 
how the human brain works. As we discussed, the brain is not one big neural net but instead consists of hundreds of 
regions, each of which is optimized for processing information in a different way. None of these regions by itself 
operates at what we would consider human levels of performance, but dearly by definition the overall system does 
exactly that. 

---

I've used this approach in my own AI work, especially in pattern recognition. In speech recognition, for example, 
we implemented a number of different pattern-recognition systems based on different paradigms. Some were 
specifically programmed with knowledge of phonetic and linguistic constraints from experts. Some were based on 
rules to parse sentences (which involves creating sentence diagrams showing word usage, similar to the diagrams 
taught in grade school). Some were based on self-organizing techniques, such as Markov models, trained on extensive 
libraries of recorded and annotated human speech. We then programmed a software "expert manager" to learn the 
strengths and weaknesses of the different "experts" (recognizers) and combine their results in optimal ways. In this 
fashion, a particular technique that by itself might produce unreliable results can nonetheless contribute to increasing 
the overall accuracy of the system. 
There are many intricate ways to combine the varied methods in AI's toolbox. For example, one can use a genetic 
algorithm to evolve the optimal topology (organization of nodes and connections) for a neural net or a Markov model. 
The final output of the GA-evolved neural net can then be used to control the parameters of a recursive search 
algorithm. We can add in powerful signal- and image-processing techniques that have been developed for pattern-
processing systems. Each specific application calls for a different architecture. Computer science professor and AI 
entrepreneur Ben Goertzel has written a series of books and articles that describe strategies and architectures for 
combining the diverse methods underlying intelligence. His Novamente architecture is intended to provide a 
framework for general-purpose AI.
179
The above basic descriptions provide only a glimpse into how increasingly sophisticated current AI systems are 
designed. It's beyond the scope of this book to provide a comprehensive description of the techniques of AI, and even 
a doctoral program in computer science is unable to cover all of the varied approaches in use today. 
Many of the examples of real-world narrow AI systems described in the next section use a variety of methods 
integrated and optimized for each particular task. Narrow AI is strengthening as a result of several concurrent trends: 
continued exponential gains in computational resources, extensive real-world experience with thousands of 
applications, and fresh insights into how the human brain makes intelligent decisions.