Microsoft Word Kurzweil, Ray The Singularity Is Near doc

Bayesian Nets.
Over the last decade a technique called Bayesian logic has created a robust mathematical foundation 
for combining thousands or even millions of such probabilistic rules in what are called "belief networks" or Bayesian 
nets. Originally devised by English mathematician Thomas Bayes and published posthumously in 1763, the approach 
is intended to determine the likelihood of future events based on similar occurrences in the past.
168
Many expert 
systems based on Bayesian techniques gather data from experience in an ongoing fashion, thereby continually learning 
and improving their decision making. 
The most promising type of spam filters are based on this method. I personally use a spam filter called 
SpamBayes, which trains itself on e-mail that you have identified as either "spam" or "okay."
169
You start out by 
presenting a folder of each to the filter. It trains its Bayesian belief network on these two files and analyzes the patterns 
of each, thus enabling it to automatically move subsequent e-mail into the proper category. It continues to train itself 
on every subsequent e-mail, especially when it's corrected by the user. This filter has made the spam situation 
manageable for me, which is saying a lot, as it weeds out two hundred to three hundred spam messages each day, 
letting more than one hundred "good" messages through. Only about 1 percent of the messages it identifies as "okay" 
are actually spam; it almost never marks a good message as spam. The system is almost as accurate as I would be and 
much faster. 

---

Markov Models.
Another method that is good at applying probabilistic networks to complex sequences of 
information involves Markov models.
170
Andrei Andreyevich Markov (1856–1922), a renowned mathematician, 
established a theory of "Markov chains," which was refined by Norbert Wiener (1894–1964) in 1923. The theory 
provided a method to evaluate the likelihood that a certain sequence of events would occur. It has been popular, for 
example, in speech recognition, in which the sequential events are phonemes (parts of speech). The Markov models 
used in speech recognition code the likelihood that specific patterns of sound are found in each phoneme, how the 
phonemes influence each other, and likely orders of phonemes. The system can also include probability networks on 
higher levels of language, such as the order of words. The actual probabilities in the models are trained on actual 
speech and language data, so the method is self-organizing. 
Markov modeling was one of the methods my colleagues and I used in our own speech-recognition 
development.
171
Unlike phonetic approaches, in which specific rules about phoneme sequences are explicitly coded by 
human linguists, we did not tell the system that there are approximately forty-four phonemes in English, nor did we 
tell it what sequences of phonemes were more likely than others. We let the system discover these "rules" for itself 
from thousands of hours of transcribed human speech data. The advantage of this approach over hand-coded rules is 
that the models develop subtle probabilistic rules of which human experts are not necessarily aware.