Sunday, November 18, 2007

Technology Notes, Vol 1, Issue 4: Machine learning.

This week, I'll write about my experiments with time, my thoughts on Machine learning, and a little bit of evolution theory.

* The speed of time
As I was writing my earlier post, "Nostalgia :80s", I kept thinking how time has flown since that time. Well, yes, but how fast does time really fly? I thought a while, and came to the conclusion that it is c. The speed of light. The speed of time should be the same as the speed of light.

What proof can I offer? Well, we know that nothing can travel faster than the speed of light - and time travels pretty fast -so, the speed of time must be c.

On a more mathematical basis: we know that t(v) = t(0) sqrt ( 1 - v^2/c^2). Therefore, when the observer reaches the speed of light, time stops. This means, that the relative velocity of the observer and time is zero - which means, v(time) - c = 0, or v(time) = c.

* The problem with Machine learning

If you are working on speech recognition, face detection, natural language processing or an other field that essentially tries to mimic the functionality of the human brain, you'd probably be using machine learning. Machine learning is the sort of intersection of the human and the computer - a science where humans tell computers how to learn based on vast amounts of solved problems. The deductions made are usually based on sound statistical methods, details of which are available at the wikipedia page.

The technique has taken huge strides in areas such as face detection, speech recognition and machine translation. However, complete success eludes it. Why?

My opinion is that the technique is trying to solve the problem by breaking it into parts, while the problem should be dealt with, at a holistic level. For instance, your brain interprets the whole image, while programs interpret colour encoded as bytes. Or does the brain really interpret the image as a whole? Hard to say. But it appears to be so - it can "understand" that a brightly lit part of a metallic object that reflects large amounts of light still belongs to the same object. It can filter a speaker's voice from the noise in the environment, almost automatically. It can translate sentences between languages taking care of ambiguities in tense, grammar, and meaning.

It'll be interesting to find out if this is really so.

* More choice isn't necessarily good

Don't believe me? See this ad for Vista choices: The sad thing about it is that it is really true. Recently a friend was making a movie in Windows movie maker and wanted to tweak a few settings. He was using Vista Enterprise, and this option - a single menu item that is default with Windows Movie Maker on XP - is available in the Home Preminum (and Ultimate) edition! How on Earth has choice helped the user?
See this talk for a more scientific explanation.

* On evolution

As some of you know, I've been reading a lot of Richard Dawkins of late. I've also been "evangelising evolution". In the process, I found that most people have misunderstandings about how the process works. This is a small attempt to clarify some of them.
  • First, evolution to create the present set of species has taken millions and millions of years, a time frame most of us cannot fathom.
  • Speciation (or the development of a new species by evolution) requires a separation between members of a species and a separation of their environments .
  • Evolution is occurring even today: Why do mosquitoes become resistant to DDT? Why do we need newer and newer strains of antibiotics? Why indeed is the AIDS virus so successful at avoiding every medicine we throw at it? Well, the answer is just one - natural (or in this case, human) selection. When we spray DDT or take antibiotics, we target and kill most mosquitoes(bacteria) in the environment. However, the few that due to some mutation survive, being selected by human selection, have the upper hand in reproduction, and spread the immune gene throughout the population, making the entire population resistant.
  • Evolution does not happen only by random events: It is indeed true that a hurricane blowing through a garage of aeroplane parts will not assemble a Boeing 747, and it is true that random mutations on their own will not lead to speciation. However, random mutation in combination with non-random natural selection, that selects the fittest, either by killing off the weak, or by having mates choose a particular trait in the other sex.

Have been reading too much evolution. I hope to start reading more Physics from now on. Keep visiting!

BTW, I know my "speed of time" hypothesis is wrong, and if it isn't, I take no credit for getting it right. It was just a random thought that popped into my head.

No comments: