AtU: Is Tin Sari possible?

[Paolo]

In real life, in this world, is programming something as efficient as Alif's Tin Sari possible?

I find it to be a fascinating idea, but I can't quite imagine that human behavior can be deduced so precisely by a computer.

Although what do I know, for all I know there might be something like this out there already...

[[deleted user]]

The beauty of machine learning is that a programmer doesn't have to understand the rules behind something in order for a computer to learn how to learn it. Right now, Artificial Intelligence can deduce to a fairly high accuracy what is spam and what is not.

I'm not an expert in machine learning so I can't say for certain but my guess is that this type of technology might be crudely possible but won't be accurately possible for a long while. Natural language processing is still struggling with teachings computers to ascertain meaning from language.

Ahaha, I'd like an episode of fact or fictional on this. : D

[Paulo]

As a professional on this field (most specifically, business intelligence & big data), I've seen some significant progress on this area.

There are some commercial software quite capable of inferring patterns after receiving sufficient amount of input data. The idea behind Tin Sari is interesting because to the extend of my knowledge, it is entirely plausible.

It would be too much, of course, to expect this to be developed by one single person working on his bedroom, or that it would be possible to identify a person with just one sentence (as at some point Alif mentions it does).

Thinking on a "conspiracy theory" level, I believe it is already possible to map patterns such as typing speed, languages, common grammatical usage and other similar parameters to identify a single person behavior, given a large enough sample available to train the system. The question is: is the government already doing it? :)

[Steve]

There are these cool little models that you can build called neural networks. What is described for Tin Sari is pretty close to a NN model - you have a series of inputs (in this case text), then you have a series of what are called hidden nodes or weights that are originally chosen more or less at random, which get used by combination to build one or more outputs. If the output is incorrect, this goes in as feedback which changes the hidden nodes and should now generate correct output for that input.

This process is repeated with millions to billions of test cases to ensure that the Neural Network generates correct output. Eventually it can get pretty good.

I shared an office with a man who was researching neural networks - he built a simple virtual robot and used a neural network to train it to dial a phone as efficiently as possible. It took a lot of failures, but it got really good.

Similarly, neural networks can be used to recognize voice by using digital sound as input and, say, a text word as output. With sufficient input, this will eventually hit on a recognizer for that word. I don't recall if this research went anywhere.

[Misti]

Upon first hearing about Tin Sari (I listened to the audio book), I instantly thought of "The Machine" on the TV show Person of Interest. It was created to weed out potential terrorist threats but it developed a personality all of its own and went rogue. However, most of the time you don't know if the number it spits out is a good person who needs saving or a bad dude who needs to be dealt with.

[Sky]

http://thehackernews.com/2015/05/swip...

[Rick]

Ask Amelia - http://www.ipsoft.com/what-we-do/amelia/