A Question About Machine Learning.

[Polymathic]

I think of machine learning as essentially a brute force trial and error bot. Am I wrong in my thinking?

 

[agentrfr]

Not brute force usually

There's some calculus involved to make the directions of variable tuning somewhat sensible (surface derivative calculus if you know what I'm on about)

It isn't random nor is it brute force. It's something in the middle

Neural networks (especially the antagonistic competitive type) are fun for the whole family!

 

[SlinkyMike]

A good place to start to try and get a grip on this is to understand stochastic processes, once you have that bedded down in theory then you can expand your understanding by reading up on backpropagation.

A clear understanding of linear algebra and calculus, as indicated above are prerequisites for discussing these things.

 

[cguy]

I think of machine learning as essentially a brute force trial and error bot. Am I wrong in my thinking?

Supervised learning is somewhat trial and error, in the sense that you run through a training set (trial), and then try to incrementally correct your parameters based on measured error (e.g., backpropagation for NN as mentioned above). It's not brute force in the sense that it actually trys to make predictions/classifications that are correct (or close to correct) even for inputs that are not in the training data. The majority of ML research focuses on improving this extrapolation ability.

 

[animal531]

I wouldn't call it brute force.

A neural network is just a math function where you know the inputs and outputs, but you're not sure how to get from the one to the other. So you use a large dataset to train your model, which modifies the function in the middle to be able to get from the input you provide to the correct outputs. Then in practice you supply it with new/unknown data, and it applies the function/model and basically guesses at the correct answer, based on your previous training.

There are also types of ML where you don't supply a large amount of training data, making it even less brute force.