4
$\begingroup$

I was looking around and I can't find a good answer, I just want to know why it can be considered as learning and is not just "calibration" or "parametrization".

I feel the word "learning" is overqualified for the things the models do.

Thanks in advance.

$\endgroup$
5
$\begingroup$

When a baby tunes the connections between neurons in his brain until he can recognize a dog and say "this is a dog!" we say that he learned.

Why wouldn't we say the same about neural networks (or other models) that happened to be inside a computer and not in a human brain?

$\endgroup$
1
$\begingroup$

First of all it's a question which can lead to some awesome discussions from the community :)

To me, calling the process of training the model "parametrization" is somewhat correct as essentially what we are doing is to make the model more optimized to solve the problem statement just like Google's Duplex which was able to pass the Turing test for a specific domain only.

But then there comes the concept of Reinforcement Learning where agent's works on maximizing the reward and take actions which could never have thought of. Take the case of AlphaGo where AI tries to understand the game and learns from it's mistake.

$\endgroup$
1
$\begingroup$

When we fit a model, with some parameters in consideration, it is called fitting. However, when we use an optimization method like gradient descent, the parameters gets updated on every iteration leading to modification of parameter just like humans.

Hence it is called learning.

$\endgroup$

Not the answer you're looking for? Browse other questions tagged or ask your own question.