Neural networks - just automatic feature engineering?

From what I’ve so far, neural networks are nothing more but a way to come up with the most appropriate features automatically. So each hidden layer’s job is to come up with new, more sophisticated features. It is just the last hidden layer that does an extra job - apart from coming up with a new set of features, it then does logistic or linear regression (depending on the problem) to calculate the hypothesis.

I am excited to get to know how exactly the neural network manages to create new features that are actually useful. This sounds so exciting.

I see NN the same way :slight_smile: - that all but the output layer compute new features, and the output layer does logistic/linear regressions with the appropriate activation function.

Hi, @popaqy!

Exactly! A neural network is just an architecture of thousands or millions parameters that tries to infer such values that reduce a loss function as much as possible. Those “features” are probably not what most people think (like corners, borders of pictures, etc) but something more “noisy” (like a sequence of specific colors, etc).

The way it manages to extract those features is with gradient descent (or equivalent algorithm).

1 Like