1
$\begingroup$

According this thread some hyperparameters are independent from each other while some are directly related.

One of the answers give an example where two hyperparameters affect each other.

For example, if you're using stochastic gradient descent (that is, you train your model one example at a time), you probably do not want to update the parameters of your model too fast (that is, you probably do not want a high learning rate), given that a single training example is unlikely to be able to give the error signal that is able to update the parameters in the appropriate direction (that is, the global or even local optimum of the loss function).

How would someone creating a neural network know how the hyperparameters affect each other?

In other words, what are the heuristics for hyperparameter selection when trying to build a robust model?

$\endgroup$
1
$\begingroup$

This is one of the most difficult and unsolved problems in machine learning and deep learning!

There are many different ways to estimate the most appropriate hyper-parameters, such as grid search, random search, Bayesian optimization, meta-learning, reinforcement learning, and evolutionary algorithms (e.g. NEAT).

However, the problem is that most if not all these approaches are typically not very computationally efficient if your model isn't anything but very small. The number of possible configurations of the hyper-parameters is very large. If you don't have good computational resources (e.g. GPUs and powerful servers), you probably out of luck or require some days to get some insight.

In certain cases, it is obvious that certain hyper-parameters are dependent on each other, e.g. in the case of the batch size and the learning rate, because we have some decent understanding of gradient descent, but, in other cases, the situation isn't so nice.

As far as I know, there isn't a very good general rule of thumb or method to solve this issue (i.e. find the dependence of hyper-parameters on each other). Maybe, as our knowledge of our models (especially, neural networks) increases, we'll get some more insights and we'll develop more efficient approaches to understand the dependence of the hyper-parameters.

Nowadays, there's automatical machine learning (AutoML), which is a fancy name to denote services that provide hyper-parameter optimization (plus some other stuff).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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