0
$\begingroup$

I have recently started looking more into autoML, where we have a "Controller" system, who outputs architectures and hyperparameters, and is given a reward based on the performance of a system trained using this configuration on a given dataset.

This is of course a very computationally expensive endeavour, since each training step of the Controller has several child networks that need to be trained. However, every time someone runs an autoML task, they will generate dozens if not hundreds of candidate child networks.

Would it be possible, or has it already been done, to have a dataset of child network performance to "bootstrap" an autoML task. This would be a great way to reduce the cost of trying a new autoML architecture.

Are there any downsides anyone can foresee? I have started with the list of Awesome-AutoML-Papers and have not found anything of this nature yet. There have been other efforts to make the process more efficient, such as Pham18 and Liu17, but we still need to start from scratch when trying a new Controller architecture.

$\endgroup$
0
$\begingroup$

(Efficient) Neural Architecture Search uses the on-policy REINFORCE algorithm to train, so unless you change this to some off-policy method, it isn't possible to make use of off-policy rollouts to train.

$\endgroup$

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.