Some neural network architectures work better with RMSprop than e.g. ADAM. So for example stated by DeepMind in their work with Atari games and reinforcement learning.

Maciej Jaskowski reproduced the experiments and stated in 2016:

DeepMind Rmsprop (instead of normal one) - improved performance by 40% in my case.

However, when comparing his implementation with the one from tensorflow I can't spot a difference. Has tensorflow updated their RMSprop? They both even cite the sixth lecture of Tieleman T. in their code.

Another explanation could be that Jaskowski mentioned the improvement in his blog, but didn't released this change in the code. Highly doubt that, since he comments in the run.py file mulitple times "deepmind rmsprop".


Your Answer

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

Browse other questions tagged or ask your own question.