eduzhai > Applied Sciences > Engineering >

NOVAS Non-convex Optimization via Adaptive Stochastic Search for End-to-End Learning and Control

  • Save

... pages left unread,continue reading

Document pages: 17 pages

Abstract: In this work we propose the use of adaptive stochastic search as a buildingblock for general, non-convex optimization operations within deep neuralnetwork architectures. Specifically, for an objective function located at somelayer in the network and parameterized by some network parameters, we employadaptive stochastic search to perform optimization over its output. Thisoperation is differentiable and does not obstruct the passing of gradientsduring backpropagation, thus enabling us to incorporate it as a component inend-to-end learning. We study the proposed optimization module s properties andbenchmark it against two existing alternatives on a synthetic energy-basedstructured prediction task, and further showcase its use in stochastic optimalcontrol applications.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×