Learning to learn by gradient descent gradient descent
Authors: Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew Hoffman, David Pfau, Tom Schaul, Nando De Freitas
Optimization algorithms today are typically designed by hand; algorithm designers, thinking carefully about each problem, are able to design algorithms that exploit structure that they can characterize precisely. This design process mirrors the efforts of computer vision in the early 2000s to manually characterize and locate features like edges and corners in images with hand designed features. The biggest breakthrough of modern computer vision has been to instead learn these features directly from data, removing manual engineering from the loop. This paper shows how we can extend these techniques to algorithm design, learning not only features but also learning about the learning process itself.
We show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms outperform standard hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including neural network training, and styling images with neural art.
For further details and related work, please see the paper https://arxiv.org/abs/1606.04474
Check it out at NIPS:
Tue Dec 6th 06:00 – 09:30 PM @ Area 5+6+7+8 #9
Thursday Dec 8th 02:00 – 9:30 PM @ Area 1+2 (Deep Learning Symposium – Poster)
Friday Dec 9th 08:00 AM – 06:30 PM @ Area 1 (DeepRL Workshop – Talk by Nando De Freitas)
Friday Dec 9th 08:00 AM – 06:30 PM @ Area 5+6 (Nonconvex Optimization for Machine Learning: Theory and Practice – Talk by Nando De Freitas)
Saturday Dec 10th 08:00 AM – 6:30 PM @ Area 2 (Optimizing the Optimizers – Talk by Matthew W. Hoffman)
Source: https://deepmind.com/blog/article/deepmind-papers-nips-part-2