Abstract: Much as replacing hand-designed features with learned functions has revolutionized how we solve perceptual tasks, learned algorithms also have the potential to transform how we train machine learning models. Learned optimizers are one such learned algorithm. Instead of writing down mathematical expressions to perform optimization, a learned optimizer learns the function to perform optimization. This talk will outline how these learned optimizers work, and discuss a number of difficulties that arise when training them. Finally I will share some interesting behaviors which are starting to emerge.
Bio: Luke Metz is a research scientist at Google Brain working on meta-learning and learned optimizers. He's interested in building general purpose, learned learning algorithms that not only perform well, but generalizes to new types of never before seen problems.