Learned optimizers: why they're the future, why they’re hard, and what they can do now

The success of deep learning has hinged on learned functions dramatically outperforming hand-designed functions for many tasks. However, we still train models using hand designed optimizers acting on hand designed loss functions. I will argue that these hand designed components are typically mismatched to the desired behavior, and that we can expect meta-learned optimizers to perform much better. I will discuss the challenges and pathologies that make meta-training learned optimizers difficult. These include: chaotic and high variance meta-loss landscapes; extreme computational costs for meta-training; lack of comprehensive meta-training datasets; challenges designing learned optimizers with the right inductive biases; challenges interpreting the method of action of learned optimizers. I will share solutions to some of these challenges. I will show experimental results where learned optimizers outperform hand-designed optimizers in many contexts, and I will discuss novel capabilities that are enabled by meta-training learned optimizers.