google-deepmind/learning-to-learn

Implements a meta-learning framework where an LSTM-based optimizer is trained end-to-end to learn task-specific update rules, using unrolled optimization trajectories and optional second-order derivatives. Built on TensorFlow and Sonnet, it enables comparison of learned optimizers against baselines like Adam across synthetic quadratic problems and real datasets (MNIST, CIFAR-10). Supports custom problem definitions through pluggable loss functions, with configurable unroll lengths and multi-optimizer setups for heterogeneous parameter groups.

4,071 stars. No commits in the last 6 months.