Download
Learning to Learn, created by Google DeepMind, is an experimental framework that implements meta-learning—training neural networks to learn optimization strategies themselves rather than relying on manually designed algorithms like Adam or SGD. The repository provides code for training and evaluating learned optimizers that can generalize across different problem types, such as quadratic functions and image classification tasks (MNIST and CIFAR-10). Using TensorFlow, it defines a meta-optimizer model that learns by observing and adapting to the optimization trajectories of other models. The project allows users to compare performance between traditional optimizers and the learned optimizer (L2L) on various benchmarks, demonstrating how optimization strategies can be learned through experience. The design supports both single-variable and high-dimensional problems, and includes tools for evaluating how well a learned optimizer performs on unseen tasks.
Features
- Trains neural networks to learn optimization strategies through meta-learning
- Supports multiple benchmark problems (quadratic, MNIST, CIFAR-10)
- Includes train and evaluation scripts with configurable parameters
- Allows easy integration of new optimization problems via TensorFlow
- Provides a learned optimizer (L2L) and baseline comparisons with Adam
- Tracks and saves optimizer performance across training epochs
Project Samples
