Notes
🔍

Reptile (Meta-Learning)

Reptile comes from a simple observation: if you want a model that adapts to a new task in a handful of gradient steps, you can look for a starting point that sits close to the good solutions of many tasks at once — and you can find it without any of the second-derivative machinery meta-learning usually demands.

The algorithm
The loop is three lines. Hold a set of initial weights, then repeat:

  • sample a task from the task distribution;
  • run k steps of ordinary SGD on that task's data, starting from the current initialisation, producing adapted weights;
  • move the initialisation a small step toward those adapted weights — the entire update is phi <- phi + epsilon(W - phi).

Nothing is back-propagated through the inner loop, so there is no Hessian, no unrolled graph, and memory cost barely rises above normal training. Introduced by Nichol, Achiam and Schulman at OpenAI, it is the first-order cousin of MAML.

Why it works, and why it matters
The update looks like plain averaging of task solutions, but its Taylor expansion carries a term that favours initialisations where different tasks' gradients agree — minima that sit near one another. That is why it matches first-order MAML on few-shot benchmarks such as Omniglot and mini-ImageNet while being far cheaper to run. For AutoML the appeal is the learned warm start: an initialisation that collapses adaptation on a new task from thousands of steps to a few, which is precisely the economy a system evaluating many tasks needs.

Reptile is a first-order meta-learning algorithm that learns a weight initialisation from which new tasks can be solved in a few gradient steps. It repeatedly samples a task, runs several steps of ordinary SGD from the current initialisation, then moves the initialisation toward the result. Because it never differentiates through the inner loop, it avoids the second derivatives MAML requires while matching its few-shot accuracy — a cheap route to the learned warm starts AutoML depends on.

A chef who has cooked across many cuisines picks up an unfamiliar dish quickly — not because they know that recipe, but because their instincts already sit somewhere useful to all of them. Reptile is a method for giving a computer model that kind of starting point. It practises on lots of small, different tasks. After each one, it nudges its starting settings a little toward whatever worked for that task. Over time the starting point drifts to a spot that is close to being right for everything. Learning a brand-new task then takes a few quick adjustments instead of starting from scratch.