Meta-learning на собеседовании Data Scientist

Идея

«Learn to learn». Train модель так, чтобы она fast adapts к new tasks с few examples.

Standard learning: many examples → one task.
Meta-learning: many tasks → fast adapt к new task с few examples.

MAML

Model-Agnostic Meta-Learning (Finn 2017).

For каждый task в meta-batch:
  Adapt model на support set (1-5 gradient steps).
  Compute loss на query set с adapted model.
Sum task losses → update meta-parameters.

Optimizes initial weights чтобы after few gradient steps на new task — good.

Pros: model-agnostic. Cons: computationally expensive, second-order gradients.

Prototypical networks

Easier alternative.

For каждый class в support set:
  prototype_class = mean(embeddings of support samples).
For query:
  predict closest prototype.

Episodic training. Works well на image few-shot benchmarks.

Reptile

OpenAI simplification MAML.

Sample task.
Adapt model on task.
Move meta-parameters towards adapted parameters.

Не second-order. Much cheaper. Comparable accuracy.

Применения

Few-shot learning. Adapt новой class с 1-5 examples.

Personalization. Per-user adaptation в recsys.

Robotics. Adapt новому task / environment fast.

Hyperparameter optimization. Learn good HP defaults.

Drug discovery. Adapt new drug class.

В практике — LLM in-context learning largely заменили meta-learning. Но research direction живёт.

Связанные темы

• Few-shot learning для DS
• Self-supervised learning для DS
• Curriculum learning для DS
• Continual learning для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на работах Finn 2017 (MAML), Snell 2017 (Prototypical), Nichol 2018 (Reptile).

Тренируйте Data Science — откройте тренажёр с 1500+ вопросами для собесов.