Model-Agnostic Meta-Learning for Fast Adaption of Deep Networks

Introduction

• Meta-learning: learning to learn
  • Goal of the trained model is to quickly learn a new task from a small amount of new data
  • Model is trained by the meta-learner to be able to learn on a large number of different tasks
• Key Idea: train model’s initial params such that model has maximal performance on a new task after the params have been updated through few gradients steps with some small data for the new task.
  • Can be viewed as building an internal representation which is suitable to many tasks, fine-tuning the params slightly for good perform.
  • the process optimizes for model that are fast and easy to fintune
• From dynamical systems view, maxing the sensitivity of the loss functions of new tasks with respect to the params
  • when sensitivity is high, small local changes to params can lead to largements in the task loss.
• Model and task-agnostic algorithm

Model-Agnostic Meta-Learning

Meta-Learning Problem Setup

• meta-learning problem treats entire tasks as “training examples”
• Training phase:
  • Sample a task from a distribution of task
  • Sample K input samples from the task
  • Get output from model
  • calculate loss of task
  • test it on unseen data from the task
  • model improved by considering the test error on unseen data changes with parameters (this is training error of the meta-learning process)
• Generally, tasks used for meta-testing are held out during meta-training.

MAML Algorithm

• Intuition: some internal representations are more transferrable than others
  • NN might learn internal features that are broadly applicable to all tasks, rather than a single task
  • How to encourage the emergence of such general purpose representations?
• Aim to find model parameters that are sensitive to changes in the task
  • Small changes in parameters will produce large improvements in loss function of any task when altered in the direction of the gradient of that loss.
• Only assumption: loss function is smooth enought that gradient based techniques can be applied.
• Meta-optimization is performed over the model params, wheras the objective is computed using updated model params.
  • MAML meta-gradient updates involves a gradient through a gradient
  • Meta Objective : \(\)