Model-Agnostic Meta-Learning (MAML) for Deep Learning Networks

By Chelsea Finn, Pieter Abbeel, Sergey Levine

A Report by Ambar Prajapati

The original research paper is available here

Human beings are good learners. They quickly adapt to new places, learn new skills and languages. Can we think of the same for machines?

This research paper proposes an algorithm for meta-learning independent of any model that can enhance the machine-learning in just a few attempts.

Learning to learn is called Meta-Learning.

Benefits of the proposed MAML algorithm

It can solve new learning tasks using only a small number of training samples
It uses generic mechanism and computation applicable to a variety of tasks
It works well with supervised classification with fewer parameters
It readily applies to regression and can accelerate reinforcement learning for task variability
It applies to different model types, e.g., fully connected and convolutional networks
It applies to various domains, e.g., few-shot regression, image classification, reinforcement learning


The fast learning is achieved by primarily modifying the top layer weights in a feedforward model

The initial parameters of the model are trained with — gradient descent procedure to produce good generalized performance — using

a small number of gradient steps and
a small amount of training data from a new task

The learning process maximizes the sensitivity of the loss functions of new tasks with respect to the parameters. Thus when the sensitivity is high, small local changes to the parameters can lead to large improvements in the task loss.

Let us consider a model, denoted f, that maps observations x to outputs a. The learning task T is defined as -

For independent and identically distributed (IID) supervised learning problem, the length H=1

The model may generate samples of length H by choosing an output at for each t.
The loss

provides task-specific feedback, in the form of another loss or a cost function

Thematic View

The below diagram provides a thematic view of this optimization -

For a given set of tasks p(T) and shots K, the model learns a new Task Ti

The MAML algorithm

The algorithm for Model-Agnostic Meta-Learning (MAML) is given below.
It calculates feedback/loss/cost LTi and corrects the earlier initialized θ

The MAML algorithm for Supervised Learning

This algorithm can be modified to adapt for Supervised Learning as below

The MAML algorithm for Reinforcement Learning

This algorithm can be modified to adapt for Reinforcement Learning as below

This paper introduces a meta-learning approach (MAML algorithm) with several benefits. It is simple and can be combined with any model using gradient-based training, including classification, regression, and reinforcement learning.



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store