sudharsan13296 / hands-on-meta-learning-with-python

Learning to Learn using One-Shot Learning, MAML, Reptile, Meta-SGD and more with Tensorflow

Hands-On Meta Learning With Python

Learning to Learn using One-Shot Learning, MAML, Reptile, Meta-SGD and more

About the book

Meta learning is an exciting research trend in machine learning, which enables a model to understand the learning process. Unlike other ML paradigms, with meta learning you can learn from small datasets faster.

Hands-On Meta Learning with Python starts by explaining the fundamentals of meta learning and helps you understand the concept of learning to learn. You will delve into various one-shot learning algorithms, like siamese, prototypical, relation and memory-augmented networks by implementing them in TensorFlow and Keras. As you make your way through the book, you will dive into state-of-the-art meta learning algorithms such as MAML, Reptile, and CAML. You will then explore how to learn quickly with Meta-SGD and discover how you can perform unsupervised learning using meta learning with CACTUs. In the concluding chapters, you will work through recent trends in meta learning such as adversarial meta learning, task agnostic meta learning, and meta imitation learning.

By the end of this book, you will be familiar with state-of-the-art meta learning algorithms and able to enable human-like cognition for your machine learning models.

Get the book

Check out my Deep Reinforcement Learning Repo here.

Awesome Meta Learning

Check the curated list of Meta Learning papers, code, books, blogs, videos, datasets and other resources here.

Table of contents

1. Introduction to Meta Learning

• 1.1. What is Meta Learning?
• 1.2. Meta Learning and Few-Shot
• 1.3. Types of Meta Learning
• 1.4. Learning to Learn Gradient Descent by Gradient Descent
• 1.5. Optimization As a Model for Few-Shot Learning

2. Face and Audio Recognition using Siamese Network

• 2.1. What are Siamese Networks?
• 2.2. Architecture of Siamese Networks
• 2.3. Applications of Siamese Networks
• 2.4. Face Recognition Using Siamese Networks
• 2.5. Audio Recognition Using Siamese Networks

3. Prototypical Network and its variants

• 3.1. Prototypical Network
• 3.2. Algorithm of Prototypical Network
• 3.3. Omniglot character set classification using prototypical network
• 3.4. Gaussian Prototypical Network
• 3.5. Algorithm
• 3.6. Semi prototypical Network

4. Relation and Matching Networks Using Tensorflow

• 4.1. Relation Networks
• 4.2. Relation Networks in One-Shot Learning
• 4.3. Relation Networks in Few-Shot Learning
• 4.4. Relation Networks in Zero-Shot Learning
• 4.5. Building Relation Networks using Tensorflow
• 4.6. Matching Networks
• 4.7. Embedding Functions
• 4.8. Architecture of Matching Networks
• 4.9. Matching Networks in Tensorflow

5. Memory Augmented Networks

• 5.1. Neural Turing Machine
• 5.2. Reading and Writing in NTM
• 5.3. Addressing Mechansims
• 5.4. Copy Task using NTM
• 5.5. Memory Augmented Neural Networks
• 5.6. Reading and Writing in MANN
• 5.7. Building MANN in Tensorflow

6. MAML and its variants

• 6.1. Model Agnostic Meta Learning
• 6.2. MAML Algorithm
• 6.3. MAML in Supervised Learning
• 6.4. MAML in Reinforcement Learning
• 6.5. Building MAML from Scratch
• 6.6. Adversarial Meta Learning
• 6.7. Building ADML from Scratch
• 6.8. CAML
• 6.9. CAML Algorithm

7. Meta-SGD and Reptile Algorithms

• 7.1. Meta-SGD
• 7.2. Meta-SGD in Supervised Learning
• 7.3. Meta-SGD in Reinforcement Learning
• 7.4. Building Meta-SGD from Scratch
• 7.5. Reptile
• 7.6. Reptile Algorithm
• 7.7. Sine Wave Regression Using Reptile

8. Gradient Agreement as an Optimization Objective

• 8.1. Gradient Agreement
• 8.2. Weight Calculation
• 8.3. Gradient Agreement Algorithm
• 8.4. Building Gradient Agreement with MAML from scratch

9. Recent Advancements and Next Steps

• 9.1. Task Agnostic Meta Learning
• 9.2. TAML Algorithm
• 9.3. Meta Imitation Learning
• 9.4. MIL Algorithm
• 9.5. CACTUs
• 9.6. Task Generation using CACTUs
• 9.7. Learning to Learn in the Concept Space

hands-on-meta-learning-with-python's Issues

Invalid link of dataset

The AT & T's the Database of Faces (https://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html) can't be downloaded, because the DTG group have be dissolved in September 2023 with his retirement. How to use the dataset now?

Wrong directory path in the prototypical network demo notebook with omniglot

The directory path in cell 13 of prototypical network omniglot demo (link) is wrong, it is missing "images" directory. As a result, the output of training is wrong (as can be seen in the notebook, the model is not learning at all during training since the train_data are all zero arrays due to it failing to load from the wrong directory. I was able to train the model correctly after fixing the wrong directory path in cell 13

source code migration from tf1 to tf2

It seems that all the source codes assume tensorflow1.x. Any plan to upgrade to tf2?

Book availability

Hello @sudharsan13296,

I ordered your book " Hands On Meta Learning With-Python" one month ago, but I haven't been delivered yet. Unfortunately this github version is still missing some sections. Do you have any idea about the book finalization and availability to the public?

Best Regards,

Weights computation for MAML in RL Setting

Hello @sudharsan13296

Thank you for this very interesting work.

I have a question regarding section 6.3 "MAML in Supervised Learning".
While in Supervised learning setting, Step 3: (inner loop) is quite obvious, I'm still not sure how to implement it for Reinforcement learning setting. In fact Di consists of K trajectories each one of horizon H. How should theta'i be computed?

A- For each of the Ks trajectories?
B- At the end of the all Ks trajectories training?

In both cases, do you have an idea on how should gradient-descent/losses be operated (eventually aggregated) to obtain theta'i?

Best Regards,