As part of an internal study group aimed at getting a good overview of artificial intelligence (AI) and machine learning (ML), we conducted a deep dive into the world of fast.ai.

What is fast.ai ?

Simply defined, fast.ai is a deep learning Python library that is primarily used for adding higher-level functionality in standard deep learning domains. It is also the name of a non-profit research group - founded by Jeremy Howard and Rachel Thomas - focused on deep learning and artificial intelligence.

The fast.ai online course (Part 1 & Part 2) provide a good introduction to a wide spectrum of machine/deep learning techniques and models along with the Python libraries involved in their implementation.

In term of libraries the course covers:

  • PyTorch: world's fastest-growing deep learning library. PyTorch works best as a low-level foundation library, providing the basic operations for higher-level functionality. PyTorch is distinctive for its excellent support for GPUs.

  • Fast.ai: popular library for adding higher-level functionality on top of PyTorch.

  • NumPy: most widely used library for scientific and numeric programming in Python. It provides very similar functionality and a very similar API to that provided by PyTorch; however, it does not support using the GPU or calculating gradients, which are both critical for deep learning.

  • Scikit-learn: popular library for creating machine learning models, using approaches that are not covered by deep learning (like decision trees).

  • Pandas: popular tabular data analysis/processing and querying.

  • Matplotlib: plotting library for the Python programming language and its numerical mathematics extension NumPy.

In term of models the course immerses us in the following:

  • Neural networks (NNs): mainly used for image processing and object detection. 

  • Linear regression: for describing data and explaining the relationship between them.

  • Random forest (combination of decision trees): for classification and regression tasks.

  • Convolutional neural networks (CNNs): current state-of-the-art approach to creating computer vision models.

  • Stable diffusion: generative AI model that produces unique photorealistic images from text and image prompts.

In addition to explaining the models and their potential use cases, the course also provides us with the inner workings of different models and thus widens our technical horizon with new AI-related concepts. Some of these concepts are not easy to grasp at first glance.

Our experience

It is preferable to practice after watching a course. For our practice we entered a Kaggle competition (Spaceship Titanic); some of us practiced the Stable Diffusion model with the DiffEdit implementation (DiffEdit: Editing Images using Generative AI).

Some parts of the course (in Part 2) are very in dept and challenging. These parts would require more than one attempt to fully comprehend them.

Looking at core concepts from multiple perspectives is generally a crucial point in the learning process. Therefore, we are sharing the list of additional material that we found helpful in fully understanding the fast.ai course.

Linear Regression & Random Forest

Introduction to Machine Learning with Python - beginner level

This course first introduces the different steps involved in a ML project and the popular Python libraries used in ML. It then uses decision trees (via the Scikit-learn library) to resolve a basic ML problem: predict the style of music a person – within a specific age group – might like.

Python Machine Learning Tutorial (Data Science)

Contrary to what the course suggests, you do not need to install the Anaconda Distribution to work with Jupyter Notebook. You can just register with Kaggle and create then run your Jupyter Notebooks there.

Build your first machine learning model in Python

The exercise in this course resemble that of the course above but it goes a little bit deeper into the different ML models (Linear Regression, Random Forest) on a quantitative data.

Build your first machine learning model in Python

Build models from scratch in Python

The courses above make use of the implementations of decision trees, linear regression and random forest already provided in the aforementioned Python libraries.

The below courses adopt an interesting approach that consists in building their custom implementations. First the theory of the concepts is spelled out, then that theory is translated into actual Python code and lastly the code is tested against an actual dataset.

How to implement Decision Trees from scratch with Python

How to implement Random Forest from scratch with Python

How to implement Linear Regression from scratch with Python

Maths

Some of the lessons explained in the fast.ai course involve lots of maths. Those not proficient in maths would probably be terrified by some of them. Thankfully there are a lot of online channels that can help. If you want to revisit maths and other related concepts (like derivatives, integrals, etc.) we found the 3Blue1Brown channel very helpful.

For example, you can revisit the essence of calculus here:

3Blue1Brown - Calculus - Visual introductions to the core ideas of derivatives, integrals and more.

You do not necessarily have to watch all of these maths lessons.

3Blue1Brown

Apart from maths, the 3Blue1Brown channel offers great visual presentations and explanations of key neural networks concepts. I highly recommend their presentations, the quality is just incredible!

But what is a neural network? | Chapter 1, Deep learning

Gradient descent, how neural networks learn | Chapter 2, Deep learning

What is backpropagation really doing? | Chapter 3, Deep learning

Backpropagation calculus | Chapter 4, Deep learning

But what is a convolution?

Convolutions | Why X+Y in probability is a beautiful mess

How to Create a Neural Network (and Train it to Identify Doodles)

This course explores how neural networks learn by programming one from scratch in C#, and then attempting to teach it to recognize various doodles and images. From a developer standpoint I believe that the approach used here has lots of merits.

How to Create a Neural Network (and Train it to Identify Doodles)

Stable Diffusion

How Stable Diffusion works

In addition to the fast.ai course we found that it was worth giving the below course a try. It expresses in different words – and examples - the process of image generation with the same core concepts (convolutional layer, U-Nets, autoencoders, etc).

How Stable Diffusion Works (AI Image Generation)

DiffEdit technique

Some of us worked on the Stable Diffusion model via the DiffEdit implementation. The following are some of the papers we found very valuable.

DiffEdit Paper

Stable Diffusion with Diffusers

How does Stable Diffusion work?

How does negative prompt work?

As mentioned earlier, we also created our own implementation of the DiffEdit technique and explained the process in a blog, which you can find here:

DiffEdit: Editing Images using Generative AI

Coding challenges

Lastly, I think you can try the following coding challenges once you have a full knowledge of the Python libraries and deep learning models. The contents are presented in a very entertaining and educational way.

Building a Neural Network with PyTorch in 15 Minutes | Coding Challenge

Building the Gradient Descent Algorithm in 15 Minutes | Coding Challenge

Building a ML Text to Image App with Stable Diffusion in 15 Minutes | Coding Challenge

Conclusion

The fast.ai course is a great introduction to the world of AI and ML. It is a dense course, covering lots of concepts and models, which requires some re-watching and additional material to fully grasp - some of which we presented here.

When we started the study group, my knowledge of AI and ML was very limited. Of course, as a developer, I had heard about - and used - chatbots (ChatGPT, Bard); I had used GitHub Copilot and found it amazing at “guessing” the next line of code. But I had no idea how those apps worked. Now, after the insights I got from the fast.ai course and other online material, I know that they are based on a concept called neural networks. This was one of many eye-openers for me. I am hoping that you get the same insights as I did.

Needless to say that we thoroughly enjoyed the course and we hope that you will too! As previously mentioned, we recommend that you practice after watching the course. Kaggle and Colab are great platforms to write AI code. The Hugging Face platform provides the means to collaborate with the AI community - if you want to.

Thanks for reading!