Key parts of teaching Machine Learning

Build a 2024 ML Course from Scratch: Insights & Teaching Strategies with JetBrains

Alex Avdiushenko

Hello everyone! I'm Alex Avdiushenko, and I am happy to have a talk here. Today we will discuss the key components of teaching Machine Learning. First, let me briefly tell you that I am currently working on a computer science and artificial intelligence program at Neapolis University in Pafos and at JetBrains company, where I also design and teach courses. We are always glad to share our ideas and thoughts with you and ready for collaboration on different projects related to machine learning and programming, among other topics.
In the past, I worked as a developer at a Yandex company, I was writing code and developing machine learning models. Earlier in my career, I was a research scientist and completed my Ph.D. in Math modeling. This is the kind of experience I bring to the table.

1. Some examples of applying AI to modern problems

2. Practical aspects of creating ML course

Source: kdnuggets.com

Let's start with an image that shows us the simple convolutional neural network, which is fascinating in its own right. So we take an image, convert it into a set of pixels in computer memory, which can be represented by either three numbers for RGB format or even one number, the intensity of color for grayscale images.
We then process it as a matrix, performing simple matrix operations, such as multiplication. After these operations, the neural network model outputs a class label, identifying what is in the image — be it a bird, a dog, the sunset, or something else. The accuracy of these models is now so high that they not only outperform humans but also make all previous methods obsolete. This approach allows us to take an image and get an answer directly, using model as a black box.

Modern machine learning models

And also LLama, Mistral, Grok…

Moreover, these models can now handle a variety of tasks beyond just image classification. For example, they can translate text. You can provide a sentence, let's say, in the Greek language, and the model can translate it to English, and vice versa. The model "understands" what is required and performs the translation.
For sure, you have heard ABOUT the GPT-4 model from OpenAI company, which for a long time was the smartest of all.
Now, OpenAI HAS released GPT-4o1 model, which is even more advanced! It is capable of generating text, code, and even answering hard questions.
Gemini from GOOGLE is another model that can do almost the same and has a really large size of context.
Another competitor in this field is Claude3, which now often is used with Cursor editor to develop software applications ten times faster.
Also, there are many other models, like LLama, Mistral, Grok, and so on.

We can guess that, for instance, autonomous vehicles are being developed and might replace human drivers in the future. THIS is an image of a real Hellenic autonomous vehicle in Cyprus.

Factories, like the one where TESLA cars are assembled, predominantly use robots, though humans are still involved in programming and maintenance. Almost all the work is done by machines, and humans manage the processes from the office.

Of course, machine translation has ALSO advanced significantly. We all use translators on our phones or on the internet, and their quality is often comparable to or even better than human translators.
They allow easy translation of messages, emails, and even speech now, providing more than enough accuracy for everyday use. This development will definitely continue, bringing progress to many areas.
There are many other examples of how machine learning and AI are improving our lives; we should have another talk to discuss them all.

And what about teaching routine automatization?

AI can help you create a draft of one slide or transform a short course description into a nice-looking syllabus
But the key content of the course is up to you
Currently, it cannot create interesting, well-thought-out, and detailed homework assignments
It cannot check the students' code and provide valuable feedback
Maybe the only way of using such technology in teaching is to create/use a chatbot that will answer the most common questions from students

So it will not replace a real teacher, but it can help to save some of your time

Motivation for teaching (and studying) ML

AI now is third literacy, after classical literacy and programming
Machine learning is becoming more and more common everywhere, including in cutting-edge science and business
Since 2024, there is even International Olympiad in AI for school students: IOAI
The opportunity to deeply understand the material that you are teaching 😁
And, of course, it is fun to teach and study ML!

Now I suggest discussing a very important topic, maybe the most important =) It is motivation for education in Machine Learning.

So this is the first reason. In my opinion, Artificial Intelligence NOW should be considered as the third form of literacy, aligning itself right after classical literacy and programming skills.

MOREOVER, machine learning disrupts various fields at an impressive pace. Particularly in the areas of cutting-edge science and business.

Also, since 2024, the significance of AI has been underscored even more by the start of the International Olympiad in AI for school students. To learn more about this nice initiative, you can visit their website shown on this slide.

Another possible REASON, fairly speaking, one of the main ones for me, is the opportunity to deeply understand the material that you are telling.

Finally, it's worth mentioning that the process of teaching and studying Machine Learning is enjoyable and fascinating!

3-5 min break for the questions and then..

1. Some examples of applying AI to modern problems

2. Practical aspects of creating ML course

Now let's pause for a couple of minutes here before the second part of the talk. You are free to ask me any questions about the first part, of course.
What is the best way to prepare a team of schoolchildren for a performance at an international AI competition? Thank you for the question. Firstly, you need to read the competition rules, criteria, and deadlines. This is an ICPC-style competition, where a team of four school students get two laptops and 8 hours for three problems (in computer vision, natural language processing, and machine learning). Secondly, you need to build foundational knowledge. Organize sessions on AI basics, machine learning, and relevant technologies. At JetBrains, we launch online club Youth AI for school students, I will be glad to see you there and share the information in the end. Thirdly, you need to simulate the competition environment. Try to mimic the actual competition setting during practice sessions.
So now let's switch gears to discuss the practical aspects of teaching machine learning.

Creating a Machine Learning course from scratch

🛠 ⚡️ 🔬 🔄 🇨🇾

So, creating a Machine Learning COURSE from scratch is a challenging task. I tried to highlight some of the key aspects that are most clear and understandable for us.
The RAPID development we see in many fields is also happening in machine learning.
If you take a modern course in BIOLOGY, for example, you'll see that about one third of the material changes each year because science and progress keep advancing.
So we need to provide students with the most up-to-date knowledge, which means frequently UPDATING our courses. For sure, this is also true for machine learning.
Last year, I had the opportunity to develop a new machine learning course from scratch at Neapolis University. I will talk mostly about this course, but suggest starting with a short survey of other possibilities.

One of the most popular courses in ML is Andrew Ng's course on Coursera. Many modern machine learning educators and researchers have grown up on this course.

ANOTHER course I know is the Open Data Science community's course, which is available on GitHub. It includes well-crafted lecture notes, theory descriptions in PDF format, and practical assignments. Everything is available for free, and nowadays, this is a common practice.
Of course, there are many other resources available, like the interactive TEXTBOOK "Dive into Deep Learning" by the big team of authors, mostly working at an Amazon company.
Also, you can watch and code with VIDEO courses on YouTube. Such resources allow us to build a comprehensive course using a variety of materials.

What is the main difficulty?

What do you need to succeed?

Assemble a strong teaching team with both academic and industry experts
Thought out and resolved yourself assignments
Trust and openness, chat with students
Kaggle-style competitions
Better to record and quickly share video lectures
Interactive and emotional atmosphere, questions for the audience

And how do you think, what is the main difficulty in teaching machine learning? It seems to me that we NEED to combine math theory and programming practice, and this is not an easy task.

To ensure success, the following STEPS are crucial:
1. You NEED to assemble a strong teaching team with both academic and industry experts to ensure students receive high-quality education and practical insights.
2. It IS crucial to have well-thought and resolved assignments. This is essential for mastering machine learning concepts.
3. You NEED to establish trust and openness with students by engaging in regular conversations and addressing their concerns.
4. We NEED to create a competitive learning environment, for example, by incorporating contests like Kaggle competitions to motivate students.
5. Nowadays, we MUST record and promptly upload lectures to accommodate different learning paces and ensure students who miss classes can catch up.
6. We NEED to foster an interactive and emotional atmosphere, even in online settings, by asking questions and encouraging participation.
In conclusion, it's important to maintain an open and accessible teaching environment.

How do we teach Machine Learning at Neapolis University?

Part 1: Fundamentals of Machine Learning

ML tasks types, examples, quality evaluation
Linear models, gradient descent, decision trees
Ensembles, random forest, and gradient boosting
Fully connected neural networks and backpropagation
Building GPT-2 from scratch

Part 2: Advanced Machine Learning

Clustering and EM-algorithm
LLM: tokenization, interpretations, embeddings
Building AlphaZero from scratch
Generative and discriminative models: VAE, GANs, flows, and diffusion
RAGs and multi agents LLMs

And how to address the question about course updates?

In our Machine Learning course at Neapolis University students write their code for different machine learning algorithms to understand the inner details deeply. The course also includes theoretical and practical tasks and culminates in an exam consisting of six problems for three hours.

The whole one-year course consists of two parts. The FIRST part covers the fundamentals of machine learning, including linear models, decision trees, ensembles, neural networks, and even building GPT-2 from scratch.

The SECOND part delves into advanced machine learning topics like clustering, tokenization, building AlphaZero from scratch, and generative and discriminative models like Variational Auto Encoders, Generative Adversarial Networks, and Retrieval Augmented Generation.

To address the question about course UPDATES: indeed, each year, the course needs to be revised by at least one third. While it might be possible to retain the same content in the second year, by the third year, significant changes are almost always necessary.
Our course structure is limited to 13 lectures, so adding new material means removing some of the old. This limitation helps us maintain a manageable course size.

Ok, let's go on.

Assignments and grading

Mastering NumPy and Pandas
Theory of linear models
Machine Learning 101 via projects
Table binary classification on Kaggle
Two-layer neural network from scratch in numpy
Image classification on Kaggle

Implement clustering and EM-algorithm
Implement minLLama
Build AlphaZero from scratch
Build Your Own Jarvis
Final project with blogpost

Grading rule-of-thumb: 50%-70%-90% 👍

Assignments are the crucial part of the course, and we have a variety of them.

In the FIRST part, we have about six or seven assignments. The first one is about mastering NumPy and Pandas. Then we move on to the theory of linear models and studying ML with practical projects.
Also, we have two competitions as a valuable tool for motivation. Public leaderboards are used to encourage a healthy competitive spirit without discouraging students.

In the SECOND part of the course, students have four assignments and a final project. We try to make the assignments as practical and as up-to-date as possible. So students can apply their knowledge to real-world problems, such as implementing a minimal version of LLama or creating your own voice assistant Jarvis to talk with on your own laptop. At the end of the course, students present their final project and write a blog post about it. It is also an important part of education, which can be used as an advertisement for the student.

To address the issue of assessing criteria: at the beginning of a course we set clear and fixed points for each assignment in advance. Generally, achieving a passing grade is relatively straightforward, requiring about HALF of the total points. This ensures that students who complete the assignments diligently, can pass, while higher grades require more comprehensive work. Sometimes we need to slightly adjust our grading system at the end of the course.

Zoo of platforms

1. JetBrains Marketplace for all code assignments because of IDE

2. Kaggle Community for all competitions

3. Cogniterra for theory and exam tasks and GitHub for final project assignment

My setup for online teaching in 2024

Slides in JavaScript (reveal.js): for animations, video, and interactivity
Formulas and hand-drawn pictures in Notability on iPad
PyCharm and Jupyter notebooks for coding demos

OBS Studio for combining different scenes and streaming it

Now I'd like to share with you my own setup for online teaching in 2024. Let's go through components step by step.

FIRST, as you see here in these slides, I used to make them in JavaScript with Reveal.js framework. This tool is great for creating dynamic and interactive presentations.

NEXT, for formulas and hand-drawn pictures, I use Notability app on iPad. This combination allows for real-time, intuitive drawing and writing, which is particularly useful for explaining complex concepts. Now let me show a live demo of how this works. You see, I can switch between slides and whiteboard with one click.

For CODING demonstrations, I prefer to turn to PyCharm and Jupyter Notebooks inside PyCharm. PyCharm is an excellent IDE for Python, which offers great coding support, autocompletion and debugging tools.

FINALLY, to bring everything together, I prefer using Open Broadcast Software Studio. It is a powerful tool for combining different scenes, such as your slides, whiteboard, and coding environment.
This setup not only enhances the visual quality of your lectures but also allows you to switch between different teaching materials smoothly.

Any complex topic can be explained naturally in this way. And of course I can switch to any material and explain it.

Build a 2024 ML Course from Scratch: Insights & Teaching Strategies with JetBrains

1. Some examples of applying AI to modern problems

2. Practical aspects of creating ML course

Modern machine learning models

And what about teaching routine automatization?

Motivation for teaching (and studying) ML

3-5 min break for the questions and then..

1. Some examples of applying AI to modern problems

2. Practical aspects of creating ML course

Creating a Machine Learning course from scratch

What is the main difficulty?

What do you need to succeed?

How do we teach Machine Learning at Neapolis University?

And how to address the question about course updates?

Assignments and grading

Grading rule-of-thumb: 50%-70%-90% 👍

Zoo of platforms

My setup for online teaching in 2024

https://lp.jetbrains.com/youth-ai-club

🥳🎉📚 Thank you for attention and text me anything! 📚🎉🥳