Ed+AI: How AI can change the way we study

Don’t waste your time skimming this article to decide whether or not you want to read it. Just read this first.

This newsletter has three main clear goals:

1. Show how AI technologies work as I study and develop my skills in the fields.
2. Explore the implications of these technologies in education, from the students’ to companies’ perspectives
3. Share how these implications and uses are already being developed with company examples, interviews with people actively impacting/researching this field, or both at once.

So, if you read this article, what will you get?

1. A logical, non-mathematical understanding of how a neural network learns (what it is, in the first place).
2. Reflection — and maybe some hope — on how AI intelligence can solve one of the biggest problems in traditional education: all students are different, but they all receive the same educational tools.
3. How Khan Academy is exploring this field and trying to make the most of AI in its education platform

If you are interested in AI and education, this article is for you!

Artificial Neural Network (ANN): how does it learn?

ANN is a “machine learning model inspired by the human brain’s neural structure. It is a type of machine learning process called deep learning.” Okay, what does that mean? It means that it is made up of nodes that act similarly to human neurons: they receive information, filter its importance, and pass it forward to result in an action or output. This process is the way your brain, and AIs, learn patterns and figure out how to identify them in new situations and respond to them in the best way.

However, if you don’t have any background knowledge on it, or if you know how to use it but you don’t know the logic behind it, that explanation above is not very helpful. So, let’s go backward and understand the logic first, and then explore how it works in technology.

Imagine this: You are part of a biology research team that needs a method of classifying any given new species that is discovered. I know, this example seems pretty random, but keep reading; it’ll make sense. So, your team needs to look into hundreds of species to find what features make it possible to classify animals into the different existing groups (like mammals, birds, reptiles).

So how do you start? Looking into every species in alphabetical order? Of course not. You:

1. Divide all the species that are already known into the existing groups.
2. Divide the team into sectors; each sector will look into one group (e.g., “the reptiles sector” will learn only about reptiles’ characteristics).
3. Sectors will learn the most meaningful features in each group and where each one of them is located.
4. Test the accuracy of their learnings, testing if the team can correctly classify species they have never seen before.
5. Upgrade their approach and test more!

Now, if I give a random animal picture to your team, each group will look at a different part of the body, identifying the features and saying things like “this feature doesn’t belong to my group” or “this one does,” and discussing until concluding on one, probably correct, classification.

This is an example of how we could logically approach this situation. That’s very similar to what a neural network does. The team is the NN; the “species” are actually data; the sectors are nodes; and the attention they give to different features is “weights.” Read carefully:

1. The NN receives training data and a final goal (In the previous example, the data would be thousands of species, and the final goal is to be able to classify them into the right groups).
2. The NN starts randomly doing the process. How? Imagine all the sectors look randomly at different features in different animals, searching for similarity that could be relevant to the group. Just like that, the nodes attribute random values to the information they receive.
3. By comparing their learning with the right classification examples that already exist, the network is able to identify which features are actually relevant to each sector. By trial and error, the nodes can identify to what information the weight needs to be higher.

I know this is complicated, but think of it this way: If you are in the “mammals team,” the features you are looking for are fur, milk production, body temperature. So you are not actually looking at how many paws the animal has; this information doesn’t have almost no weight on your final classification. That’s what the nodes are doing. They are training themselves to know where to look. Each node looks at a different aspect and gives its contribution to the whole analysis process.

Each sector would divide into layers to save time and energy. For example, the first layer looks at big features, like if it has fur or feathers. If it has feathers, then it’s passed to the next layer that classifies different birds. In this way, the sector that classifies types of mammals should not receive an animal with feathers and doesn’t waste energy with that. The neural network learns it by comparing with the right examples in training data and seeing that no animal with feathers is a mammal.

Some more technical informations:

1. The ANN is divided into layers: The first one receives the information (input), the hiden ones does all this process that I am describing, and the last layer give the classification (output).
2. ANN does not look at each piece of data at one time. Just like the animal groups, it divides the data into groups (called mini-batches, but don’t worry about that) to increase the speed of its learning process.
3. Each node figures out what’s important by comparing its first random classifications to the right answers in training. Then, it adjusts itself to be more accurate through a process called Back Propagation. In other words, each node changes its weights to make its output closer to the correct answer, shaping the hiden layers to match the expected classification.

So, this ANN would be able to classify even animals that it has never seen before, looking at their individual features and measuring the weight each one of them has in the group’s classification.

This is just a brief understanding on the logic behind ANN training, for a more detailed and technical aproach, check this free course, made by the 3Blue1Brown YouTube channel, you are a visual learner. If you prefer to read, I recommend this course by IBM.

ANN layers ilustration

How could this ever be useful in a educational context?

Great question!

Instead of classifying animals, it could classify student by looking and diferent aspects of their personality and learning process

Why? Everyone learns in a diferent way. Our brain development and life experiences influence how we receive and store information, and make logical conections — check out this BBC article (after finishing this one!) . But, even with all the diferences, we can divide students in some main types defined by how they learn.

For instance, consider that some students may prefer writing texts about their studies because they find that organizing information in a written format makes more sense to them. In this scenario, an AI could classify them as individuals with a “linguistic learning style” and suggest written texts and articles to enhance their learning experience. On the other hand, individuals with a more visual learning style might prefer learning through videos or mind maps in other cases.

This technology could be used to identify the student’s learning style and personalize how the information reaches them, maximizing their learning process. It can target individuals’ abilities and difficulties, so that everyone can achieve the same type of knowledge using different paths. And this is not even the best scenario.

Okay, great… so what?

Saying that “Personalized education values a student’s passion by tailoring learning journeys to their unique potential” (Bard generated sentence) is completely meaningless in real life. If I wrote “potato, potato, tomato” instead of that sentence, it would have the exact same impact on you and the issue itself — that is, none. When talking about the potential impact, we have to be as practical as possible.

So, what do we do with a technology like that? Make school easier? So that students don’t have to adapt their own unique brains to a structured system? Well, no. First, it’s not possible for a school to be completely flexible. Second, I personally believe that making school “easier” won’t actually help students. They need to be challenged so that they can make mistakes and grow. AI’s personalized tools can fill another space in the student’s life, keeping the challenge but optimizing growth and learning.

As I see it right now — and with the research (with technology and people) that I am doing, I hope I change my mind and learn an even better way — the ideal personalized AI educational tool should not actually change the school format but complement it with at least one of these things:

• Help you adapt your studying to your learning style.
• Find the best order for you to learn about a subject (e.g., what type of information you can learn faster, so that you should learn it first).
• Transform any type of information given into the format that best fits your way of organizing logical ideas (e.g., bullet points, paragraphs, infographics, images, mind map, etc.).

Imagine this: A 13-year-old boy spends most of his day at school. He pays attention to most classes and takes some notes but doesn’t actually learn anything. Why? He has difficulties learning by the reading and writing method, so he prefers to listen and say out loud what he is learning. When studying for tests, instead of forcing himself into reading the textbook over and over again, he simply uses an AI tool that reads with a human-like voice and intonation, the entire textbook and his own class notes. In this way, he can study in the way that he learns the most, saving time and energy, and maximizing his grades. Sounds great, right? I think that this will be one of the very common ways of studying in a really short time. That’s not the best part.

How is the best way to power education with this technology?

Khan Academy’s Khanmigo is a great example of how this personalized learning machine can be used: as a tutor.

Khanmigo can guide you in a path based on your previous data. Basically, the NN is constantly geting better, and getting trained by your learning patterns.

https://blog.khanacademy.org/learner-khanmigo/

That’s amazing! What can you do with this? Anything from learning a new, cool mathematical concept to how to take over the world. I do not actually know how to approach the future implications of AI tutors like this, as they are growing and getting better. And, as contradictory as it seems, this is the perfect place to end this article — at the point that we don’t know what’s next. Because this leaves room for exploring those questions in other articles, with diferent perspectives and examples.

How should you feel?

Yes, it’s great to have the possibility of learning in the best way possible for you, but after reading this, you should have some questions:

• Is this technology taking up teachers’ jobs?
• What about students that don’t have access to technology? Won’t it make the gap between privileged and underprivileged people bigger?
• How fast are new tools like this growing?
• Should I throw away all my textbooks and find an AI tutor?

I don’t have the answer to any of these questions, not even the last one. That’s why this is the main goal of this newsletter. In the next months, I’ll be interviewing different people in this field and analyzing more deeply how different companies and EdTechs are approaching these questions.

You don’t read all of this just to skip the last part! This is important too:

• I hope you’ve learned something but that you leave this article with more questions than you had before.
• If you like this content, give some claps, share it, and leave a comment!
• Follow me for weekly articles, with the same simple structure but completely new ideas, in a constant attempt of finding good answers and developing better questions.
• Follow me on X for uncomplicated content on different technologies, AI, education, and interviews: @Leticia0anjos
• Connect with me on LinkedIn so you can learn more about me, and I can learn more about you :)

Thank you for reading! See you next week 😊