Transformers Demystified

March 19, 2023
3 min read

What if I told you that the future of AI is here, and it's called Transformers? No, I'm not talking about the Michael Bay movies with the robots that turn into cars.

I'm talking about a type of neural network architecture revolutionizing machine learning. And guess what? It's not just for the tech nerds like me anymore.

So, what's the big deal about Transformers? Why are they the new darling of the AI world? And why should you care?

Let's start with the basics. Remember when we used to rely on Recurrent Neural Networks (RNNs) for language tasks? Yeah, those days are over. RNNs were like the old Nokia phones - they did the job, but they had their limitations. They struggled with large text sequences and were as slow as molasses when they came out. But one of their biggest limitations was their susceptibility to the vanishing/exploding gradient problem. In layperson's terms, this problem refers to the difficulty RNNs have in learning from information many steps back in time. Imagine trying to remember the first move in a chess game when you're already 50 moves in. It's tough, right?

Enter Transformers, the iPhone of the AI world. Introduced in 2017, these bad boys changed the game. They're like RNNs on steroids - they can handle large text sequences, the gradient problem does not plague them, and they're highly parallelizable. That means you can train some really big models, and I mean massively big.

So, how do these Transformers work? It's all about three key concepts: Positional Encodings, Attention, and Self-Attention.

Positional Encodings are like the GPS of the AI world. They tell the model where each word is in a sentence, allowing it to understand word order. Attention, on the other hand, is like the model's ability to focus on every word in the input when deciding the output. And Self-Attention? That's the model's ability to understand a word in the context of the words around it.

Now, you might be thinking, "Okay, Bora, that's all well and good, but what can these Transformers actually do?" Well, let me tell you, they can do a helluva lot.

From text summarization, question answering, and classification to composing music, generating images from text descriptions, and predicting protein structure, Transformers are the Swiss Army knife of the AI world. They're the magic hammer for which everything is, in fact, a nail.

But here's the real kicker: we don't fully grasp what Transformers can actually do. They exhibit emergent abilities that we don't fully understand yet. It's like we've been handed a magic wand, but we're still figuring out all the spells it can cast. These emergent abilities range from multiplication to generating executable computer code to decoding movies based on emojis. The emergence of these abilities is unpredictable, and researchers are now trying to understand why and how they happen. Understanding emergence could reveal answers to deep questions around AI and machine learning in general, like whether complex models are truly doing something new or just getting really good at statistics.

And the best part? You don't need to be a tech whiz to use them. Anyone can get in on the action with products like ChatGPT, DALL-E, and Midjourney.

So, there you have it, folks. Transformers, demystified. They're not just a buzzword; they're the future of AI. And this future isn't on the horizon - it's right here, right now, and it's looking pretty darn exciting.

Get in touch.

©Bora Nikolic 2023

Make something great.