Deep learning is the part of machine learning that uses neural networks with many layers. "Deep" refers to those layers: each one takes what the previous one found and builds something more complex, from edges to shapes and from shapes to "cat". That is how it learns useful patterns from photos, audio or text without defining every feature by hand.
A question from the app. Pick an answer.
What makes deep learning "deep"?
The name is not about understanding. A deep network finds patterns in numbers and can fail on a case only slightly different from the ones it saw.
That is all it means: stacked layers, each building on the one before. In photos, from edges to shapes and from shapes to "cat".
It learns from data. "Deep" describes the layers, not the number of examples.
Picture four dolls, one inside the other. The biggest is artificial intelligence: any attempt to get a machine to do tasks that seem to need intelligence. Inside it is machine learning, which learns from examples instead of hand written rules. Inside that, deep learning, which does it with neural networks that have many layers.
The smallest is an LLM, like the one behind an AI chat: a huge deep network trained on text. Every LLM is deep learning, but not all deep learning is an LLM: face unlock on your phone and automatic captions are deep learning too.
You give a photo to a network trained to recognize cats. The first layer sees only numbers: how light each dot is. The early layers tend to react to edges and blobs; the middle ones to shapes like a pointy ear or an eye; the last ones put it all together and give the answer: cat.
That is the "deep" part: many stacked layers, each building on the one before. It does not mean deeper understanding. Nobody programmed what an ear is: the network found it on its own by adjusting its weights over thousands of photos.
To estimate delivery time with a classic model, you can prepare columns such as distance, time and rainfall. The model learns patterns from that data.
A photo has no obvious columns: which number says "whiskers"? Deep learning takes the raw data and works out on its own what it needs to look at. That is why it shines with photos, voice and text. With a tidy table and few examples, though, a classic model is often enough, costs less and is easier to explain.
Layered neural networks are an idea that goes back decades. What they lacked was food and muscle: huge numbers of examples to learn from and huge amounts of math to train. The internet brought the first, with millions of photos, recordings and texts.
GPUs brought the second. They are the chips built to draw video game graphics, and they do thousands of multiplications at once: exactly what training a network needs. That is why training a big model today takes rooms full of GPUs, and that shows up in the cost of AI.