AI 101 — What is deep learning?

Deep learning is the branch of machine learning where a model stacks many layers of simple math, and each layer learns to turn the output of the layer below into something slightly more abstract — so the system builds its own features instead of waiting for a human to spell them out.
Both words in the name are doing work. "Deep" means many layers, not a large amount of knowledge. "Learning" means the layers' internal numbers are set by examples and error correction, not written by a programmer.
Why it matters right now
Every model in the news this month is a deep network. Alibaba's Qwen4-Max, Meituan's LongCat-2.5, Google's Stitch agent skills, the GPT-6 Astra run that helped close a 50-year gap in the Standard Model — all of them are stacks of layers with billions of tuned numbers inside. When a company says "frontier model," "parameters," or "context window," it is describing furniture in a house deep learning built.
This matters for a practical reason: it tells you where the capability comes from and where it doesn't. The model is not a database of facts and it is not a set of rules someone wrote down. It is a feature-building machine, graded a few trillion times until its internal representations got useful — which is why it generalizes to inputs nobody anticipated, and why it fails in ways no one can fully predict.
The mental model
Picture a stack of representations. At the bottom is something raw: pixels, characters, audio samples. Layer one turns raw input into a slightly less raw thing — in an image, patches of light and dark. Layer two combines those into edges and corners. Layer three assembles corners into textures. Somewhere higher, shapes appear; higher still, objects.
Nobody designed the "corner" layer. The only thing humans designed is the top of the stack — the output format and the scoring rule — and the plumbing that sends blame downward when the answer is wrong. That plumbing is backpropagation, and it nudges every layer's numbers a little in the direction that would have made the error smaller. Repeat a few trillion times and the middle layers end up holding representations that look, to a human inspecting them, suspiciously sensible.
Yann LeCun, Yoshua Bengio and Geoffrey Hinton put the definition in a 2015 review in Nature: deep learning lets models composed of multiple processing layers learn representations of data "with multiple levels of abstraction," and a key property is that the features in those layers "are not designed by human engineers" but learned from data, per the review.
That is the actual break with the previous era. For decades, computer vision meant a specialist hand-crafting a feature extractor: edge detectors, corner detectors, histogram tricks, tuned by domain experts over years. Deep learning replaced the craftsperson with a scoring function.
The analogy: an inspection line that grades itself
Think of a factory line with four inspectors and a supervisor at the end.
The first inspector is nearly blind. She can only tell you where the light is brighter than the dark. The second inspector never sees the photo; he only sees her notes, and he reports corners and curves. The third reads his notes and reports textures and shapes. The fourth reports "dog," "not dog."
Now the key part: nobody wrote the rule for "ear-shaped" or "fur-like." Every time the fourth inspector gets it wrong, the supervisor sends a correction back down the line, and each inspector adjusts their personal rulebook by a tiny amount. Run the line over a million graded photos and the inspectors invent a private vocabulary of shapes that works.
Depth is what makes the vocabulary possible. No single inspector can go from pixel brightness to "dog" in one hop — useful concepts sit at different levels of abstraction and need intermediate stops. What changed between 1986 and 2012 was not the idea but the ability to run a very long line without the correction signal fading to nothing on its way back to the first inspector.
Where the numbers come from
Backpropagation arrived in a 1986 paper by David Rumelhart, Hinton and Ronald Williams, according to the ACM's Turing Award citation; Hinton shared the 2018 award with Bengio and LeCun.
The moment the field noticed was September 2012. A University of Toronto entry in the ImageNet image-recognition contest, a network nicknamed AlexNet, cut the top-five error rate to 15.3 percent against 26.2 percent for the runner-up — with 60 million parameters trained on two consumer graphics cards, per the AlexNet paper. Progress on that benchmark had been incremental for years; eleven points arrived in one contest. Within two years the field had switched.
Depth kept climbing after that. Adding layers naively stopped working — the correction signal degraded and accuracy got worse — until skip connections let information jump over layers, which made 100-plus-layer networks trainable, as Microsoft's Kaiming He and colleagues showed in 2015. Today's models go much further, but the shape of the thing is the same one from 2012.
Common misconceptions
"Deep learning means the model knows a lot." No. Depth is architecture. A tiny deep network that recognizes handwritten digits is deep learning; a huge lookup table that stores answers is not.
"It works like a brain." The inspiration is real but thin. Real neurons are vastly more complex, and the training method here — grade the output, nudge the numbers — has no known counterpart in biology at this scale.
"More layers is always better." Depth without the right tricks makes training worse, not better, and a model with more layers than its data can support memorizes instead of generalizing.
"It replaced every other method." On a small table of numbers — sales, prices, sensors — gradient-boosted trees still match or beat deep networks, train in seconds, and are far easier to explain. Deep learning wins when the input is messy and high-dimensional: pixels, audio, language.
Where to learn more
The 2015 Nature review by LeCun, Bengio and Hinton remains the clearest short tour and is free to read. The AlexNet paper is unusually readable for a landmark result — nine pages, and it explains its own tricks. For the layer beneath this one, read how the numbers get tuned in what gradient descent does and what backpropagation is. When you are ready for the design that made depth scale to language, read what a transformer is.
Related reading:
- AI 101 — What is a neural network?
- AI 101 — AI vs machine learning: what's the difference?
- AI 101 — What is overfitting?
If a model's knowledge lives in millions of tuned numbers nobody wrote, is calling it "trained" or "taught" the right word at all? Write to us via our contact page.







