Every field has a moment where the story splits into "before" and "after." For deep learning, that moment has a year attached to it: 2012. This is the first post in a series where I'll be working through my Deep Learning course notes and turning them into something more digestible — starting at the very beginning, with the question of why this field exploded when it did.
The problem nobody could crack
Picture the state of computer vision before 2012. Researchers had a benchmark called ImageNet — a database of roughly 14 million images, organized into about 20,000 categories. A subset of this became the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), which asked systems to sort images into one of 1,000 classes, based on nothing but images scraped from the internet, each carrying a single label.
At the time, classifying images into a thousand categories wasn't just hard — it was considered close to impossible. Error rates on the challenge had been stuck around 25% (measured as "Top-5 error," meaning the correct label had to appear among a model's top five guesses) for years. Progress had stalled. Nobody had a clear path forward.
Enter AlexNet
In 2012, a team led by Alex Krizhevsky entered the competition with something different: a convolutional neural network (CNN). Instead of relying on hand-engineered rules for what to look for in an image, the network learned its own representations directly from the pixels.
The result nearly halved the error rate in a single year. This wasn't an incremental improvement — it was the kind of jump that made the rest of the field stop and pay attention. And it kept going: in the years that followed, ILSVRC error rates continued to drop, eventually approaching — and some claimed surpassing — human-level performance.
That claim is worth pausing on, though. "Superhuman performance" sounds impressive, but how many humans had actually gone through the entire test set to establish a real baseline? Barely any. One researcher, Andrej Karpathy, famously did sit down and manually label the whole test set himself — which led to the joke that what these systems achieved wasn't quite "superhuman," but "super-Karpathy-an." It's a good reminder to look closely at benchmark claims rather than taking headline numbers at face value.
It's also worth noting ImageNet wasn't a perfect benchmark. Some images were genuinely ambiguous — a photo labeled "cherry" that also happens to show a dog, for instance. When a dataset only allows one label per image, it inevitably runs into cases where reality doesn't fit neatly into a single box.
Why a GPU company became one of the most valuable companies in the world
Here's a connection that isn't obvious at first: why did NVIDIA's stock price start climbing around the same time deep learning took off?
The answer is compute. Training neural networks means doing enormous numbers of matrix multiplications, and GPUs — originally built to render graphics — turned out to be extremely good at exactly that kind of math. As deep learning adoption grew, so did demand for GPU hardware.
But the story isn't purely a straight line. There's a noticeable dip in NVIDIA's stock around 2018–2019, and deep learning demand alone doesn't explain it. Around the same time, Bitcoin's value dropped sharply, and cryptocurrency mining had also been a major driver of GPU demand. So NVIDIA's rise reflects two overlapping trends — AI compute and crypto mining — not deep learning in isolation. It's a useful reminder that market signals are rarely caused by just one thing, even when the more exciting explanation is tempting.
Deep learning leaves the lab
Once the ILSVRC breakthrough proved CNNs worked, adoption spread fast. A few examples from the era:
Netflix — the Netflix Prize, a $1 million challenge to build a better recommendation engine, was partly solved using deep learning techniques.
Siemens and GE — healthcare imaging and diagnostics.
Daimler and other automakers — the push toward autonomous driving.
Google, Microsoft, IBM, Apple, Samsung — deep learning woven into core products across the board.
This is the shift from "interesting research result" to "technology reshaping industries" — and it happened remarkably quickly after 2012.
A different kind of proof: games
Around the same time, deep learning was also proving itself in a very different arena: games.
Chess had already fallen to computers back in 1997, when Deep Blue beat Garry Kasparov. But chess is, in a sense, a more tractable problem — engines could lean on a database of known opening moves, brute-force search through the middlegame, and another database for endgames.
Go is a different beast entirely. On a 19×19 board, a player can place a stone on almost any open point on any turn. That means the number of possible game states explodes far faster than in chess — so fast that even today's compute power can't brute-force it. Go required something smarter than search.
That "something smarter" arrived in 2016, when AlphaGo beat a professional Go player for the first time. A year later, AlphaGo Zero surpassed every human player — having learned entirely through self-play, without any human game data at all. Then AlphaZero generalized the same approach to other board games, and by 2019, AlphaStar was beating professional players at StarCraft, a real-time strategy game with far messier, less discrete decision-making than Go.
Why this history actually matters
It's tempting to treat this kind of timeline as trivia — dates and milestones to memorize for an exam. But there's a real reason to understand it before diving into the technical machinery of neural networks: it tells you what problem deep learning was actually built to solve.
The throughline across ImageNet, AlexNet, and AlphaGo is the same: traditional approaches relied on humans encoding the rules or features by hand, and that approach hit a ceiling. What changed in 2012 — and what will show up again and again as we get into convolutional layers, architectures, and training techniques — is systems learning their own representations directly from data, at a scale humans never could have hand-engineered.
That's the thread I'll be pulling on for the rest of this series. Next up: what's actually happening inside a neural network when it "learns" a representation.
This post is part of a series working through my Deep Learning coursework (FAU Erlangen-Nürnberg). Notes are adapted from lecture materials by the FAU.
Top comments (0)