Layers stacked deep
Deep learning is a machine learning method that stacks neural networks — structures loosely modeled on the brain’s neurons — many layers deep. The depth is where the name comes from, and essentially all the rapid progress in AI since the 2010s rests on it.
Finding the features by itself
Earlier machine learning required a human to decide what to pay attention to: to recognize a cat, look at the shape of the ears. The breakthrough of deep learning is that it discovers those features automatically from large amounts of data.
Inside the network, layers near the input pick up simple things like lines and colors, and deeper layers assemble them into eyes, ears, and eventually whole faces. That composition is why more layers means more capability.
Image and speech recognition accuracy jumped close to human level as a result. Face unlock on a phone and automatic photo sorting are both downstream of this.
The turning point was 2012, when a team using deep learning won a major image recognition competition by a wide margin and research worldwide converged on the approach almost immediately. That is the usual starting point for the current AI era.
Every headline event since — the Go victories, self-driving, practical voice assistants — has had this technique behind it.
How it leads to generative AI
The LLM inside ChatGPT is built with deep learning too. An LLM is a neural network using an architecture called the Transformer, well suited to text, trained on an enormous corpus.
The stack runs machine learning, then deep learning, then LLMs, then generative AI. You do not need the internals to use the tools, but knowing the lineage makes the news far easier to follow.