Machine learning using neural networks with many layers. ‘Deep’ simply means lots of layers. This is the approach behind nearly every AI advance of the past fifteen years, from image recognition to speech transcription to chatbots.
Technical view
The underlying ideas date to the 1980s. What changed around 2012 was not theory but three practical things arriving together: enough labelled data, graphics processing units (‘GPUs’) that could do the enormous number of parallel multiplications required, and a handful of training refinements that made deep networks converge reliably. The field has since been driven substantially by scale. Empirical scaling laws show performance improving predictably as model size, data volume and computing power increase together, which is why the industry has invested so heavily in ever larger training runs.
Related
Part of AI Terminology Index, section: How the Machinery Works.