In July 2006, Geoffrey Hinton, Simon Osindero, and Yee-Whye Teh published “A Fast Learning Algorithm for Deep Belief Nets” in Neural Computation (volume 18, issue 7, pages 1527 to 1554). The full text, hosted on Hinton’s own University of Toronto page, states the goal directly: to derive “a fast, greedy algorithm that can learn deep, directed belief networks one layer at a time.”
At the time, networks with many layers were considered too hard to train. The paper showed you could train them one layer at a time without supervision to get a good starting point, then fine-tune the whole network. On the standard handwritten-digit benchmark this produced a very strong generative model and better classification than the leading methods of the day.
This paper is widely credited with restarting the deep-learning era. It made “deep” networks practical again and rebuilt the research community’s confidence that depth was worth pursuing, setting the stage for the breakthroughs of the early 2010s. Hinton would go on to share the 2024 Nobel Prize in Physics for foundational work of this kind.