TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 16—18 · FACTUAL · DURABLE

2012: the examples won

The ideas were decades old, and enough labelled data and cheap computing arrived around 2012.

The idea

In 2012 a neural network built by three researchers in Toronto — Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton — won the ImageNet competition, a yearly test of sorting over a million photographs into a thousand categories, by a margin that shocked the field. It had been given no rules for recognising cats or cars. It had learnt from the examples, trained on two consumer graphics cards. Within a few years, learning from examples had replaced hand-written rules across vision, speech and language. Most tools on this map descend from that turn.

Why it matters

It explains why all of this happened now. The ideas were decades old. What arrived around 2012 was enough labelled data and enough cheap parallel computing to make them work.

See it in the studio

When a tool tags your site photographs by building type with no rule you ever wrote, you are using a descendant of the 2012 network. When it labels a chhatri as a gazebo, you are seeing what its examples did and did not contain.

Watch for this

"Deep learning was invented in 2012." It was not. Neural networks trained by backpropagation date from the 1980s, and convolutional networks from the late 1980s and 1990s. 2012 is when the method won in public, on a hard benchmark, against everything else.

Try it

Pick an image-recognition tool; your phone's photo search will do. Give it ten photographs of one building type from your town. Count the hits and misses, then ask: was it likely shown buildings like these, or not? You are auditing the examples.

Prove it

Explain what changed in 2012, what did not change, and why a method from the 1980s needed 2012 to win.

How it works

The three ingredients were a large labelled dataset (ImageNet, assembled from 2009), graphics processors built for games that could do the arithmetic of training in parallel, and a network deep enough to use both. Deep networks learn layers of features — edges, textures, parts, objects — from the data itself, which is what made hand-written feature rules obsolete. The same recipe — more data, more compute, deeper networks — ran through the following decade into the generative models of today.

What this idea builds on

What this idea opens up

Sources

Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON