perception
Find every face
It is 2000, and consumer digital cameras are about to ship with a promise the hardware can barely back: find the faces before the shutter clicks. Detect all faces in family photos — any pose, any lighting, grandmother in shadow, toddler mid-turn — fast enough to run many times a second on a camera's embarrassment of a processor. The design decision is the battle: hand-engineered features composed into a cascade that rejects easy negatives in microseconds, or learned features that need training data and compute the year 2000 does not have in a camera. Justify the choice against the budget. The stakes are the first mass deployment of computer vision into a billion pockets — and every missed dark-skinned face is a bias shipped at consumer scale.
Who this problem belongs to
The two figures whose methods fit it best, out of 58 in contention.
Kanade wrote the first doctoral thesis on face detection and recognition (Kyoto, 1973), running actual pixel programs on 1970s hardware, and by 2000 his CMU group's Schneiderman-Kanade detector was among the best published face finders in the world, handling profile views that frontal-only systems missed. He is one of the few people in history who had, in the year the problem is set, a working answer to exactly this question. His career-long habit of engineering vision for real-time constraints — optical flow with Lucas, video systems for the Super Bowl's EyeVision — means the camera's miserable compute budget is a familiar constraint, not a surprise. The statistical-feature approach his group used sits precisely at the problem's hand-engineered/learned design fork. This is his home turf at the exact date.
LeCun's convolutional networks are the learned-features side of this problem's central design fork, and he had already demonstrated them under brutal compute budgets: by the mid-1990s his AT&T conv-nets read a significant fraction of American bank checks on DSP hardware, and Vaillant, Monrocq and LeCun published conv-net face localization in 1994. So in 2000 he could argue, with running systems as evidence, that learned features fit in small processors if the architecture shares weights and the arithmetic is fixed-point. The honest gap: training data for all poses and skin tones was scarce in 2000, and a Viola-Jones-style cascade would beat conv-nets on raw frames-per-second for several years. He would lose the immediate deployment race but correctly predict the field's endgame — and he could build both contenders.
Fought here
In the mind map
The same ideas, as concepts rather than history — in John's ML knowledge map.
58 figures are scored on this problem. Draw it in a battle to see where you land.