Cognitron
Fukushima built a multilayer network that organises itself without a teacher: the receptive fields of deeper layers widen on their own.
Why it matters
A deep network trained without labelled data and without back-propagation, five years before the neocognitron and eleven before the 1986 paper.
Learning is local: a connection strengthens when input and output are both active, while competition between neighbouring cells stops them all learning the same thing. The author compares the scheme to visual cortex directly. The cognitron's limitation, sensitivity to a shift in the pattern, is the problem the neocognitron then solved.