Atlas record: https://www.aitimeline.dev/en/events/evt-0004/
Research · October 1986
Placed by the contemporary primary publication. The exact event date is not known; its documented interval appears below.
A method for training multilayer networks was described.
It supplied a practical training rule that later neural-network work could reuse.
The paper reports experiments; it does not resolve every difficulty of deep learning.
Nature · Published October 1986
MIT Press (scan of the chapter hosted on gwern.net) · Published 1986
It is the first rule for training hidden layers, preceding the 1986 paper from the same group.
Both describe training multilayer models by propagating error backwards; the later publication drew the wider attention.
Both build multilayer models; the way layers are selected and trained differs.
The two volumes set out the full connectionist programme; the Nature paper of the same year is only its core.
The network is trained by the back-propagation described a year earlier.
The network is trained by back-propagation.
It shows why the 1986 method does not work at depth.
The evaluation function is a network trained by back-propagation.
A recurrent architecture is trained by the gradient method popularised earlier.
The model is trained by back-propagation.
Both concern training deep networks; this one adds layer-by-layer initialisation before fine-tuning.
Replacing the sigmoid removes the gradient decay inherent in the 1986 method.
The state backed the school that grew around the authors of back-propagation.
Work by the same author that gave a practical way to train deep networks.