Forward-Forward Algorithm
A neural network training algorithm proposed by Geoffrey Hinton in 2022 that replaces the conventional forward-plus-backpropagation process with two forward passes. Each layer updates its weights using local signals, eliminating the need for global gradient backpropagation.
Timeline (last 90 days)
The Forward-Forward algorithm replaces one forward pass plus one backward pass with two forward passes: the first runs on positive samples to increase the goodness of each layer, and the second runs on negative samples to decrease goodness
The Forward-Forward algorithm and target propagation are considered superior to standard backpropagation in biological plausibility, because there is no mechanism for transmitting precise backward gradient signals in the brain
Hinton proposed the Forward-Forward Algorithm in 2022, replacing one forward pass plus one backward pass with two forward passes that process positive samples and negative samples respectively.
The Forward-Forward Algorithm currently performs close to backpropagation on simple tasks such as MNIST but still has a significant gap on complex tasks, and is inherently still supervised.
All Facts (4)
The Forward-Forward algorithm replaces one forward pass plus one backward pass with two forward passes: the first runs on positive samples to increase the goodness of each layer, and the second runs on negative samples to decrease goodness
50%UnverifiedThe Forward-Forward algorithm and target propagation are considered superior to standard backpropagation in biological plausibility, because there is no mechanism for transmitting precise backward gradient signals in the brain
50%UnverifiedHinton proposed the Forward-Forward Algorithm in 2022, replacing one forward pass plus one backward pass with two forward passes that process positive samples and negative samples respectively.
50%UnverifiedThe Forward-Forward Algorithm currently performs close to backpropagation on simple tasks such as MNIST but still has a significant gap on complex tasks, and is inherently still supervised.
50%