Seonglae ChoGradient Descent through the Denoising loss using Energy Function
Closer to how the cerebral cortex works than Back Propagation, proposed by Geoffrey Hinton
Signals form some kind of loop as they pass through cortical layers before returning to the start of the synapse.
As the number of layers grows, FF uses less memory than backprop. However, in thin networks with fewer than 20 layers, FF uses far more memory than backprop.
Unlike Automatic Differentiation, it does not compute gradients.
Forward-Forward gives up on computing global weight credit exactly, and instead has each layer update its weights from its own local objective.
Forward Forward Algorithm Notion
![[딥러닝] The Forward-Forward Algorithm: Some Preliminary Investigations (FF)](https://static.velog.io/favicons/apple-icon-152x152.png){kind=icon}
![[딥러닝] The Forward-Forward Algorithm: Some Preliminary Investigations (FF)](https://velog.velcdn.com/images/nochesita/post/a5854fae-195f-4f61-aaf4-9b04c6330386/image.png)
