Table 1 Summary of related works.

Backpropagation⁴

Learning using Backpropagation that consists of two passes through a neural network. The forward pass works as a prediction, and the backward pass minimizes the error based on derivatives.

Computationally inefficient; biologically implausible; vanishing gradient issue; requires large amounts of labeled data; sensitive to hyperparameter settings.

Forward-Forward algorithm⁵

Bridging the gap between learning in the human brain and learning in machines, mitigating downsides of backpropagation with greedy multi-layer learning.

Implementation complexity; relatively unexplored.

Self-supervised learning^{2,10,11,12,13,14,15,16,17,18,19}

Utilization of large unlabeled datasets, and learning representations with pseudo-labels, examples include SimCLR, MoCo, BYOL, SwAV, Barlow Twins, and VICReg.

Identical output representation generation problem known as ‘Collapse’.

Layer-wise learning^30,31

Attenuating the drawbacks of backpropagation.

Requires careful choice and management of layer-wise loss that leads to generalization.