Skip to main content
[Preprint]. 2024 May 31:arXiv:2405.20594v1. [Version 1]

Table 1:

Detailed comparison of algorithms.

BP FA DFA SF KP WM PAL PFA

No need to transport weight sign
No need to transport weight magnitude
No separate feedback weight learning phase
No explicit weight symmetry after training
Accurate approximation to BP (path alignment)
BP-level task performance

BP: backpropagation. FA: feedback alignment. DFA: direct feedback alignment. SF: sign-concordant feedback. KP: Kollen-Pollack algorithm. WM: weight mirror. PAL: phaseless alignment learning. PFA: product feedback alignment.

: It is unclear how the feedback weights in SF can be learned in a biologically plausible way.

: these algorithms reduce, but do not fully eliminate explicit weight symmetry.

: These algorithms significantly outperform FA and DFA, but still underperform compared to BP in more challenging tasks (CIFAR10 for PAL and ImageNet for SF).