The case for backpropagation as a biologically plausible learning rule
A researcher says searches across thousands of learning rules failed to outperform backpropagation on ImageNet, despite hopes that their approach could improve on it.
TLDR
A computational neuroscience researcher says their work convinced them that backpropagation is biologically plausible and can be implemented through vector error feedback. They say efforts using a library of learning rules from their ICML 2020 paper failed to improve on backpropagation on ImageNet, despite searches across thousands of rules on TPUs. The researcher also cites Francioni et al.’s 2025 Nature paper as confirmation of vector-error feedback.
Combined views
—
1 Source, first seen ago
