The authors combine dendritic neural networks (structured neuron architectures inspired by biology) with equilibrium propagation, a biologically plausible alternative to backpropagation. Standard equilibrium propagation struggles on harder tasks and deeper networks. They tested dendritic EP on three image classification datasets and found it matches standard EP on simple tasks but significantly outperforms it on harder datasets (KMNIST, Fashion-MNIST) and deeper models. Analysis shows dendritic EP produces higher activation magnitudes and more distributed hidden-state activity.
Main takeaways:
- Dendritic structure improves equilibrium propagation, especially on challenging datasets and deeper architectures
- On KMNIST and Fashion-MNIST, dendritic EP significantly outperforms standard EP and approaches backpropagation performance
- Dendritic EP exhibits higher activation magnitudes and more distributed hidden-state activity during the "free phase"
- Architectural design matters for biologically plausible learning algorithms
- Suggests structured, biology-inspired architectures can enhance learning beyond what backpropagation offers