Fig. 5. ${\mathcal{F}}^{\mathrm{eHebb}}$ improves learning through fixed feedback pathways.

Meta-training an image classifier network with ${\mathcal{F}}^{\mathrm{eHebb}}$ plasticity rule (eHebb; Eq. (8)) on 5-way classification tasks sampled from EMNIST dataset⁴⁹. a Meta-accuracy and b meta-loss plots for ${\mathcal{F}}^{\mathrm{eHebb}}$ compared to ${\mathcal{F}}^0$ via feedback alignment (FA) and backprop (BP), c alignment angles α_ℓ for modulating signals across the network (ℓ = 1, 2, 3, and 4; see Fig. 1 for ℓ = 0) compared with backprop model. Comparing panels a and c indicates that ${\mathcal{F}}^{\mathrm{eHebb}}$ improves the model’s performance by rendering the modulatory signals to be more backprop-like. d Convergence of the plasticity coefficients Θ = {θ₀, θ₂} using the meta-learning model (Alg. 1). The meta-optimizer starts converging after 200 episodes.