Fig. 4. Meta-learning biologically plausible plasticity rules.

Performance of a classifier network trained with a pool of biologically plausible plasticity rules ${\mathcal{F}}^{\mathrm{pool}}$ (pool) through fixed feedback pathways. a Meta-accuracy and b meta-loss for ${\mathcal{F}}^{\mathrm{pool}}$ compared to ${\mathcal{F}}^0$ via feedback alignment (FA) and backprop (BP). While learning by ${\mathcal{F}}^{\mathrm{pool}}$ initially resembles ${\mathcal{F}}^0$, continued meta-optimization raises accuracy. This increase reflects the discovery of plasticity terms that can improve learning with random feedback pathways to level with the backprop method in the given classification task. c Alignment angle α_ℓ of the teaching signals of ${\mathcal{F}}^{\mathrm{pool}}$ with the ones for backprop for ℓ = 1, 2, 3, and 4. As discussed in Fig. 1, α₅ = 0. d Convergence of the plasticity coefficients Θ = {θ_r∣0 ≤ r ≤ R − 1} with R = 10. Using L1 regularization in meta-loss (Eq. (6)) sparsifies the set of meta-parameters and helps with identifying the most influential plasticity terms in learning (see Supplementary Note 6 for a discussion on alternatives).