Optimal structure of metaplasticity for adaptive learning - Fig 3

<p>(<b>a</b>) Schematic of the reservoirs, buffers, and transient meta-states, and how synapses occupy these meta-states according to reward probability. (<b>b</b>) Plotted are the fractions of synapses in the weak reservoir (dashed lines) and buffer (solid lines) as a function of reward probability. The inset shows the fraction of transient meta-states for <i>N</i> = 6 and <i>N</i> = 8 models. Fractions in the strong reservoir, buffer, and transient meta-states are the mirror image (along <i>p</i>_<i>r</i> = 0.5) of their weak counterparts. As reward probability approaches 0.5, more synapses occupy transient and buffer meta-states, making the model more adaptable. As <i>p</i>_<i>r</i> deviates from 0.5, more synapses transition to reservoirs, enabling the model to protect the signal.