ABSTRACT Three‐factor learning rules extend Hebbian plasticity by introducing a global modulatory signal that triggers local synaptic weight changes and enables reward‐, surprise‐, or novelty‐based learning. Because synaptic weights update only when this third factor is present, adopting it in reinforcement learning provides a biologically plausible implementation that avoids unnecessary updates, enabling more energy‐efficient and faster computing. Implementing three‐factor learning in neuromorphic hardware requires a global signal that increases the weights of multiple, locally activated (flagged) memristive devices, while leaving inactive ones unaffected. Concretely, the combined presence of local and global signals must induce non‐volatile conductance changes that scale with the intensity of the local flag. Here, we demonstrate an optoelectronic memristor technology based on liquid–processed phase–change telluride inks and synthesized via a scalable solution approach that fulfills these requirements. By leveraging plasmonic, photonic, and electronic control, our integrated cells can be electrically, optically, or electro–optically switched between binary or multilevel states, providing optoelectronic synaptic functionalities. In particular, the optical signal broadcast through a silicon waveguide can act as a global factor modulating the electrical conductance of several cells, emulating biological neuromodulators. Our work advances solution–processed chalcogenide memory devices and their integration into memristive circuits capable of implementing bio–inspired learning rules.
Portner et al. (Thu,) studied this question.