Nonmonotonic Plasticity: How Memory Retrieval Drives Learning

What are the principles that govern whether neural representations move apart (differentiate) or together (integrate) as a function of learning? According to supervised learning models that are trained to predict outcomes in the world, integration should occur when two stimuli predict the same outcome. Numerous findings support this, but – paradoxically – some recent fMRI studies have found that pairing different stimuli with the same associate causes differentiation, not integration. To explain these and related findings, we argue that supervised learning needs to be supplemented with unsupervised learning that is driven by spreading activation in a U-shaped way, such that inactive memories are not modified, moderate activation of memories causes weakening (leading to differentiation), and higher activation causes strengthening (leading to integration).