Inverse Models Predict Mirroring Offsets and Explain the Acquisition of New Syntax in a Sequential Song Learning Task

Control-theoretic inverse models are very useful for learning and generating flexible sensorygoal directed motor behaviors. We have recently proposed a simple eligibility-weighted Hebbian learning rule capable of provably forming inverse models in high dimensional linear networks by associating random motor explorations with their future sensory consequences [1]. In this theory the inverse model forms in the synaptic connections from sensory to motor neurons, allowing the conversion of a sensory memory (for example a tutor song template) into the necessary motor pattern required to reproduce the sensory memory. Mirror neurons naturally arise during inverse model learning. Mirroring of motor and sensory evoked activity is either in precise temporal register, reflecting predictive inverses associated with stereotyped motor codes as in HVC, or temporary delayed, reflecting causal inverses associated with variable motor codes as in LMAN.