During auditory fear conditioning, animals learn that an auditory tone predicts an aversive outcome (electric shocks). However, learning finally reaches a steady state at certain memory strength (termed the learning asymptote) beyond which further training is ineffective at producing learning unless the strength of the aversive outcome is increased. We previously found (1) shock-evoked responding in pyramidal neurons in lateral nucleus of amygdala (LA) instructs fear conditioning and strengthens LA neural responses to auditory input, (2) learning dependent reduction in the shock-evoked response in LA (termed prediction error coding). Prediction error coding also occurs in the periaqueductal gray (PAG) which relays aversive shock signals to LA. PAG receives inhibitory input from central nucleus of amygdala (CeA) which is activated by tones after fear learning. Here, we hypothesized that a CeA-PAG pathway provides a negative feedback on PAG to produce prediction error coding in LA and that this neural coding in LA regulates fear learning asymptotes. We first developed a 4-day fear conditioning paradigm in which rats were trained (days 1, 3) and tested (days 2, 4) twice. We found that rats reached learning asymptote after initial training (day 1) as learning was not enhanced by overtraining (day 3). We found that: (1) optogenetic inhibition of CeA-PAG afferents at learning asymptote increased learning levels and that this manipulation also disinhibited predicted shock-evoked responses in LA neurons (i.e. reengaged prediction error coding). (2) optogenetic additive activation of shock-evoked responses in LA pyramidal cells during overtraining increased learning asymptotes. (3) optogenetic inhibition of shock-evoked LA neuronal activity during overtraining abolished the increase in fear learning induced by a higher shock intensity. These results support our hypotheses and suggest a circuit mechanism which may contribute to the production of anxiety disorders which are typified by exaggerated aversive learning.