Representational geometry of punishment learning across basolateral amygdala neurons

Abstract

Punishment learning allows animals to suppress actions that cause harm while maintaining actions that earn reward. Although the basolateral amygdala is necessary for punishment learning, how its population activity changes across learning, and how this activity reflects an action's consequences, remain unknown. We used longitudinal calcium imaging to track 1,147 basolateral amygdala neurons in male mice across instrumental training, punishment, choice test, and punishment extinction. Punishment learning transformed basolateral amygdala neuron activity associated with the punished action. This activity diverged from that associated with the unpunished action and became aligned with activity evoked by the punisher itself. This alignment was action and contingency specific. It preserved the direction of individual neurons' responses to the punisher: neurons activated by the punisher become more activated by the punished action alone, whereas neurons inhibited by the punisher become more inhibited by the punished action alone. We then asked whether reward and punishment were encoded separately or on a common value dimension across basolateral amygdala neurons. Punishment added a strong shock component to action-related activity without reducing its reward component, while reward and shock themselves occupied largely orthogonal population directions. Thus, during punishment learning, basolateral amygdala population activity can encode opposing consequences of the same action instead of simply changing what that action is worth.