Mesolimbic dopamine release conveys causal associations

Learning to predict rewards based on environmental cues is essential for survival. It is widely believed that animals learn to predict rewards by updating predictions whenever the outcome deviates from expectations. Such violations of predictions are called reward prediction errors (RPEs). RPEs are the critical teaching signal in the most widely accepted model for associative learning—temporal difference reinforcement learning (TDRL). As TDRL RPE has been successful at explaining the activity dynamics of dopamine, it has become the dominant theory of dopamine’s role as the critical regulator of learning.
An alternative approach to learn cue-reward associations is to infer the cause of meaningful outcomes such as rewards. Since causes must precede outcomes, a viable approach to infer whether a cue causes reward is to learn whether the cue consistently precedes reward. This approach is advantageous in the real world where cues often outnumber the meaningful outcomes (e.g., rewards). Using this intuition, here we propose a causal inference algorithm that infers whether a cue is a cause of reward. Based on this algorithm, we denote stimuli (cues or rewards) whose cause should be learned by the animal as “meaningful causal targets” and propose that mesolimbic dopamine signals whether a current event is a meaningful causal target. We found that our model makes similar predictions as RPEs under commonly studied experimental settings.
Hence, to distinguish between the two hypotheses (RPEs or causal associations), we performed experimental tests by measuring dopamine release in nucleus accumbens core. We found that mesolimbic dopamine is consistent with signaling causal associations but not RPE in every case, thereby challenging the dominant theory of reward learning in the brain. Our results provide a new conceptual and biological framework for associative learning.