Dopamine, Model-Based Learning, and the Neural Basis of Learning with Calle Machado Costa

Summary

In this episode of Science Quickly, Calle Machado Costa discusses how learning is encoded in the brain, contrasting two main frameworks: model-based learning, which builds a mental map of the world to guide choices, and model-free learning, which updates decisions through trial and error based on reward history. Costa describes the idea of being “slightly cursed” in predictions, which keeps his research exciting by ensuring expectations are continually challenged. He recounts an influential PhD finding using the DAT-Cre mouse line that revealed sex-dependent background phenotypes complicating behavioral results, underscoring the need for careful strain selection in neuroscience experiments.

• Key insight: Dopamine signals can reflect prediction errors beyond mere rewards, suggesting a model-based component to learning.
• Key insight: The orbitofrontal cortex stores and links information about events, and its disruption can produce imprecise models rather than a simple switch to a different learning system.
• Key insight: Computational perspectives illuminate mental illness such as schizophrenia and addiction as disorders of how models of the world are formed and used.
• Key insight: Predictions that turn up unexpected results are a productive part of science, driving new questions and discoveries.

Introduction and scientific setting

The podcast features Calle Machado Costa, an assistant professor of psychology at the University of Alabama at Birmingham and a 2026 Young American Scientist honoree. Costa explains that his career has been driven by a pattern of predictions that did not align with outcomes—a trait he jokingly calls a curse, but one that keeps his research dynamic and informative. He emphasizes the value of starting projects with a clear hypothesis and prediction, noting that even failed predictions provide meaningful data that contribute to a broader understanding of neuroscience.

Two dominant frameworks for learning

The discussion centers on two broad schools of thought about learning. The first, model-free learning, updates decisions by tracking the value of cues or actions based on received outcomes. It is computationally simpler and relies on learning from actual experiences. The second, model-based learning, posits that the brain constructs a rich representation or cognitive map of the external world. This allows inferences about unobserved events and the prediction of future outcomes beyond direct experience, at the cost of greater computational and neural resource demands.

Costa explains that modern neuroscience often casts these ideas in terms of prediction errors: how the brain updates beliefs when outcomes differ from expectations. He updates this framework with a nuanced view that highlights how dopamine signals, traditionally tied to reward prediction errors, may also reflect more complex, model-based prediction errors. In addition, Costa discusses the orbitofrontal cortex (OFC) as a critical region in linking specific events and forming task-related cognitive maps. His work suggests that the OFC does not merely mediate all model-based learning but helps construct precise links between events, with inactivation leading to confused or degraded models rather than a simple switch to model-free strategies.

Key findings and their implications

One pivotal area of Costa’s research challenges the idea that model-based and model-free learning are strictly separate systems. His data indicate that dopamine teaching signals can carry information beyond reward value, aligning more with model-based prediction error signals in certain contexts. This reframes how researchers think about learning signals in the brain and invites a more integrated view of how different neuromodulatory systems interact during learning.

Another important thread concerns the orbitofrontal cortex. In experiments with rodents, transient inactivation of this region disrupted model-based learning, but the effect was not a simple elimination of model-based processing. Instead, animals formed a model that was “confused” or imprecise, suggesting that the OFC contributes to the sophistication and specificity of internal models. Costa argues that a nuanced perspective—where behavior is explained by models of varying complexity and precision rather than a binary dichotomy—may better account for how people learn and why some mental illnesses emerge.

Applications to mental health and addiction

Costa extends his findings to the neuroscience of mental illness, including schizophrenia and addiction. He discusses latent inhibition, a process measuring attentional filtering, and notes that OFC involvement is essential for efficient filtering of irrelevant information. Deficits in latent inhibition are common in schizophrenia, potentially explaining how individuals with the disorder form spurious associations leading to cognitive difficulties and hallucinations. In addiction, the idea shifts away from a simple shift from model-based to model-free control, toward the concept of disordered models of the world, which may help to explain why behavioral interventions like contingency management can be effective even when classic reward-based accounts fall short.

Open questions and future directions

The podcast closes with several directions Costa is excited to pursue. He is investigating the informational content of dopamine teaching signals to determine which dimensions contribute to their information-carrying capacity beyond traditional reward prediction errors. He is also exploring how acetylcholine interacts with dopamine signals to support different aspects of learning, particularly the motivational versus reward-related components. Finally, Costa expresses interest in how environmental factors shape whether neural systems generate highly detailed or more generalized models of the world and how cellular and molecular processes implement these computations in neurons. These topics point toward a broader aim: understanding how learning, cognition, and behavior emerge from the interactions of neural circuits and neuromodulators.

Conclusion

The podcast highlights how a computational lens can illuminate core questions about learning and behavior while offering practical insights into mental health and illness. Costa’s work exemplifies the value of precise, testable hypotheses, even when outcomes defy expectations, and showcases how modern neuroscience seeks to integrate multiple learning theories with dynamic neural signaling to understand the brain’s adaptability in a complex world.