Psychedelic therapy has shown remarkable promise for treating mood and anxiety disorders, inspiring both scientific exploration and theoretical innovation. Despite its growing clinical validation, the mechanisms underlying its therapeutic effects remain under active investigation. This talk synthesizes insights from machine learning theory and computational neuroscience to shed additional light on these mechanisms and present them in a way accessible to a wide audience.First, we explore how psychedelics influence belief structures in the brain through the lens of the CANAL model, which expands on the well-known “RElaxed Beliefs Under pSychedelics” theory. Using tools from deep learning, we delineate two distinct optimization landscapes that underpin belief representation, illustrating how both overly rigid and overly flexible beliefs may contribute to different forms of psychopathology. Our model predicts that psychedelics dynamically re-calibrate these landscapes in ways that are often potentially therapeutic, but may not be helpful in all cases.Next, we present a concrete predictive processing model based on our framework that simulates how serotonergic psychedelics, such as LSD and psilocybin, affect cortical dynamics via receptor-level mechanisms. By modeling the effects of neuromodulation through 5-HT2a and 5-HT1a receptors, we propose computational explanations for their impact on cognition and emotion, and ultimately their psychotherapeutic effects. This work identifies potential avenues for designing novel psychotherapeutics, emphasizing the unique potential of biased agonists like 5-MeO-DMT as future therapeutic tools.By bridging machine learning and neuroscience, this talk offers an accessible yet scientifically rigorous exploration of how abstract computational models can illuminate the profound effects of psychedelics on the mind. It aims to engage scientists, clinicians, and therapists alike in a shared understanding of this rapidly evolving field.
