The goal of this project was to test the theory of predictive processing, the idea that the brain functions like a prediction machine, constantly comparing an internal model of the world against incoming sensory information. Over the five-year duration of this grant, we successfully identified the biological hardware in the mouse brain that performs these computations.
Key Achievements:
Mapping the Prediction Circuit: We identified the specific types of neurons responsible for signaling surprises or errors. These include positive prediction error neurons (which fire when something unexpected happens) and negative prediction error neurons (which fire when an expected event fails to occur). Through advanced microscopy and optogenetics, we mapped how these cells communicate within the local circuitry of the visual cortex (Jordan & Keller, 2020).
Multisensory Integration: We discovered how the brain uses one sense to predict another. Our research showed that with experience, sounds can directly influence visual processing. We identified the inhibitory mechanisms that allow the brain to cancel out predictable visual information based on auditory cues (Garner & Keller, 2022; Solyga & Keller, 2023), providing a potential window into how sensory imbalances might lead to hallucinations.
Genetic Discovery: A major breakthrough was the identification of unique molecular markers for these functional cell types. By linking specific genes to specific neural responses, we developed new viral tools that allow researchers to target and manipulate only the neurons involved in prediction errors (O’Toole et al., 2023). This provides a powerful toolkit for the neuroscience community.
A New Model for Cortical Logic: In the final phase of the project, we investigated the role of deeper brain layers (Layer 5). Our results led us to propose a sophisticated new model of brain function called the Joint Embedding Predictive Architecture (JEPA). This model learns to predict the abstract meaning or essence of an image or event rather than trying to recreate every tiny, irrelevant pixel of detail (Vasilevskaya & Keller, 2026).
Broader Impact:
By defining the exact cells and circuits that maintain our internal model of reality, this work provides a foundational roadmap for understanding neurological conditions. Disorders such as schizophrenia and autism are increasingly viewed as disorders in prediction error computations, where the brain either trusts its internal expectations too much or too little. Our discovery of the genetic and cellular basis for these processes opens new avenues for targeted therapeutic interventions.
