The expanded theory assigns roles to connections among brain regions that help combine sensory information, movement and learned object models.
The Thousand Brains Theory holds that the neocortex is made of repeated computing units called cortical columns. Each column is proposed to learn models of objects by combining sensory information gathered across movements, while long-range connections help the columns reach agreement.
This extension addresses connections and theoretical problems left out of earlier versions. It argues that the neocortex works as a heterarchy—a network of interacting parts rather than a simple hierarchy—and proposes roles for pathways involving the thalamus and connections that carry information forward and backward. The authors also discuss how these pathways could support models made from smaller parts and convert information from the body’s viewpoint into a more general model of an object.
How brain links may work
The authors propose that the thalamus helps convert information about feature orientation and movement from an egocentric perspective—the viewpoint of the sensors—into an allocentric perspective used by learned models. They also propose that hierarchical feedforward and feedback connections, along with pathways linking the cortex through the thalamus, allow cortical columns to learn compositional models, in which larger objects or structures are built from smaller components.
The letter presents these proposed roles as an extension of the Thousand Brains Theory. It reviews how long-range connections are organized, relates the proposals to findings from anatomy, neurophysiology and behavior, and identifies predictions that could be tested in future experiments.
Evidence and open tests
This is a theoretical letter, not a report of a new experiment. Its claims draw on a review and interpretation of anatomical, neurophysiological and behavioral findings, and it proposes roles for different brain pathways along with predictions for future testing. The abstract does not report new data, performance measurements or experimental confirmation of the proposed mechanisms, so whether these roles accurately describe brain function remains an open question.
// Source
Neural Computation · 2026 · DOI: 10.1162/neco.a.1579
Authors: Jeff Hawkins, Niels Leadholm, Viviane Clay