Transferring active inference to newly encountered prism-shifted environments
Rapid adaptability to sudden changes in environmental contingencies is a hallmark of biological intelligence. However, how animals flexibly adapt to novel environments remains largely unclear. To investigate the brain's generalizability, this work focuses on prism adaptation, a classical visuomotor phenomenon reflecting rapid recalibration of perception-action coupling. Modeling is conducted based on canonical neural networks that instantiate active inference at the circuit level. Namely, we develop canonical neural networks equipped with multiple policy matrices in parallel, enabling the reuse of previously acquired sensorimotor knowledge. We demonstrate that the proposed model recapitulates key features of prism adaptation in an arm-reaching task, including rapid error reduction and after-effects following prism removal. Networks employing policies pretrained for different target positions adapt substantially faster than naive networks that relearn policies from scratch. These results suggest a biologically plausible mechanism by which neural circuits flexibly reuse prior knowledge to support rapid adaptation.
Authors
- Takuya Isomura (ORCID: https://orcid.org/0000-0003-2960-4919)
Institutions
- Kyoto University (JP)
- RIKEN Center for Brain Science (JP)
- RIKEN (JP)
- YKK (Japan) (JP)
Publication Details
- Journal
- Communications Biology
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1038/s42003-026-11152-5
- Citations
- 1
- Primary Topic
- Embodied and Extended Cognition
- Type
- article
- Field-Weighted Citation Impact
- 1.86
Funders
- Japan Agency for Medical Research and Development
- Japan Society for the Promotion of Science
- Japan Science and Technology Agency
- RIKEN
- Core Research for Evolutional Science and Technology