Study type: In vitro · Status: Verified against declared source
Biologically grounded neocortex computational primitives implemented on neuromorphic hardware improve vision transformer performance.
Proceedings of the National Academy of Sciences of the United States of America · 2025
Study scale: We extended the implementation to fully spiking dynamics by mapping a 10-neuron excitatory TN population with shared inhibition.
Abstract only: Open source record
Product or molecular entity relationships
- VIP: Exact entity relationship. Legacy citation custody associates this source with the catalog record; no product-relevance conclusion is implied.
Plain-language verified summary
Question
Importantly, sWTA dynamics align naturally with the constraints and objectives of neuromorphic engineering, which seeks to emulate brain-like computation in energy-efficient hardware (22–24).
Methods
To implement the sWTA motif, a canonical cortical computation that enables selective amplification while preserving representational stability, on IBM TN, we developed an automated parameter-mapping algorithm that translates biophysical model properties into TN’s discrete neuromorphic space (SI Appendix, Algorithm 2).
Scale or participants
We extended the implementation to fully spiking dynamics by mapping a 10-neuron excitatory TN population with shared inhibition.
Key findings
We implement a biologically grounded cortical circuit motif in neuromorphic hardware and AI architectures to show how experimentally informed neocortical computations, realized through cell-type-specific soft winner-take-all (sWTA) dynamics, can enhance AI.
Limitations and uncertainty
Second, we implemented this circuit motif on IBM’s TN neuromorphic chip using a gain-matching framework to translate continuous biophysical dynamics into discrete hardware parameters.
Verified against declared source. Verification is limited to the declared source and review scope. It does not mean independent replication or establish efficacy, safety, or suitability.