This research proposes a framework for generating realistic, multi-scale neural spike trains using hierarchical coupling architectures. It integrates population dynamics with single-neuron spike generation through a multi-level coupling mechanism, introducing a new training objective based on maximum mean discrepancy.
Key findings
Introduces a three-tier hierarchical coupling architecture for neural spike train generation.
Proposes a maximum mean discrepancy training objective for spike train statistics.
Validates the method against benchmarks including the Allen Brain Observatory dataset.
Limitations & open questions
The framework's scalability to very large neural populations is yet to be demonstrated.
Real-time constraints for applications like brain-computer interfaces are not fully addressed.