Learning neural evolution operators: From decoding to identifiable causal state-space models

Area of Science:

• Systems Neuroscience
• Computational Neuroscience
• Neural Dynamics

Background:

• Neural encoding, decoding, and representation learning predict variables from neural activity.
• Dynamical systems model neural computation via evolving latent population states.
• Current frameworks predict observations but struggle to identify unique computational mechanisms.

Purpose of the Study:

• To propose a unifying framework integrating representational models, latent neural dynamics, and perturbation-based validation.
• To address the challenge of observational data providing insufficient constraints for mechanistic validity in neural circuits.
• To shift neuroscience from prediction-oriented to perturbation-constrained mechanistic dynamical approaches.

Main Methods:

• Reviewing advances in learning neural evolution operators (RNNs, latent state-space models).
• Analyzing identifiability of latent dynamical models and their evolution operators.
• Discussing perturbation, intervention, and closed-loop interfaces as causal constraints.

Main Results:

• Neural representations may emerge from latent dynamical processes on low-dimensional manifolds.
• Latent trajectories and predictive performance alone do not guarantee mechanistic validity due to observational equivalence.
• Perturbation-based validation is crucial for falsifying candidate dynamical explanations and reducing ambiguity.

Conclusions:

• Distinguishing mechanistically valid neural dynamics requires more than observational recordings.
• Evolution operators should be treated as experimentally testable hypotheses, not just descriptive models.
• Integrating latent dynamical modeling with perturbation validation enables a mechanistic understanding of neural computation.