Neural population dynamics without an explicit generative model: RP-GPdyn recovers interpretable latent dynamics from synthetic spikes and real recordings
Synopsis
Models of neural population dynamics usually specify three parts explicitly: the nonlinear flow field, the mapping from state to firing, and observation noise, and misspecifying any of them biases estimates of dynamics and latent trajectories. RP-GPdyn replaces the flow field with a Gaussian-process transition function and links latent states to neural activity implicitly through a recognition-parametrised model, avoiding a generative observation model. On synthetic spiking data, a random-dot motion decision simulation, mouse anterior lateral motor cortex delay activity and Neural Latents Benchmark motor-cortex data, it recovered behaviourally relevant latent variables and dynamics.
Interpretation
A neural population dynamics model can drop the explicit generative mapping from latent state to firing and still recover interpretable latent dynamics from recordings. Earlier approaches typically parametrise separately the nonlinear flow field and its stochasticity, the nonlinear mapping from dynamical state to neural activity, and the variability of individual responses given that state; misspecification of any component can bias estimates of dynamics and latent trajectories. On synthetic spike trains generated from a latent Van der Pol oscillator with 50 neurons, 50 training trials and 30 held-out trials, RP-GPdyn recovered the latent trajectories, dynamics and phase portrait, including a fixed point consistent with the unstable spiral; compared with SING, which also uses GP transitions but keeps explicit generative modelling, it was more robust to spike-history effects, with the largest gains in dynamics prediction.
Recognition factors carry the link between latent states and neural activity, so the model need not specify a spike-count or calcium-imaging noise distribution. The recognition-parametrised model builds the joint from the empirical marginal over observations plus parametrised recognition factors, so learning focuses on the recognition factors and prior rather than fitting an observation distribution. On simulated spikes from the Wang random-dot motion decision network, the model was not told that left-selective, right-selective and non-selective subpopulations existed; using only population spike counts of all excitatory neurons, it recovered the bistable dynamics, its latent states predicted instantaneous firing rates of the selective excitatory populations, and trial choice could be predicted from observation prefixes, including ambiguous zero-coherence and low-coherence error trials.
Interpretability of the latent dynamics is retained by conditioning the Gaussian-process flow field on learned fixed points and local Jacobian matrices. Flexible recurrent-network or neural-ODE dynamics can predict neural activity well, but their vector fields, fixed points and bifurcation structure often require substantial post-hoc analysis. On delay-period activity from four fixed-delay sessions in mouse anterior lateral motor cortex, the model supported the previously proposed bistable attractor mechanism, and licking-direction classification from predicted neural activity was on par with SING; on the Neural Latents Benchmark MC_Maze data, inferring from 137 held-in neurons and scoring across 182 neurons, posterior reconstructions achieved good PSTH scores, and autonomous rollouts from inferred states came close to full-posterior reconstructions.
Perspective
The demonstrated scope is synthetic spiking data, a Wang random-dot motion decision-network simulation, mouse anterior lateral motor cortex delay-period recordings and Neural Latents Benchmark MC_Maze motor-cortex data; the authors propose calcium imaging, where observation noise is more complex, as a future direction.
The model assumes autonomous latent dynamics, excluding the influence of exogenous feedforward inputs such as contextual or sensory signals; the authors note that inferring latent dynamics and exogenous inputs together is inherently degenerate and leave it for future work. On the ALM data, despite hinting at a bistable attractor structure, one fixed point was estimated erroneously and consistently across random seeds, which the authors suggest may reflect the scarcity of data (59 neurons and 163 single trials).
