Identifying Latent Neural Dynamics with Recognition-Parametrised Gaussian Process Dynamical Systems

Arielle Rosinski, Qianqian Feng, Lior Fox, Maneesh SahaniarXiv 2026 · arXiv

TL;DR

The paper introduces RP-GPdyn, a new framework for uncovering the hidden "rules" (dynamics) governing neural population activity. By moving away from traditional generative models that require strict assumptions about neural noise, RP-GPdyn uses a "recognition-parametrised" approach to learn interpretable, non-linear dynamical systems that are robust to complex, real-world data.

What Problem Does This Paper Address?

Neuroscientists believe the brain computes through the collective "dance" of neurons—a hypothesis called computation through dynamics. To understand this, researchers build models that map high-dimensional neural recordings (like spikes) to a low-dimensional "latent" state space where the actual computation happens.

However, existing models face a major hurdle: they are typically generative. They require the researcher to specify exactly how latent states produce observed data (e.g., assuming neurons fire according to a Poisson distribution). If this assumption is wrong—for instance, if neurons have "memory" like refractory periods or bursting—the model's estimate of the underlying dynamics becomes biased. Furthermore, integrating different types of data (like spikes, calcium imaging, and video) into one generative framework is mathematically cumbersome.

Methods

The authors propose RP-GPdyn (Recognition-Parametrised Gaussian Process dynamics), which combines two powerful ideas:

  1. Recognition-Parametrised Model (RPM): Instead of modeling how latents generate data ($p(\text{data}|\text{latent})$), RPMs focus on the recognition factor ($g(\text{latent}|\text{data})$). This "generation-free" approach makes the model agnostic to the specific noise distribution of the sensors, making it much more robust to misspecified noise models.
  2. Interpretable GP Dynamics: The model uses a Gaussian Process (GP) to learn the transition rules of the latent states. To ensure the results aren't just a "black box," the authors explicitly condition the GP on learned fixed points and Jacobians (local linearizations). This allows researchers to visualize the "flow" of the system, identifying attractors and stability points that have clear biological meaning.

The model architecture is summarized in Figure 1: RP-GPdyn defines a joint model over observations and latents, using a GP for dynamics and recognition factors to link them.

RP-GPdyn defines a joint model over observations and latents, using a GP for dynamics and recognition factors to link them.

Datasets

The model was validated across a range of increasingly complex scenarios:

  • Synthetic Oscillators: The Van der Pol and Damped Duffing oscillators were used to test the model's ability to recover known non-linear phase portraits.
  • Simulated Decision-Making: A spiking "Excitatory-Inhibitory" (EI) network model was used to simulate a perceptual "winner-take-all" task.
  • ALM (Anterior Lateral Motor Cortex): Real neural recordings from mice performing a delayed-response licking task.
  • MC_Maze: A large-scale benchmark dataset (Neural Latents Benchmark) consisting of motor cortex recordings from monkeys performing reaching movements in a virtual maze.

Key Insights and Ideas

  • Robustness to "Messy" Noise: In simulations where neurons exhibited bursting or refractory periods (violating standard Poisson assumptions), RP-GPdyn significantly outperformed traditional generative models (like SING) in recovering the true latent trajectories.
  • Interpretability is Built-In: By parameterizing the GP with fixed points, the model doesn't just predict the next state; it reveals the underlying "skeleton" of the computation, such as the saddle points and stable attractors that drive decision-making (see Figure 3).
  • Multimodal Integration: The framework naturally allows for "behavior-guided training." By adding a second recognition arm for behavioral data (like hand velocity) during training, the model is "nudged" to find latent dynamics that are specifically relevant to behavior, even if that behavioral data is unavailable during testing.

Key Results

  • Superior Prediction: On synthetic data with complex noise, RP-GPdyn achieved an $R^2$ of ~0.98 for trajectory prediction, compared to ~0.63 for generative alternatives (Figure 2).
  • Recovering Biological Structure: In the ALM dataset, the model successfully identified a bistable attractor structure, where different "wells" in the state space corresponded to the animal's choice to lick left or right (Figure 4).
  • Benchmark Performance: On the MC_Maze benchmark, RP-GPdyn produced latent trajectories that were highly predictive of both future neural activity and hand kinematics, even when trained on neural data alone.
Motor cortex results showing that different reference measures and recognition models can capture distinct dynamical features like rotational organization or reach separation.

Limitations and Future Work

  • Autonomous Assumption: The current model assumes "autonomous" dynamics, meaning it doesn't explicitly account for external sensory inputs (like a sudden flash of light) that might "kick" the system into a new state.
  • Fixed Point Selection: While the model can learn fixed points, determining the number of points to include still requires some manual oversight or post-hoc pruning.
  • Scalability: While sparse GPs are efficient, scaling to extremely high-dimensional latent spaces (e.g., hundreds of dimensions) remains a computational challenge common to most GP-based methods.

The authors suggest that future versions could incorporate control-based inference to handle exogenous inputs, potentially making RP-GPdyn a universal tool for diverse recording modalities like calcium imaging.

Summary updated

Abstract

Neural circuits are densely interconnected systems, and many neural computations are thought to be mediated by the dynamical evolution of activity within these recurrent networks, a hypothesis often referred to as "computation through dynamics". Under this view, understanding computation within real neural populations depends on building models of the dynamical rules governing the temporal evolution of recorded high-dimensional neural activity. Most commonly, such models depend on explicit generative parameterisations of (i) the (nonlinear) intrinsic dynamical flow field and stochasticity, (ii) the (nonlinear) mapping from dynamical state to neuronal activity, and (iii) the variability of individual neural responses given that dynamical state. Misspecification of any of these components can bias estimates of dynamics and latent trajectories. In this paper, we circumvent generation-related issues by introducing the Recognition-Parametrised Gaussian Process dynamics (RP-GPdyn) model. RP-GPdyn models the dynamical flow field using a nonparametric Gaussian Process-defined transition function and models the relationship of dynamics to neural activity implicitly using the Recognition-Parametrised Model (RPM) framework. We show that this generation-free approach uncovers meaningful, behaviourally relevant latent variables and dynamics from both synthetic and experimental datasets.
BibTeX
@misc{rosinski2026identifying,
  author = {Arielle Rosinski and Qianqian Feng and Lior Fox and Maneesh Sahani},
  title = {{Identifying Latent Neural Dynamics with Recognition-Parametrised Gaussian Process Dynamical Systems}},
  year = {2026},
  eprint = {2611.17487},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17487},
}
Figure