Neural Manifold Representation

• Neural Manifold Representation is a mathematical and algorithmic framework that discovers and parameterizes low-dimensional, non-Euclidean manifolds underlying high-dimensional neural data.
• It employs manifold-specific kernels and variational inference methods, such as mGPLVM with techniques like ReLie and von Mises–Fisher, to recover latent neural trajectories and tuning functions.
• The approach enables robust model comparison and uncertainty quantification, with applications ranging from analyzing biological neural circuits to advancing geometric deep learning in artificial systems.

A Neural Manifold Representation (NMR) is a mathematical and algorithmic framework for discovering, parameterizing, and analyzing low-dimensional, typically non-Euclidean manifolds underlying high-dimensional neural or neural-network data. NMR seeks to provide both a generative or descriptive model of the neural latent space and practical tools for inference, interpretability, and comparison of representations across biological or artificial systems. The field spans statistical modeling, geometric deep learning, neuroscientific data analysis, and machine learning architectures, with canonical models illustrated by the manifold Gaussian process latent variable model (mGPLVM), among others (Jensen et al., 2020).

1. Mathematical Foundations of Neural Manifold Representation

NMR assumes that observed high-dimensional neural signals (e.g., neural population spike counts, fMRI voxels, calcium imaging data) are supported on or near a low-dimensional latent manifold $\mathcal{M}$, with $\mathcal{M}$ often possessing non-Euclidean geometry (e.g., a circle, torus, sphere, or rotation group). The generative process posits a latent variable $x_t \in \mathcal{M}$ for each "condition" (time $t$, trial, or stimulus), and each neuron's activity $y_{n,t}$ is generated from a smooth tuning function $f_n:\mathcal{M}\to\mathbb{R}$ evaluated at $x_t$, plus Gaussian noise: $$\begin{aligned} x_t &\sim p^{\mathcal{M}}(x)\ f_n &\sim \mathcal{GP}(0, k_n^{\mathcal{M}})\ y_{n,t} &= f_n(x_t) + \varepsilon_{n,t},\quad \varepsilon_{n,t}\sim \mathcal{N}(0,\sigma_n^2) \end{aligned}$$ where $p^{\mathcal{M}}$ is a uniform or Gaussian prior over the manifold (uniform for compact $\mathcal{M}$, Gaussian for Euclidean), and the kernel $k_n^{\mathcal{M}}$ encodes manifold-aware smoothness (Jensen et al., 2020).

The manifold kernels are tailored to the topology:

• For Euclidean $\mathbb{R}^d$: $d(g,g') = \|g-g'\|^2$.
• For a circle $S^1$: $d(g,g') = 2(1-\cos(g-g'))$.
• For an $n$-torus $T^n$: $d(g,g') = 2\sum_{k=1}^n (1-\cos(g_k-g'_k))$.
• For a sphere $S^n$: $d(g,g') = 2(1-g\cdot g')$.
• For $SO(3)$ (3D rotations): $d(g,g')=4\left[1-(g\cdot g')^2\right]$, where $g$ is embedded as a unit quaternion.

2. Variational Inference and Optimization

Direct inference in NMR models is intractable due to non-Euclidean latent spaces and nonlinear mappings. The mGPLVM framework constructs a structured variational posterior $$q_\theta(x_{1:T}) = \prod_{t=1}^T q_{\theta_t}(x_t)$$ over latent trajectories, employing reparameterization techniques appropriate for the latent manifold (e.g., "ReLie" reparameterization for Lie groups, von Mises–Fisher for spheres).

The model is trained by maximizing the evidence lower bound (ELBO): $$\mathcal{L}(\theta) = H[q_\theta(x_{1:T})] + \mathbb{E}_{q_\theta}[\log p^{\mathcal{M}}(x_{1:T})] + \mathbb{E}_{q_\theta}\left[\log p(Y\,|\,x_{1:T})\right]$$ The likelihood term integrates out the Gaussian process tuning functions via sparse GP variational bounds with inducing points placed on $\mathcal{M}$. Gradients of the ELBO with respect to both kernel hyperparameters and variational parameters are estimated using Monte Carlo methods and manifold-aware reparameterization (Jensen et al., 2020).

3. Manifold Topology Selection and Model Comparison

A defining strength of NMR—as instantiated in mGPLVM—is its capacity to distinguish among candidate topologies for $\mathcal{M}$. Separate models with distinct manifold topologies are trained, and their held-out predictive performance or ELBO is compared:

• Cross-validated mean-squared error or negative log-likelihood for held-out neurons or timepoints.
• Estimated marginal likelihood via ELBO or importance-weighted sampling.

The best-performing manifold topology is selected as most compatible with the data. Empirically, manifold-matched models (e.g., torus models on toroidal-simulated data, ring models on fly head-direction circuits) outperform Euclidean models by large margins (Jensen et al., 2020).

4. Empirical Validation and Applications

The manifold latent variable approach is validated on both synthetic and biological datasets:

• Synthetic datasets constructed on $T^1$, $T^2$, $SO(3)$, or $S^3$ are used to demonstrate recovery of latent state and tuning functions up to isometry, with mGPLVM matching ground-truth trajectories and out-performing Euclidean latent variable models.
• Calcium imaging from the Drosophila ellipsoid body, which is a ring attractor, is accurately modeled with a $T^1$ (circular) manifold, recovering the expected topology and showing improved uncertainty quantification when the population activity is ambiguous.
• Mouse anterodorsal thalamic nucleus recordings during natural foraging and REM sleep are shown to contain a circular manifold ($T^1$) correlating highly with tracked head direction, even without behavioral labels (Jensen et al., 2020).

Key use cases include representing grid cell activity (tori), orientation cell populations ($SO(3)$, spheres), conceptual or abstract variables with continuous state spaces, and latent dynamics in artificial neural agents.

5. Non-Euclidean and Probabilistic Extensions

NMR generalizes beyond classical latent variable models—such as the Gaussian process latent variable model (GPLVM), which is restricted to Euclidean latent spaces—by explicitly modeling non-Euclidean manifold structure through:

• Manifold-specific kernels respecting global topology.
• Non-Euclidean variational approximations (ReLie, von Mises–Fisher) for probabilistic inference.
• Scalable sparse GP inference frameworks compatible with arbitrary manifold structure.

Such generalization is essential for neural systems where ground truth dictates nontrivial topology (e.g., rings for head direction, tori for grid cells), and enables unsupervised discovery and statistical quantification of uncertainty on the latent topology (Jensen et al., 2020).

6. Implications and Theoretical Significance

The manifold latent variable approach to NMR provides:

• A unified Bayesian formalism for uncovering low-dimensional, possibly nonlinear and non-Euclidean neural representations.
• A methodology for automatic selection and quantification of latent topology from data.
• Robust quantification of neural code geometry, enabling neuroscientists and machine learning practitioners to let the data "speak" about underlying structure, as opposed to imposing Euclidean assumptions a priori.

By coupling manifold-aware kernels, manifold-constrained variational inference, and scalable computation, this framework advances the state-of-the-art in both neuroscientific and general latent variable modeling, making it possible to resolve geometric and topological properties previously inaccessible to classical approaches (Jensen et al., 2020).