Source-linked AI summary
Approximate Bayesian inference as a gauge theory
Biswa Sengupta, Karl Friston
TL;DR
The paper asks whether brain function can be framed as a gauge theory of variational free-energy minimization. It sketches a Schild’s-ladder algorithm for parallel transport on statistical manifolds, and argues that information-geometric curvature may shape neuronal dynamics, plasticity, and attention.
Problem
The paper asks whether central nervous system function can be formulated as a gauge theory, building on the proposed role of free-energy minimization in self-organizing systems.
Source: §1 · p. 1Method
The paper sketches a Schild’s-ladder algorithm for discrete parallel transport on statistical manifolds without requiring knowledge of their tangent structure.
Source: Variational inference on manifolds · p. 4, §4 · p. 5Results
Curvature in information geometry may shape neuronal dynamics and plasticity, while attention may couple neuronal-state motion according to curvature measured by Fisher information.
Source: §4 · pp. 5–6Takeaways & Limitations
A properly formulated gauge theory should in principle provide exact forms of neuronal dynamics and plasticity, potentially revealing familiar phenomena such as attentional gating and adaptive learning rates.
Source: §4 · p. 6The agent cannot calculate sensory-state marginal likelihood directly because this requires marginalizing over the external states that cause sensory input.
Source: Variational free-energy formalism · p. 2
Abstract
from arXiv · showhide
In a published paper [Sengupta, 2016], we have proposed that the brain (and other self-organized biological and artificial systems) can be characterized via the mathematical apparatus of a gauge theory. The picture that emerges from this approach suggests that any biological system (from a neuron to an organism) can be cast as resolving uncertainty about its external milieu, either by changing its internal states or its relationship to the environment. Using formal arguments, we have shown that a gauge theory for neuronal dynamics -- based on approximate Bayesian inference -- has the potential to shed new light on phenomena that have thus far eluded a formal description, such as attention and the link between action and perception. Here, we describe the technical apparatus that enables such a variational inference on manifolds. Particularly, the novel contribution of this paper is an algorithm that utlizes a Schild's ladder for parallel transport of sufficient statistics (means, covariances, etc.) on a statistical manifold.
1. Introduction
The introduction frames gauge symmetry as a way to describe biological self-organization through variational free-energy minimization. For the brain, it casts free-energy reduction as an interaction between internal and external states, mediated by perception and action.
1. Introduction
- A gauge theory describes interactions between physical fields and matter through a Lagrangian invariant under local transformations.Such invariance is called gauge symmetry; the dynamics do not depend on the frame of reference.Source: §1 · p. 1
- Embryogenesis and dendritic-spine self-organization are presented as distinct contexts that can both be formulated as gradient ascent on variational free energy.The introduction then asks whether this framework can sketch a gauge theory of the central nervous system.Source: §1 · p. 1
- At nonequilibrium steady state, a self-organizing system is proposed to minimize variational free energy, maintaining its states and form amid environmental change.This is linked to homeostasis and allostasis.Source: §1 · p. 1
- In the brain-environment framing, local sensory perturbations are countered through action and perception, treated as functionals of the gauge field.The brain supplies internal states, while the environment supplies external states and continuous forces.Source: §1 · p. 1
- Agents suppress free energy by changing sensory input, acting on external states, or modifying internal states through perception.The introduction presents gauge theory as a logical-mathematical formalism for this free-energy minimization.Source: §1 · pp. 1–2
2. Methods
The method formulates inference about hidden external states as minimizing variational free energy over a parameterized distribution, then extends that optimization to a statistical manifold. It uses geodesic descent and parallel transport to combine successive directions, with Schild’s ladder providing a generic transport procedure.
2. Methods
- Variational free energy bounds surprise by approximating hidden external states with a parameterized distribution; minimizing it is equivalent to maximizing model evidence.For a Gaussian distribution, its sufficient statistics are the mean and covariances.Source: Variational free-energy formalism · p. 2
- On a Riemannian manifold, the Fisher-information metric augments free-energy gradient flow, replacing Euclidean precision-weighted prediction errors with asymptotic dispersion and precision-weighted errors.The natural gradient uses this geometry and is invariant under re-parameterization of the approximate distribution.Source: Variational inference on manifolds · pp. 2–3
- The Fisher-scoring metric used in SPM is the score-function variance, whereas this derivation also includes the prior metric, instantiated as the prior Hessian.The paper notes that Fisher-scoring premultiplies the free-energy gradient by the inverse Fisher-information metric.Source: Variational inference on manifolds · p. 3
- Conjugate-gradient descent follows geodesics, but successive directions occupy different tangent spaces and must be parallel-transported before combination.The method uses the Levi-Civita connection; Fletcher–Reeves is the selected update rule, while other conjugate-gradient rules can also be lifted by altering the implicit norm.Source: Variational inference on manifolds · p. 3
- For complicated distributions, the transport procedure uses Riemann exponential and logarithmic maps and a Schild’s ladder to carry sufficient-statistic directions between points.The geodesic is first approximated by a standard projection method; the ladder repeatedly uses a midpoint and a doubled geodesic segment until the destination is reached.Source: Variational inference on manifolds · pp. 3–4, Figure 2 · p. 3
3. Results
The paper frames sensory entropy as a Lagrangian and maps sensory perturbations to forces and gauge fields. On this curved free-energy manifold, information geometry links neuronal descent and precision-weighting to attention.
3. Results
- The gauge-theory mapping treats sensory entropy as the Lagrangian, external sensory stimuli as local forces, and gauge fields as compensators.The free-energy account also frames internal states as a generative model of the immediate environment.Source: Sensory entropy as a Lagrangian · p. 4
- A Riemann gradient replaces Euclidean descent on curved free-energy landscapes by weighting the Euclidean gradient with asymptotic variance.The Fisher information metric accounts for curvature when measuring distance on the manifold.Source: Sensory entropy as a Lagrangian · p. 4
- With a Gaussian posterior approximation, perception and action follow precision-weighted prediction errors, while curved-space flows also reflect dispersion.The resulting precision limit is the Cramér–Rao bound and, in this account, its attainment translates into attention.Source: Sensory entropy as a Lagrangian · p. 4
- In hierarchical predictive coding, bottom-up messages carry prediction errors and top-down messages send predictions that explain away errors at lower levels.The gauge-theory interpretation casts prediction errors as local forces that activate gauge fields.Source: Sensory entropy as a Lagrangian · pp. 4–5
- Attention is interpreted as the precision-weighting of prediction errors that arises from curvature in the information geometry of least-free-energy paths.The paper describes attention as a force manifesting from that curvature.Source: Sensory entropy as a Lagrangian · p. 5
4. Discussion
The discussion presents neuronal activity as trajectories of sufficient statistics on a curved statistical manifold and advances a parallel-transport algorithm for such manifolds. It connects curvature to possible implications for neuronal dynamics, plasticity, and attention.
4. Discussion
- The proposed algorithm computes discrete approximations of parallel transport without requiring knowledge of the manifold’s tangent structure.The authors say this makes it tractable for manifolds without an assumed ambient space.Source: §4 · p. 5
- Neuronal activity is represented as trajectories of sufficient statistics, such as a Gaussian distribution’s mean and variance, on a curved manifold.The gauge-theoretic account treats movement between neuronal activity distributions as a shortest-path problem.Source: §4 · p. 5
- In deterministic variational schemes, manifold coordinates are distribution sufficient statistics; in stochastic Monte Carlo schemes, they are latent causes.The paper identifies the sufficient statistics as quantities such as means and covariances.Source: §4 · p. 5
- Manifold curvature may link neuronal activity to experience-dependent synaptic plasticity and couple neuronal states in attention.The proposed coupling depends explicitly on curvature measured by Fisher information.Source: §4 · pp. 5–6
- Figure 3 shows the SPM toolbox, which primarily uses variational free-energy minimization across neuroscience, nonlinear dynamics, and reinforcement learning problems.The passage describes the figure as a screenshot of the toolbox.Source: Figure 3 · p. 6
- A properly formulated gauge theory is proposed to provide exact forms of neuronal dynamics and plasticity, potentially illuminating familiar phenomena such as attentional gating.These are presented as possible implications, not as demonstrated outcomes in this passage.Source: §4 · p. 6
A. Software
The software section describes variational and stochastic methods available through the SPM toolbox, while noting that parallel transport for variational models is planned for a later release.
A. Software
- SPM provides variational algorithms for nonlinear regression, probabilistic graphical models, and variational reinforcement learning.These methods are available through the statistical parametric mapping toolbox.Source: A. Software · p. 6
- The SPM suite also includes stochastic methods such as Langevin, Manifold, Riemannian, Hamiltonian, and population MCMC.It additionally includes geometric Annealed Importance Sampling algorithms.Source: A. Software · p. 6
- Parallel transport for variational models is planned for a subsequent SPM release.Source: A. Software · p. 6