Networks of spiking neurons can represent uncertainty by sampling. Each moment's spike pattern is one guess about the hidden state of the world, and the guesses visited over time trace out a probability distribution. These networks mix slowly when the distribution is stretched or strongly correlated. Natural gradient enables fast sampling in spiking neural networks showed that reshaping the dynamics by the right geometry fixes this: the network moves along a metric given by the inverse Fisher information, which "rounds out" narrow valleys. That paper fixes the metric in advance and derives it only for Gaussian targets. For real posteriors, which curve differently in different places, the right metric changes with position, and no one has said how a network could compute it locally.
We propose to take the metric from an optimization method that already builds local, well-behaved curvature estimates on the fly. Successive convex approximation (SCA) is an optimizer that replaces a hard non-convex objective, at each step, with a convex "surrogate" that matches the objective's value and slope at the current point and lies above it everywhere. Our claim is that the curvature of this surrogate is a sound, position-dependent preconditioner for a spiking sampler. It is positive-definite by construction, it can be built from first-order information, and it bounds the true curvature so that step sizes stay stable. With the noise switched off, the resulting sampler reduces exactly to an SCA optimizer.
Core Claim
Let $U(z) = -\log p(z \mid x)$ be the energy of the posterior over latent state $z$ given input $x$. Around the network's current estimate $z_k$ we build a convex quadratic surrogate
$$\tilde U_k(z) = U(z_k) + \nabla U(z_k)^\top (z - z_k) + \tfrac{1}{2}(z - z_k)^\top B_k (z - z_k),$$
where $B_k$ is a positive-definite matrix chosen so that $\tilde U_k \ge U$ near $z_k$. This is the majorization condition that SCA requires. The sampler then runs Langevin-type dynamics preconditioned by $B_k^{-1}$, meaning a drift down the gradient plus injected noise, both rescaled by the local geometry. It includes the correction term that position-dependent preconditioning needs to keep the target distribution exact, and in the spiking version this correction takes the form of a Metropolis–Hastings acceptance rule.
We make three specific claims. First, when the target is Gaussian and $B_k$ is exact, this reduces to the natural-gradient sampler of Masset et al. Second, in the zero-noise limit the drift is an SCA iteration, so it inherits the monotone decrease $U(z_{k+1}) \le \tilde U_k(z_{k+1}) \le U(z_k)$. This matches Masset et al.'s finding that efficient balanced networks, a class of spiking networks that track a signal with as few spikes as possible, are the deterministic greedy limit of their sampler. Third, and this is the part that is actually new, $B_k$ can be learned online from gradient differences along the network's own trajectory. The update is a local outer-product rule projected onto positive-definite matrices, and it gives fast mixing on non-Gaussian, position-dependently curved targets where any single fixed metric fails.
The idea differs from the closest SCA work in purpose. Stochastic Training of Neural Networks via Successive Convex Approximations. linearizes a network's output inside a convex loss to get strongly convex surrogates, but uses them only to find one weight vector. We reuse its composite construction, which keeps the convex outer loss exact and linearizes the non-convex inner map, and use it to shape a sampler instead. Escaping Saddle Points with the Successive Convex Approximation Algorithm adds isotropic noise to SCA only when the gradient becomes small, and only to escape saddle points. Here noise is always present and is shaped by the surrogate, because the goal is a correct distribution rather than a better stationary point.
Research Program
The work has three layers.
The first layer is an algorithm at the level of firing rates: surrogate-preconditioned Langevin with a learned $B_k$. We study two ways to build $B_k$. The generic one is a secant update: pairs of recent position changes and gradient changes update $B_k$ by a rank-one outer product, and the result is clipped to lie between a small floor and a local Lipschitz bound. The structured one applies to posteriors that are a convex likelihood of a nonlinear map, such as Bayesian regression with a small network. There $B_k = J_k^\top \nabla^2 \ell \, J_k + \tau I$, where $J_k$ is the Jacobian of the inner map, $\nabla^2 \ell$ is the curvature of the convex loss, and $\tau$ is a small damping term. This is exactly the curvature of the Scardapane–Di Lorenzo surrogate.
We will prove that the continuous-time dynamics leave the target invariant when the correction term is included. We will also prove a mixing bound in terms of how tightly $B_k$ sandwiches the true curvature, adapting the "inexact gradient step" argument that Song et al. use for P-SCA.
The second layer maps the algorithm onto spiking networks. We extend the spike coding construction of Masset et al. so that the effective recurrent weights, which set the metric, are modulated by $B_k$, and we implement the $B_k$ update as a slow plasticity rule driven by products of prediction errors. We keep the two-timescale structure that Combining Sampling Methods with Attractor Dynamics in Spiking Models of Head-Direction Systems found necessary: fast connections keep excitation and inhibition balanced, and slow connections integrate and learn. We will report honestly which parts need non-local information. The Jacobian in the structured variant is the prime suspect.
The third layer is the constrained case, handled only as an application. For targets on non-convex sets such as the head-direction circle, the SCA habit of convexifying the constraint around the current point gives a local feasible region. We compare this against the soft quadratic penalty used by Pjanovic et al.
Baselines
The do-nothing reference is the Euclidean spiking sampler with a tuned step size. Masset et al. report that its Wasserstein error explodes as correlation rises and at dimension 64. Any method should beat it easily, and we include it to calibrate the benchmarks.
The strongest existing spiking method is the fixed natural-gradient sampler of Natural gradient enables fast sampling in spiking neural networks. On equicorrelated Gaussians it estimates mean and variance within approximately 50 ms, keeps stable firing and tracking at $\rho = 0.99$, and scales to $n_p = 64$. On non-Gaussian targets we give it its most favorable fixed metric, the inverse Hessian at the posterior mode (a Laplace metric), computed offline. Our method must match it on Gaussians, where it is optimal, and beat it on the other targets to justify itself.
The strongest non-spiking competitor is Riemannian-manifold Langevin with the exact Hessian made positive by eigenvalue clipping. It is the digital gold standard for position-dependent geometry.
Two further baselines isolate the key ingredient. Adaptive covariance MCMC learns one global preconditioner from the running covariance of the samples, a rule that can be written in Hebbian form. The second is an ablation of our own method in which $B_k$ is replaced by the raw, possibly indefinite secant estimate without the majorization bound.
All methods get the same targets, the same budget of gradient evaluations, and the same simulated neural time. Every method gets the same step-size and learning-rate search, a fixed grid of 20 configurations each. Beating a baseline means a higher effective sample size per gradient evaluation and per 100 ms of network time, together with a lower final Wasserstein distance, across seeds and with non-overlapping intervals.
Why not just…?
Why not just learn a global covariance online, as adaptive MCMC does? For Gaussian targets the sample covariance equals the inverse Fisher metric, so this simple rule may match us exactly. This is a real threat, and the Gaussian benchmark will likely show a tie. The proposal stands only if position-dependent curvature matters. The funnel-shaped, banana-shaped, and Bayesian-network posteriors below are chosen because a single global metric is provably wrong in some region of each. If adaptive covariance matches us there too, the core claim fails.
Why not just use the exact Hessian with eigenvalue clipping? On digital hardware this may mix as well as or better than we do. It needs second derivatives and an eigendecomposition at every step, neither of which a network of local synapses can compute. We therefore compare at equal gradient-evaluation budgets and report the gap. A small gap argues that surrogate curvature is a cheap, local stand-in, and a large gap limits the contribution to plausibility rather than performance.
Why not just use the fixed metric at the mode? When the posterior is close to Laplace-shaped this is enough. Masset et al.'s results imply it will do well on Gaussians. Where curvature changes by orders of magnitude, as in the neck of a funnel, it will take unstable steps in some regions and tiny steps in others. The Wasserstein distance measured per region of the funnel will show whether this happens.
Why not just enforce the circle constraint with a soft penalty? Pjanovic et al. note that the activity bump wanders off the manifold as input noise grows. We measure off-manifold deviation against noise level for both methods. If the penalty holds up at the noise levels that matter for behavior, we drop the constrained application.
Proposed Experiments
The first check is replication. On Masset et al.'s equicorrelated Gaussians, with $\rho$ up to 0.99 and $n_p$ up to 64, the learned $B_k$ must converge to the true Fisher metric and match the reported 50 ms tracking time. Failing this invalidates the construction.
The decisive check uses non-Gaussian synthetic targets: Neal's funnel, a banana-shaped density, and a two-mode mixture. We found no reported spiking-sampler result on Lacuna for any of these. We measure Wasserstein distance, effective sample size, and spike count as an energy proxy against all baselines. The ablation that removes the majorization bound should show instability or bias in regions of negative curvature. That would demonstrate that the SCA property itself, and not just adaptivity, carries the benefit.
On real data, we sample posteriors over the weights of small networks trained on the UCI regression sets CASP, Parkinsons, SkillCraft1, and Wine, which Stochastic Training of Neural Networks via Successive Convex Approximations. used for optimization with no numerical scores reported. We found no reported sampling result on Lacuna for these datasets. We use the structured Gauss–Newton surrogate and evaluate held-out log-likelihood and calibration.
For dynamic inference, we run the stochastic-volatility hidden Markov model tasks of Neurons as Monte Carlo Samplers: Bayesian Inference and Learning in Spiking Networks and test whether surrogate preconditioning speeds up filtering when the posterior's shape changes over time. Finally, the head-direction circle task tests the constrained variant against the soft penalty.
References
- Natural gradient enables fast sampling in spiking neural networks
- Stochastic Training of Neural Networks via Successive Convex Approximations.
- Escaping Saddle Points with the Successive Convex Approximation Algorithm
- Combining Sampling Methods with Attractor Dynamics in Spiking Models of Head-Direction Systems
- Neurons as Monte Carlo Samplers: Bayesian Inference and Learning in Spiking Networks