Research analysis · Learning algorithms

Training without backprop, but not blindly

Zeroth-order methods estimate gradients from output changes under random perturbations, which makes them the natural fit for hardware, and tissue, that cannot run backpropagation. Huebotter and colleagues show that the forward pass already tells you which perturbation directions matter, and using that geometry turns a blind search into a biased, far better aligned one.

Source: An Educated Guess: Deriving Statistically Aligned Gradient Estimates for Zeroth-Order Optimisation, arXiv:2611.08712, 5 November 2026. Primary source. Read the full arXiv HTML text.

What the work claims

The paper introduces Spectral Zeroth-Order estimation, a family of methods (SZO-kappa) that replace the isotropic random directions of classical perturbation learning with directions shaped by each layer's input second moment, the activation geometry that is available for free during the forward pass. The central claim is that this geometry tells you which parameter directions can actually move the network's response on the current data, so perturbing along those directions produces informative loss differences instead of noise. A single parameter, kappa, interpolates between known methods: kappa equals zero recovers classical weight perturbation, any nonzero kappa recovers node perturbation in fully connected layers under per-sample credit assignment, and kappa equals one matches a recently proposed hybrid method, which the authors formalize as node-through-weight perturbation, in expectation under per-batch credit.1

This is a theory paper with supporting experiments: derivations with explicit bias-variance structure, plus gradient-alignment and training studies on a multilayer perceptron and a convolutional network. The empirical claims are about alignment angles and relative accuracy ordering, not state-of-the-art results.

How it works

The curse of dimensionality in zeroth-order optimization is simple to state: if you perturb a weight vector with independent random noise, the perturbation is mostly orthogonal to the true gradient, and the estimate's variance grows with the effective dimension of the search space. Classical fixes (simultaneous perturbation stochastic approximation, weight perturbation, node perturbation) accept that geometry and average over many probes. This paper asks a different question: which directions are even worth probing? The answer it derives from the forward pass is the layer's input covariance. Directions in weight space that align with high-variance input features produce larger and more reliably measurable changes in the layer's output, so the estimator biases its random search toward those directions, in proportion to the feature spectrum raised to a power controlled by kappa.

The trade is explicit and the authors quantify it: the estimation error decomposes into a variance term that scales as the effective number of searched directions over the number of probes, plus a linearization error from the finite-difference step size. Shaping directions buys a variance reduction over weight perturbation that the authors describe as substantial in wide, overparameterized layers, at the price of a bias that is known in closed form because it is built from observable input statistics. That bias can then be whitened away, or kept and treated as a preconditioning direction that adaptive optimizers already know how to handle.

The unification result is the elegant part. The family does not just resemble existing perturbation methods; it contains them. At one end is blind isotropic search; in the middle lies node perturbation, which perturbs activities rather than weights and assigns credit more efficiently; at the other extreme, search concentrates on the dominant input direction. Node-through-weight perturbation, a heuristic hybrid from earlier work, falls out as the kappa-equals-one member under per-batch credit. Several apparently distinct tricks in the perturbation-learning literature turn out to be points on one dial.

Experiments bear the theory out at moderate scale. On a two-layer multilayer perceptron trained on Fashion-MNIST and a six-layer convolutional network (five convolutional layers and a linear readout) trained on CIFAR-10, the spectral estimators improve gradient alignment with the backpropagation gradient over all isotropic baselines. On the convolutional network, SZO-1 leads node perturbation by a gap that grows from 6 to 10 accuracy points as the number of probes per step rises from 1 to 32. Under a compute-matched trade of batch size for extra probes, SZO-1 improves from 45.7% to 48.7% and 49.9% validation accuracy on CIFAR-10, while weight perturbation improves from 37.8% to 40.5% and 41.2%. Backpropagation, for reference, reaches 89.9% on Fashion-MNIST and 83.6% on CIFAR-10 with SGD in the same setup. Early in training, the spectral methods and the hybrid gain 4.5 to 5.7 accuracy points over chance-initialized baselines where node and weight perturbation gain at most 2.3, consistent with their better alignment on convolutional layers.

Where a skeptic should push

The load-bearing assumption is that the input second moment is a stable and accessible quantity. The estimator's bias is only correctable because the geometry is known; in any substrate where the effective input statistics drift, are partially unobservable, or depend on the perturbation itself, the bias term becomes an unknown error rather than a dial. The authors themselves flag where the dimensional advantage weakens: aggregate batch rewards, simultaneous multi-layer perturbation, and extensive weight sharing, such as convolutions, all change the effective ratio of parameters to nodes. A convolutional layer is precisely the regime the CIFAR-10 experiment occupies, and the alignment gains there, while real, are the ones that shrink.

Second, the scale is modest and the ceiling is visible. These are a two-layer perceptron and a six-layer convolutional network on standard benchmarks; backpropagation remains tens of points ahead on CIFAR-10, and the spectral estimators close part of the gap to node perturbation rather than the gap to backpropagation. Whether the geometry shaping survives the depth, recurrence and nonstationarity of real substrates is an open question the paper does not attempt. Demonstrated: better alignment and consistent accuracy ordering at small scale, plus a clean unification. Asserted: suitability for exotic hardware, which is motivation, not evidence.

Training living tissue without backpropagation

Organoid intelligence has a training problem that no amount of better electrodes solves: you cannot backpropagate through living tissue. There is no differentiable graph, no stored intermediate activations, no backward pass. Every serious proposal for improving an organoid's computation reduces to some form of perturbation learning: stimulate, observe the output, nudge the stimulator or readout, repeat. That places the field squarely in zeroth-order territory, and until now it has mostly run the naive version, isotropic stimulation noise plus reward feedback, which is exactly the estimator this paper shows to be the worst member of a family.

The actionable insight is that the forward pass in a tissue experiment already yields the geometry you need. You control the stimulation patterns and you record population activity; the second moment of that driven activity across trials is measurable with the same hardware, no extra experiments required. Shaping perturbation directions by that covariance, the kappa dial, gives a principled way to design stimulation protocols that probe the tissue where it is most responsive, cutting the number of perturbation trials needed to estimate a useful update. In a field where a single training run can take days of culture time and every trial stresses the preparation, variance reduction translates directly into feasibility.

The unification cuts both ways. It means much of what closed-loop organoid work already does unwittingly is a member of this family: patterned stimulation plus population readout is closer to node perturbation than to weight perturbation, and the paper formalizes why that assignment of credit is the efficient choice. The genuine threat is the one the authors state plainly: the dimensional advantage weakens under weight sharing, and biological tissue is the ultimate shared-parameter, nonstationary substrate, where effective input statistics drift with plasticity, fatigue and development. The bias you correct today may be wrong tomorrow. Calibrated expectation: this paper improves the search, not the substrate; if the tissue's useful computations live in directions the stimulation geometry cannot reach, no estimator shapes its way around that.

The bottom line

Established: zeroth-order gradient estimates can be systematically improved, and several known perturbation methods unified, by deriving search directions from forward-pass activation statistics; the improvement is demonstrated in alignment and small-scale training accuracy. Open: performance at depth and on nonstationary, partially observable substrates, which is exactly the organoid regime. What would confirm relevance to biological computing: a closed-loop tissue training study comparing covariance-shaped stimulation against random stimulation at matched trial counts. What would weaken it: evidence that driven tissue activity statistics are too unstable across the training horizon for the shaped bias to stay correctable.

Frequently asked questions

What is zeroth-order optimization?

Methods that estimate how to improve a system using only evaluations of its output, without access to internal gradients: perturb parameters or inputs, observe the loss change, and build a search direction from those differences.

What does SZO-kappa add over classical methods?

Instead of isotropic random perturbations, it shapes search directions using each layer's input second moment, which is available during the forward pass, reducing estimator variance at a known, correctable bias.

How did it perform against baselines?

On a CIFAR-10 convolutional network, SZO-1 outperformed node perturbation by a gap growing from 6 to 10 points as probes increased, though backpropagation at 83.6% remained far ahead of all perturbation methods.

Why is this relevant to organoid computing?

Training living neural tissue cannot use backpropagation; it is inherently a perturbation-learning problem, and the paper's recipe for shaping perturbations from measurable activity statistics maps directly onto stimulation-protocol design.

What are the limits for biological substrates?

The method assumes stable, observable input geometry; tissue is nonstationary and only partially observable, and the theoretical dimensional advantage weakens under parameter sharing, so gains must be demonstrated empirically in closed-loop tissue experiments.

References

  1. J. Huebotter, S. Thill, M. van Gerven and N. Ahmad. An Educated Guess: Deriving Statistically Aligned Gradient Estimates for Zeroth-Order Optimisation. arXiv:2611.08712. 2026. https://arxiv.org/abs/2611.08712. Accessed 2026-11-09.