The credit rule that decides whether tissue could generalize

What the work claims

This is a methods paper about optimisation, and its central claim is that the shape of the surrogate gradient, the smooth stand-in used to train spiking networks, controls not just whether training converges but how well the trained network generalises. The authors introduce a scheme with two parts. An asymmetric surrogate assigns larger gradient to neurons whose membrane potential is higher, that is closer to the firing threshold, and they prove, under stated distributional conditions, that it has a lower coefficient of variation of the local gradients than the symmetric surrogates in common use. An adaptive component adjusts an effective window over space and time to keep gradient directions consistent across timesteps, addressing what they name temporal gradient confusion. On CIFAR-10 a ResNet-19 reaches the mid-96 percent range, with corresponding gains on CIFAR-100, ImageNet and the neuromorphic CIFAR10-DVS set, at small timestep counts.¹

The bold part is the mechanism they claim underneath the numbers: an argued link from local gradient variation to the curvature of the loss landscape, and from curvature to generalisation. Their analysis bounds a curvature measure by a term that grows with local-gradient variation, so lower variation is associated with flatter minima, and flatter minima are in turn associated with better test performance. That turns a training heuristic into an argument about why some spiking networks generalise and others do not.

How it works

A spiking neuron fires when its membrane potential crosses a threshold. That firing is a step function, and a step has a derivative of zero almost everywhere and infinity at the jump, so ordinary backpropagation cannot flow through it. The standard fix is the surrogate gradient: during the backward pass, pretend the step was a smooth curve, typically a symmetric bump centred on the threshold, and use that curve's derivative. Training deep spiking networks is done with spatio-temporal backpropagation, which unrolls the network over its timesteps and backpropagates through all of them, aggregating gradient contributions from every timestep into each weight update. This is backpropagation through time with a surrogate standing in for the spike.

The authors make two observations about that process. First, a symmetric surrogate treats a neuron sitting just below threshold and one sitting far below it with gradients that are mirror images, which they argue injects unnecessary variance into the gradient. Their asymmetric surrogate instead skews credit toward neurons with higher membrane potential, the ones whose state is most informative about whether a spike will occur, and they prove, under a stated condition on the distribution of membrane potentials, that this reduces the coefficient of variation of the local gradients. Second, because each weight update sums contributions across timesteps, gradients that point in inconsistent directions at different timesteps partially cancel and confuse the update; their adaptive scheme applies spatial adaptation at the last timestep, where activations and gradients are relatively stable, and a temporal adaptation that aligns gradients between adjacent timesteps, pulling the earlier timesteps into consistency with that stable reference.

The connective tissue of the paper is the claim that these variance reductions matter because gradient variation is tied to loss-landscape curvature. They present analysis and loss-surface visualisations arguing that lower local gradient variation corresponds to flatter regions of the loss landscape, and they lean on the established finding that flatter minima generalise better. So the story is causal in intent: design the surrogate to lower variance, land in a flatter minimum, generalise better. The empirical gains across static and neuromorphic datasets are offered as confirmation.

Where a skeptic should push

The load-bearing assumption is the chain from surrogate shape to variance to flatness to generalisation. Each link is individually plausible and each is contestable. The flat-minima-equals-generalisation claim is a strong regularity in deep learning but not a theorem; there are known counterexamples where sharpness measures can be manipulated without changing the function. The variance-to-curvature link is argued analytically under assumptions and illustrated on selected layers, which is suggestive rather than conclusive. And the gains, while consistent, are incremental over a crowded field of surrogate-gradient and flat-minima methods; the paper competes at the level of fractions of a percentage point, where architecture, data augmentation and training budget can swamp the effect being measured. The honest reading is that A2SG is a well-motivated refinement with a coherent theoretical story, not a demonstration that surrogate shape is the dominant driver of generalisation.

A second push matters more for the interpretation that follows. The asymmetric rule is often described in biological language, more credit to neurons near threshold, but it is not a local learning rule. It is a factor computed inside backpropagation through time and multiplied into a gradient that has been propagated globally from the output error across all layers and all timesteps. The membrane-potential dependence modulates a nonlocal signal; it does not replace it. Anyone tempted to read the asymmetry as a ready-made biological plasticity rule is reading past what the method actually is.

What flat minima ask of noisy tissue

Start with the deflationary implication, because it is the important one. This paper is, in effect, a detailed inventory of what the dominant spiking-network training paradigm requires, and every item on the list is something living tissue lacks. It needs a differentiable surrogate for the spike, an unrolled computation graph over time, exact per-timestep membrane potentials, a global error signal backpropagated through every layer and every timestep, and fine control over the gradient's statistical shape. A cortical organoid on a microelectrode array supplies none of these. It does not expose a differentiable spike, it cannot run backpropagation through time, and it offers no channel by which a global output error could be delivered to each synapse with the right sign. When the field speaks loosely of training an organoid the way one trains a spiking network, this is the machinery being borrowed, and it does not port.

The non-obvious implication is sharper, but it has to be flagged as an analogy that reaches past what the paper establishes, not as a result the paper proves. The authors argue that generalisation in these networks is shaped by gradient variation, where variation means the inconsistency of the surrogate-gradient estimate across neurons and timesteps, a quantity defined entirely inside backpropagation through time. It is tempting to map that onto living tissue, which is intrinsically high-variance in a quite different sense: trial-to-trial variability, channel noise, ongoing spontaneous activity, and drift. The temptation should be resisted in its strong form, for two reasons. First, these are not the same variance. The paper's is a property of a computed gradient; biology's is a property of neural activity; equating them quietly assumes the very learning rule and loss landscape that tissue does not have. Second, the mapping can invert. The mainstream result on stochastic gradient noise, from small-batch and high-learning-rate training, is that more gradient noise biases learning toward flatter minima and better generalisation, which is the opposite of the naive reading. So the honest statement is a hypothesis, not a prediction: if a defined wetware learning signal turned out to carry the destructive, direction-conflicting kind of variation the paper targets, rather than the flatness-promoting kind, then one would expect trained tissue to land in sharper, worse-generalising minima. That is a claim to test, and the test is concrete, giving a living network a defined closed-loop training signal and checking whether its variability tracks off-distribution failure. Nothing in this paper settles the direction.

There is a possible opportunity hiding in the same construction, and it must be stated as a structural analogy only. The one part of A2SG with a biological echo is the asymmetric surrogate: giving more weight to neurons near threshold is a voltage-dependent gating, and real synapses do something structurally similar, since NMDA-receptor and calcium-dependent plasticity are gated by postsynaptic depolarisation. That raises a question worth pursuing: could a local, voltage-gated plasticity rule that tissue can actually implement bias a living network toward a low-variance, flatter-minimum regime, without any backpropagation at all. The paper provides no evidence for this. Its asymmetry is only a modifier of a globally backpropagated gradient, and nothing in it shows that biological voltage-gated plasticity reduces gradient variation or reaches flat minima. What the paper offers is two ingredients, flatness as a target and voltage-gating as a variance-reducing move, that a wetware learning rule could try to combine. The opportunity is a research direction, nothing more.

The bottom line

Established: shaping the surrogate gradient to favour neurons near threshold lowers gradient variance and, on standard image and neuromorphic benchmarks, improves deep spiking-network accuracy at low timestep counts. Argued but not settled: that this works specifically by flattening the loss landscape, and that flatness is the operative cause of the generalisation gain. For organoid intelligence the durable lessons are two. First, the accuracy of trained spiking networks is inseparable from a training apparatus, backpropagation through time with a differentiable surrogate, that living tissue cannot run, so benchmark results transfer but mechanisms do not. Second, and more tentatively, the paper's variance-to-flatness argument invites a hypothesis rather than a prediction. Whether intrinsically noisy tissue generalises poorly under training depends on a defined learning rule and landscape the paper does not supply, and on which kind of variance a tissue learning signal would carry, since stochastic gradient noise can as easily promote flat minima as sharp ones. What would make the worry concrete is a closed-loop wetware study relating a defined training signal's variability to off-distribution failure; what would open a path around it is a local, voltage-gated plasticity rule shown to steer a living network toward flatter, more transferable solutions.