The leftover membrane charge is why converted spiking nets lose accuracy

What the work claims

This is a methods paper in the ANN-to-SNN conversion lineage: take a well-trained artificial network, replace its activations with integrate-and-fire spiking neurons, and run the result for only a few time steps instead of the hundreds that early conversion schemes required. The authors' diagnosis is that accumulated conversion error at short time windows can be attributed to the distribution of residual membrane potential, the charge still sitting on a neuron's membrane when the simulation ends, which reflects a systematic truncation bias: spikes quantize activation into integer counts, and the leftover charge is the part of the true activation that never made it into a spike.¹

The proposed fix, called SRMP, adds a regularization loss during a short fine-tuning stage that shifts the initial membrane potential of the integrate-and-fire neurons and pushes the residual membrane potential distribution toward the shape that minimizes conversion error. A companion layer, SCR-Conv2d, uses lateral inhibition, a grouped convolution in which active neurons suppress their neighbors, to sharpen features and remove redundant spikes. Combined with the strongest existing conversion baseline, the method reports accuracy gains at two, four and eight time steps on CIFAR-10, CIFAR-100 and ImageNet, and the authors show the same recipe works on converted transformers and on multi-threshold spiking variants, with what they describe as negligible extra computation.

How it works

An integrate-and-fire neuron integrates incoming spikes into a membrane potential and emits a spike when the potential crosses a threshold; its activation is therefore an integer count over a finite window. A converted artificial neuron has a real-valued activation. The integer counts can approximate the real values, but each neuron finishes the window with some residual charge below threshold, and because the residual correlates with where the activation sat relative to the threshold boundaries, the errors accumulate layer by layer rather than averaging out. The authors derive the residual-membrane-potential distribution that would minimize this bias and then regularize toward it by adjusting each neuron's initial membrane potential, which is a learnable offset rather than a structural change. During the second-stage fine-tuning, which runs for ten epochs with weights and firing thresholds frozen and only the initial potentials and regularizer active, the network keeps its learned function while its internal state geometry is corrected.

The reported numbers make the point concrete. On ImageNet with a VGG-16 backbone, the combined SCR-SRMP method reaches 56.46 percent at 16 time steps, which is 5.54 percentage points above the QCFS baseline under the same conditions. On CIFAR-10 with VGG-16, the converted network retains 92.84 percent at just two time steps against a 95.65 percent source network, versus 90.67 percent for the baseline at that step count. The competitive refinement layer helps specifically when input distributions vary, because its inhibitory neighborhoods suppress co-firing and stabilize the encoding when the time window is tiny.

Where a skeptic should push

The most load-bearing assumption is that conversion is the right frame at all. Conversion inherits the artificial network's function; it cannot exceed the source network's accuracy, and the paper's own framing concedes this upper bound. Everything reported is an improvement in closing the gap, not in surpassing it, and the gains are largest precisely where the absolute accuracies are lowest, at two to eight time steps. A reader should also weigh that fine-tuning, even ten epochs with frozen weights, is still an extra training stage whose hyperparameters, including a regularization weight that the authors set differently per dataset, carry real tuning burden.

Second, the evaluation is vision classification on standard benchmarks. Conversion error on transformers is demonstrated, but the harder question, whether short-window converted spiking nets hold up on temporal, streaming tasks where spikes should shine, is left open. Third, the claimed negligible overhead is an architectural-count argument, not a chip measurement; no energy or latency figure on deployed neuromorphic hardware appears in the paper. Finally, the theoretical characterization of the optimal residual distribution rests on assumptions about activation distributions that the authors themselves note break down at larger time steps, where the residual becomes nearly uniform and the method stops helping.

Programming versus training the spiking substrate

The deeper significance of this paper for organoid intelligence is philosophical as much as technical. Conversion embodies one extreme of a spectrum: do not teach the substrate anything, compile a finished solution into it. That approach has always looked hopeless for living tissue, and it still does, no one is going to fine-tune the initial membrane potentials of a hundred thousand cells in a dish with a per-dataset regularization weight. But the mechanism of the fix translates, and it is the non-obvious part. The authors show that the dominant conversion error is not in the weights but in the mismatch between the distribution of internal state at inference time and the distribution the computation assumes. Close that gap and a few discrete events suffice. The analogous claim for biological computing is sharp: when a spiking organoid fails to reproduce a trained behavior, the first suspect should be baseline state mismatch, tonic activity, initial conditions, homeostatic set points, not the learned connectivity. Priming the network's state before each trial may matter as much as any stimulation pattern, because the residual error lives in what the tissue is holding, not in what it has wired.

There is a second, quieter lesson in the method: the fix is distributional, not parametric. SRMP does not change what the network computes; it aligns the statistics of a latent variable so that quantization stops accumulating bias. For closed-loop organoid training, where the training signal is a stimulation pattern and the readout is spikes, the same principle suggests that calibrating the statistics of background activity, for instance through adaptation of the culture's excitation-inhibition balance, could reduce the effective bit cost of every subsequent computation. That is a hypothesis, not a result, and the paper contains no biology, but the mechanism is substrate-generic: state distribution governs quantization error.

The threat is the usual one from a fast-improving neighbor. Every result like this narrows the gap between a trained artificial network and its spiking shadow. If converted spiking networks reach near-parity at two to eight time steps on ImageNet-class problems with negligible overhead, then low-latency, energy-frugal spiking inference becomes a solved engineering problem on deterministic silicon, and one more plank of the case for biological computing, that only neuromorphic hardware can do efficient spike-based inference, is removed. The opportunity for the field is to internalize the correction and stop competing on that axis. What living tissue offers that conversion never will is in-situ adaptation: the substrate rewires itself during use. The honest position after this paper is that organoid intelligence should be sold on plasticity, self-repair and continuous learning, not on spike efficiency, because the spike-efficiency argument is being steadily won by compilation.

The bottom line

Established: for converted spiking networks, a large share of short-window accuracy loss is attributable to residual membrane potential statistics, and a short fine-tuning stage that aligns those statistics, plus a lateral-inhibition refinement layer, recovers several percentage points at two to sixteen time steps across standard vision benchmarks, including a 5.54 point gain over the strongest baseline on ImageNet at 16 steps. Hypothesis, for organoid intelligence: state-distribution alignment, the tissue-level analogue of residual membrane potential alignment, is an underexploited lever for making trained behavior robust in living substrates. What would confirm it: closed-loop studies showing that baseline-state priming reduces the stimulation cost or trial count needed to elicit a trained response. What would break the broader claim: chip-level measurements showing the overhead is not negligible, or direct-spiking training surpassing conversion at the same latency, which would make the whole conversion frame obsolete.