Fourier graph operators go spiking for forecasting

What the work claims

Bakhshaliyev and Landwehr at the University of Hildesheim claim that the missing ingredient in spiking time series forecasting is explicit cross-variable modeling. Prior spiking forecasters, including spiking transformers and the Fourier-domain SpikF, process each variable independently; correlations between variables, which in domains like traffic networks or power grids carry much of the predictive signal, go unexploited. SpikF-GO imports the hypervariate graph formulation from the ANN model FourierGNN, in which every scalar observation at every variable and every time step of the input window is a single node, and combines it with spike-driven spectral processing: a Spiking Fast Fourier Transform moves the graph signal to the frequency domain, a Hard Concrete gate learns which frequency bins to keep (encouraging sparse, hardware-friendly frequency usage), and a Complex LIF gate applies independent leaky integrate-and-fire neurons to the real and imaginary parts of the spectrum, so even the frequency-domain arithmetic stays binary and event-driven. A variant adds Central Pattern Generator positional encodings to strengthen long-range temporal structure.¹

The evidence claim: on eight public benchmarks under a unified experimental protocol, SpikF-GO with CPG achieves the best average rank among all compared spiking methods on both the coefficient of determination (average rank 2.4) and mean absolute error (2.3), the plain SpikF-GO ranks second on R-squared (2.8), and both surpass the ANN baseline FourierGNN. The energy claim: a theoretical analysis on a 45 nm hardware model reports 1.89 times lower energy than FourierGNN, rising to 7.86 times at a compact embedding dimension of 8, with little accuracy loss. The authors also release SpikingTSF, a unified comparison of spiking forecasting architectures under one protocol, which the field arguably needed more than another model. This is a preprint, not yet peer reviewed.¹

How it works

The hypervariate graph is the conceptual engine. A conventional spatiotemporal model keeps two modules: one for time, one for cross-variable space. The hypervariate formulation collapses them: a node is a (variable, time step) pair, and the fully connected graph over these nodes encodes intra-series temporal dependencies, inter-series spatial dependencies, and time-varying cross-variable interactions in a single object. Multiplication by a complex linear operator in the Fourier domain of this graph is equivalent to graph convolution in the node domain, so spectral mixing is implicit graph convolution, and because the operators are gated by spiking neurons on both the real and imaginary components, the equivalence survives spike-driven computation. The Hard Concrete gate makes the frequency selection sparse and learnable, which matters twice: it cuts energy, and it means the model itself decides which frequency bands of the array carry forecast-relevant information.¹

The benchmark suite is broad: Traffic, METR-LA, and PEMS-BAY traffic flow; Solar and Electricity energy systems; COVID-19 case records; ECG biomedical signals; and Wiki web traffic, spanning 55 to 2,000 variables at granularities from 5 minutes to daily, with chronological splits (7:2:1, and 6:2:2 for COVID-19). Nine spiking baselines across the major architectural families, including spiking RNNs, TCNs, transformers, and the prior Fourier-domain SpikF, are compared under the same protocol, which is the study's real public good: before this, spiking forecasting results were scattered across incompatible experimental setups. Notably, the comparison includes the observation that wall-clock runtime on GPU actually favors FourierGNN, because spiking simulation incurs multi-step overhead that would not arise on neuromorphic hardware.¹

The energy accounting is explicit about its assumptions: 4.6 pJ per floating-point operation and 0.9 pJ per synaptic operation on 45 nm technology, with memory-access and addressing costs decomposed following standard models, and the Spiking FFT costs estimated from the prior SpikF analysis. Under those assumptions SpikF-GO at embedding 128 uses 1.89 times less energy than FourierGNN while forecasting better, and shrinking the embedding to 8 pushes the factor to 7.86; the lighter prior model SpikF reaches 4.27 times but does not model cross-variable dependencies and does not beat the full model.¹

Where a skeptic should push

The most load-bearing assumption is that average rank over benchmarks is an adequate proxy for superiority, and it deserves pressure. An average rank of 2.4 means the model is usually second or third on any given dataset, not that it wins everywhere; ranks compress absolute margins, hide variance, and reward consistency more than peak performance. The authors mitigate this with a unified protocol, multiple runs with reported standard deviations, and results across forecast horizons, which is good practice, but a reader should still want the absolute-error tables, not just the rank summary, before concluding the ANN baseline is beaten in any operationally meaningful sense.

Second, the energy advantage is theoretical and the authors say so plainly. It rests on operation-count prices from a 45 nm model and on the assumption that synaptic operations are cheap; real neuromorphic hardware is neither uniformly available nor uniformly efficient, and the GPU wall-clock result in the same paper, where the ANN is fastest because spiking simulation costs extra on conventional hardware, is a live demonstration that theoretical energy and realized energy can point in different directions. The 7.86 times figure at embedding 8 is additionally a configuration choice: smaller embeddings trade model capacity for energy everywhere, in ANNs as well, so part of that headline is the familiar accuracy-energy frontier rather than a spiking-specific dividend.

Third, scalability cuts against the design. A fully connected hypervariate graph has nodes equal to variables times window length, so cost grows with their product; the authors themselves flag complexity growth in high-dimensional long-horizon settings as a limitation. ECG at 55 variables and PEMS-BAY at 2,000 variables are very different regimes, and the graph formulation will feel that difference. And one benchmark family here, biomedical ECG, is the closest any of this gets to physiological data; it is still a public tidy benchmark, not noisy implanted-tissue recording.¹

Spectral spiking models for organoid readouts

The non-obvious implication for organoid intelligence is that this paper exposes a shared blind spot. Multichannel electrode array recordings are multivariate time series; the standard analysis pipeline, per-electrode thresholding, firing-rate features, and then independent channel statistics, commits exactly the sin SpikF-GO was built to correct: it processes each variable independently and discards cross-electrode dependencies. If co-moving channels carry predictive signal in traffic and grid data, they carry it in a dish too, where the interesting computation is precisely the coordinated structure across electrodes. SpikF-GO shows that spike-driven spectral graph mixing can capture that structure in principle while remaining binary and event-driven throughout the spectral domain, which is the natural currency of a spiking readout, and it beats the floating-point ANN doing the same graph math under a unified protocol. That is a blueprint for an edge-side readout front end: spectral-spatial compression of an electrode array performed in spikes, near the array, rather than shipping raw traces to a GPU cluster.

The opportunity has a second layer in the Hard Concrete gate. Because the gate learns which frequency bins matter and reports them, it doubles as an assay: run it on MEA data and the selected bands are an empirical answer to which oscillatory structure of the culture actually carries forecast-relevant information, as opposed to which bands merely look prominent in a power spectrum. For a field that debates whether organoid oscillations are computation or epiphenomenon, a sparse, task-driven band selection is a sharper instrument than most of what is currently measured, and it comes as a byproduct of a forecasting model.

The threat is the obsolescence argument, now with better tools. The strongest version of "why tissue" has always been temporal efficiency; this paper is another data point that silicon can execute spectral, graph-structured, event-driven temporal computation directly, in spikes, at claimed fractions of the floating-point energy, with a training recipe that tissue cannot run. Whatever organoids contribute to temporal processing will have to be something a Hard-Concrete-gated spiking network cannot: most plausibly adaptive internal weights rather than efficient static operators. There is also a quieter warning in the CPG variant's success: long-range temporal structure needed explicit positional scaffolding to materialize, and living tissue gets that scaffolding from developmental patterning, which dissociated cultures largely lack. A random cortical organoid may be a SpikF-GO without the CPG module: competent at local mixing, structurally impoverished at long range, for reasons no amount of readout engineering will fix. Readouts can be upgraded; ontogeny cannot be retrofitted.¹

The bottom line

Established within this preprint: under a unified protocol across eight benchmarks, spiking spectral graph operators rank first and second among spiking forecasters and surpass the ANN FourierGNN, with cross-variable modeling identified as the differentiator and a reusable benchmark suite released alongside. Established only on paper: the energy advantage, which is an operation-count model at 45 nm with no neuromorphic hardware deployment, and which the paper's own GPU timing table complicates. For organoid intelligence, the durable contributions are the diagnosis and the instrument: independent-channel analysis of multivariate signals is provably leaving signal unexploited, and a spiking spectral graph front end with a learnable band gate is a plausible near-tissue readout architecture that would also tell you which of the culture's rhythms deserve attention. What would confirm both claims is deployment on real neuromorphic hardware with measured energy, and a demonstration on genuinely noisy multivariate physiological recordings rather than tidy public benchmarks; what would break the energy story is evidence that on available hardware the spiking overhead erases the theoretical dividend, as the GPU runtime already hints it might.