A reservoir that hears frequencies, not just amounts
Nearly all spiking reservoir computers run on leaky integrate-and-fire neurons, which are temporal low-pass filters: they register how much input arrived, not at which frequencies. Micheli, Gallicchio and Tomen replace them with a heterogeneous population of resonate-and-fire oscillators and show, in a controlled neuron-for-neuron comparison, that the reservoir becomes a distributed bank of band-pass filters. Representations get measurably more class-discriminative before any training, accuracy rises by up to 26.6 percentage points over the matched integrator baseline, and the spiking formulation lands 9 to 19 times cheaper in energy than a non-spiking oscillatory reservoir of the same design.
Source: Beyond Low-Pass Dynamics: Frequency-Selective Spiking Reservoirs with Resonant Neurons, arXiv:2610.37906v1, 27 October 2026. Primary source. Read: the full PDF, including methods, Tables 1 to 3, the energy cost model, and the appendix controls.
What the work claims
This is a primary methods-and-results paper, a NeurIPS 2026 submission from TU Delft and the University of Pisa, and its central claim is a mechanistic one: the dominant neuron model in spiking reservoir computing throws away temporal frequency information, and restoring it, via intrinsic oscillatory dynamics, is enough to change what the fixed, untrained reservoir can represent.1 The authors build HRF-Res, a liquid state machine in which a conventional leaky integrate-and-fire encoder layer drives a reservoir of Harmonic Resonate-and-Fire neurons, all weights random and frozen, with only a linear readout trained. Against a purpose-built control, LIF-Res, identical in every respect except that the reservoir neurons are integrators with equally heterogeneous time constants and independently re-tuned hyperparameters, the oscillatory version wins on three of four datasets and matches on the fourth.
The headline numbers deserve precision. On sequential MNIST, HRF-Res scores 93.64 plus or minus 0.25 percent against 67.00 plus or minus 1.04 for LIF-Res, a 26.6 point gap. On the Heidelberg spiking audio dataset it scores 83.48 against 72.95, and on the FordA vibration series 69.72 against 67.15. DVS Gesture is the honest exception: 71.81 plus or minus 1.59 versus 73.48 plus or minus 1.12, overlapping error bars, no significant difference. Against the non-spiking oscillatory reservoir RON, which shares the same architecture but computes densely, HRF-Res consumes 13.8, 12.0, 18.7 and 9.6 times less energy on sMNIST, permuted sMNIST, FordA and npCIFAR-10 respectively, while giving up 1.8 to 4.9 accuracy points. On a broader benchmark sweep it reports the best accuracy among spiking reservoir methods, for example 89.78 percent on the speech dataset SHD with 6,000 neurons where the strongest prior baseline used 16,000 neurons to reach 89.30.
How it works
A leaky integrate-and-fire neuron decays monotonically toward rest, so its membrane potential is a smoothed, low-pass-filtered copy of the input; rapid fluctuations are attenuated and the neuron responds most strongly to slow components regardless of the signal's actual frequency structure. The Harmonic Resonate-and-Fire neuron replaces this with the dynamics of a damped harmonic oscillator: two coupled state variables, a velocity and a position, where a stiffness parameter sets the natural oscillation frequency and a damping parameter sets how fast oscillations decay. The neuron spikes when its position crosses a threshold. Because the dynamics include a restoring force, the membrane accumulates energy from inputs that arrive in phase with its natural frequency and attenuates inputs at other frequencies. It is a band-pass filter, not a low-pass one.
HRF-Res draws the stiffness and damping of each reservoir neuron independently from uniform distributions, so the population behaves like a bank of tuning forks, each neuron preferentially reporting activity in its own frequency band. The paper verifies this is not just a single-neuron curiosity that dies under recurrent coupling and discrete-time integration: sweeping 30 log-spaced sinusoids through the full trained-against network, 98, 100, 62 and 83 percent of HRF-Res neurons are band-pass resonant on sMNIST, FordA, SHD and DVS Gesture, against 23, 9, 15 and 4 percent in the matched LIF-Res, and the mean sharpness of tuning is consistently higher, for example a Q-factor of 4.6 versus 2.0 on FordA. Crucially, the untrained reservoir's states are more aligned with class structure, measured by centred kernel alignment between the temporal-mean reservoir state and the label kernel: the gap runs from plus 9 percent on sMNIST to plus 322 percent on FordA, and the ranking of datasets by this gap matches the ranking by sharpness of frequency tuning. Frequency selectivity is doing representational work before a single readout weight is learned.
Energy accounting is done with the standard multiply-accumulate versus accumulate-operations model of the SNN literature, using 45 nm CMOS estimates where a synaptic operation costs about a fifth of a multiply-accumulate. The saving is structural: the dominant recurrent cost, which scales with the square of the neuron count, executes only when spikes arrive rather than densely at every timestep.
Where a skeptic should push
The single most load-bearing assumption is that class-relevant information lives in temporal frequency content. The paper's own data proves this is an assumption and not a law: on DVS Gesture, where class identity plausibly depends on spatial trajectories rather than stable spectral structure, the resonant reservoir confers no advantage over integrators. The authors say this plainly, and it is the most valuable sentence in the paper for anyone tempted to generalise. A second caveat is that the absolute kernel-alignment scores of untrained reservoirs are low by design; the claim is about relative representational quality, not that the frozen reservoir is already a good classifier. Third, the energy figures are model-based estimates, not silicon measurements: the authors themselves list hardware validation as an open gap. The 19-fold figure compares against a specific non-spiking baseline at matched size, and the real deployment gap on actual neuromorphic chips will move with spike rates and implementation details.
There are also robustness asymmetries worth flagging. HRF-Res degrades more gracefully than LIF-Res under additive input noise, but it is more sensitive to changes in the input sampling rate, because resampling shifts the signal spectrum away from the neurons' preferred bands, a sensitivity that follows directly from the mechanism itself. Experiments run over three to five seeds with modest spread, and the benchmark suite is classification only. None of this undermines the mechanism claim; all of it bounds where the mechanism buys anything.
Frequency selectivity as an organoid asset
For organoid intelligence this paper lands in an unusual place: the field's substrate already does the interesting thing, and has mostly been treating it as noise. Brain organoids on microelectrode arrays produce oscillatory network activity, bursts with structured spectral content that shifts over maturation. HRF-Res supplies a concrete computational reading of such dynamics: an oscillatory neuron population is not a defective integrator that happens to ring; it is a frequency-selective feature extractor whose tuning is set by the distribution of intrinsic time constants across the tissue. If an organoid's recurrent network behaves anything like this reservoir, then its endogenous rhythms are doing band-pass decomposition of whatever stimulation it receives, and the readout problem becomes a question of which frequency channels carry task-relevant structure, not whether the tissue can integrate.
The paradigm fit is the deeper point. Reservoir computing is the natural frame for organoid computing because you cannot backpropagate through living tissue: the frozen-recervoir, trained-linear-readout structure of HRF-Res is exactly the training regime a wetware system is confined to. This paper improves the substrate half of that bargain without touching the readout half, and it quantifies the representational payoff of one biological feature, heterogeneous resonance, that organoid tissue possesses in abundance. Its readout ablation is directly transferable: membrane-based statistics of the oscillatory state outperformed spike-count readouts on every dataset, which is a strong argument for reading organoids with field-potential-like or subthreshold measures rather than pure spike counts wherever the hardware allows it.
The threats are equally concrete. The DVS Gesture lesson applies with force: much of what organoid systems are asked to do, pattern recognition after patterned stimulation, may live in spatial and amplitude structure rather than stable frequency structure, in which case endogenous oscillation adds bias, not capability. The sampling-rate sensitivity has a wetware twin in electrode bandwidth and binning choices, which silently move the stimulus spectrum relative to whatever the tissue is tuned to. And the frequency content of organoid rhythms drifts over weeks of maturation and varies batch to batch, so any frequency-channel readout inherits that non-stationarity as calibration debt. The opportunity is real: a reservoir whose filter bank is alive, self-maintaining and reconfigurable by development. So is the trap of mistaking an epiphenomenon for a computation.
The bottom line
Established: in matched spiking reservoirs, oscillatory resonate-and-fire neurons produce sharper frequency tuning, more class-aligned untrained representations, and large accuracy and energy gains over integrator neurons, with the gains concentrated where tasks carry spectral structure. Not established: hardware-measured energy, performance beyond classification, or any benefit when the discriminative information is not frequency-encoded. The result that would confirm the organoid reading is a demonstration that stimulation spectra aligned to an organoid's endogenous rhythmic bands yield better linear-decodable separability than off-band stimulation, controlled against the same tissue at matched firing rates. The result that would break it is the DVS Gesture pattern repeating: no representational advantage over matched integrator-like dynamics on organoid-relevant tasks, in which case oscillation is scenery, not substrate.
Frequently asked questions
What is a resonate-and-fire neuron?
A spiking neuron model whose membrane dynamics follow a damped harmonic oscillator instead of simple leak-and-integrate. It has a natural oscillation frequency set by a stiffness parameter and selectively amplifies inputs arriving near that frequency, behaving as a band-pass filter rather than a low-pass filter.
Is the claimed 19-fold energy saving measured on hardware?
No. It is a model-based estimate comparing synaptic operations in the spiking network against multiply-accumulate operations in a matched non-spiking reservoir, using standard 45 nm CMOS cost figures. The authors list hardware validation as future work.
Does HRF-Res beat fully trained spiking networks?
The paper claims state of the art among spiking reservoir methods, where only a linear readout is trained, not against gradient-trained deep SNNs. Reservoir models trade peak accuracy for training efficiency and are aimed at settings where backpropagation through time is impractical.
Why did the resonant reservoir fail to help on DVS Gesture?
The authors argue that class identity in that dataset is carried by spatial gesture trajectories rather than stable temporal frequency content. Frequency selectivity helps only when the discriminative signal has spectral structure, which is the paper's central boundary condition.
What does this have to do with brain organoids?
Organoid neural networks are oscillatory and cannot be trained by backpropagation, so they sit naturally in the frozen-reservoir regime this paper studies. The work suggests organoid rhythms may function as a filter bank for stimulus processing, and that reading subthreshold or field-potential states may beat spike counts.
What would prove the organoid connection?
Evidence that stimulation aligned to an organoid's endogenous frequency bands produces better separable readouts than off-band stimulation, with firing rates controlled, would support it. Repeated failure to beat integrator-like baselines on organoid tasks would argue oscillation is incidental.
References
- A Micheli, C Gallicchio, N Tomen. Beyond Low-Pass Dynamics: Frequency-Selective Spiking Reservoirs with Resonant Neurons. arXiv:2610.37906. 2026. http://arxiv.org/abs/2610.37906v1. Accessed 2026-11-01.