Beyond Low-Pass Dynamics: Frequency-Selective Spiking Reservoirs with Resonant Neurons

AI summary

Overview

Research area: Neuromorphic computing and reservoir computing — specifically spiking neural networks (SNNs), liquid state machines, and oscillatory neuron models for temporal signal processing.

Technical level: Intermediate. The paper assumes familiarity with reservoir computing, leaky integrate-and-fire (LIF) neurons, and standard sequence-classification benchmarks, but the core argument is conceptually simple: swap low-pass neurons for band-pass neurons in a fixed spiking reservoir.

Scope: The paper proposes HRF-Res, a spiking liquid state machine built from Harmonic Resonate-and-Fire neurons, and evaluates it on ten benchmarks across four modalities (time series, sequential images, audio spike trains, neuromorphic vision), with controlled comparisons against an equivalent LIF reservoir and a non-spiking oscillatory reservoir (RON).

What This Paper Is About

Spiking reservoirs inherit the training efficiency of reservoir computing and the event-driven energy profile of spiking networks, but they almost universally use LIF neurons, which act as low-pass filters and therefore encode how much input arrives rather than at which temporal frequencies. The authors replace LIF units in the reservoir with a heterogeneous population of oscillatory HRF neurons that resonate at different intrinsic frequencies, aiming to capture frequency-specific temporal structure. The goal is to show that frequency selectivity is itself a computational principle for spiking reservoirs — improving representational quality and accuracy without giving up spiking energy efficiency.

Key Contributions

1. Frequency selectivity survives in a full recurrent spiking reservoir. The authors show that HRF neurons act as distributed band-pass filters even under discrete-time integration, recurrent coupling, and random heterogeneous connectivity — not just as isolated single neurons.

2. Intrinsically more class-discriminative representations before any readout is trained. Using Centred Kernel Alignment (CKA) on frozen reservoirs, HRF-Res states align better with class labels than an equivalent LIF reservoir (LIF-Res) on all four analysed datasets, with the largest gap on FordA (+322%). A controlled ablation holds everything fixed except neuron dynamics, with independent hyperparameter tuning for LIF-Res per dataset.

3. Higher classification accuracy across modalities from the same architectural change. HRF-Res outperforms LIF-Res on three of four datasets with gains up to +26.6% (sMNIST: 93.64% vs. 67.00%), while consuming less energy on most datasets.

4. Substantial energy reduction relative to a non-spiking oscillatory reservoir. Against RON, HRF-Res achieves 9–19× lower inference energy (up to 19×) with accuracy losses of only 1.8–4.9 percentage points, and state-of-the-art accuracy among spiking reservoir methods on the reported benchmarks.

Main Findings

• Band-pass behaviour is measurable in the recurrent network. The fraction of resonant neurons (Q > 0) is 98%, 100%, 62%, and 83% for HRF-Res on sMNIST, FordA, SHD, and DVS Gesture respectively, versus 23%, 9%, 15%, and 4% for LIF-Res (mean over three seeds, standard deviations at most 2 percentage points).

• HRF filters are sharper, not just more numerous. Among resonant units, mean Q-factors are 4.2 vs. 3.6 (sMNIST), 4.6 vs. 2.0 (FordA), 2.5 vs. 0.6 (SHD), and 3.4 vs. 1.4 (DVS Gesture).

• LIF neurons collapse to the lowest frequency. Preferred-frequency standard deviations for HRF are 0.049, 0.041, 0.021 and 0.016 Hz, versus 0.002–0.008 Hz for LIF — consistent with pure integration. On FordA, HRF responses concentrate around 0.42 Hz.

• Representational advantage precedes readout training. CKA gaps favour HRF-Res on all four datasets: +322% (FordA), +13% (DVS Gesture), +9% (SHD), +9% (sMNIST), with the sMNIST difference lying within seed-to-seed variability. The CKA ranking matches the Q-factor gap ranking.

• Accuracy gains mirror the representational gains. sMNIST: 93.64 ± 0.25 vs. 67.00 ± 1.04 (+26.6%); SHD: 83.48 ± 0.44 vs. 72.95 ± 0.53 (+10.5%); FordA: 69.72 ± 0.68 vs. 67.15 ± 0.66 (+2.6%). DVS Gesture is the exception: 71.81 ± 1.59 vs. 73.48 ± 1.12, with overlapping standard deviations indicating no significant difference.

• Spiking is cheaper than LIF here, not just than RON. HRF-Res uses less energy than LIF-Res on most datasets despite higher accuracy, because LIF neurons appear to need higher firing rates to encode temporal information. DVS Gesture is again the exception.

• Energy savings against RON are structural. Per-benchmark reductions: 13.8× (sMNIST, 95.44% → 93.64%), 12.0× (psMNIST, 91.33% → 87.96%), 18.7× (FordA, 73.15% → 69.72%), 9.6× (npCIFAR-10, 38.50% → 33.57%), at matched N = 800 and final-state readout. Including the activity-independent O(N) neuron updates, the reduction remains 9–17×.

• Membrane readouts beat spike readouts. Across six readout strategies, the best membrane-based readout on the HRF position state outperforms the spike-rate readout on every dataset; temporal averaging beats the final state by ~4% on FordA; the combined 3N-dimensional readout (RMS + standard deviation + final state) is best or near-best on all four datasets at triple the parameter count.

• Competitive against prior spiking reservoirs with fewer neurons. SHD: 89.78 ± 0.35 with 6,000 neurons vs. 89.30 for a 16,000-neuron ELSM. FordA: 82.05 ± 0.55 with 2,500 neurons vs. 80.37; FordB: 70.55 ± 0.14 with 800 neurons vs. 64.32; Wafer: 99.33 ± 0.04 vs. 98.85; Earthquakes: 77.22 ± 1.10 vs. 71.94. N-MNIST: 97.13 ± 0.12 vs. 94.43, but below 98.10 from an ensemble method; DVS Gesture: 81.06 ± 0.55 vs. 85.20.

Methodology in Plain English

The model has two fixed, randomly connected populations. An input signal first drives a layer of standard LIF neurons. Those LIF spikes feed a reservoir of HRF neurons, each behaving like a damped tuning fork with its own natural frequency and damping set by two per-neuron parameters. The HRF neurons spike back to the LIF layer, closing a recurrent loop. No weights are trained inside the reservoir; only a linear readout on the final HRF state is fitted, exactly as in standard reservoir computing.

The key comparison is a deliberately matched one. The authors build LIF-Res, identical in every respect except that the oscillatory population is replaced by LIF neurons with heterogeneous time constants and thresholds drawn from log-uniform distributions — mirroring the HRF heterogeneity — and give it its own independent hyperparameter search per dataset. Any difference in results is therefore attributable to oscillatory versus integrative dynamics.

To verify that frequency selectivity actually exists, they sweep 30 log-spaced sinusoidal inputs through the full network, identify each neuron's preferred frequency via the Discrete Fourier Transform, and compute a Q-factor from the half-power bandwidth. To test whether that selectivity matters, they compute CKA between temporal-mean reservoir states and label kernels without training a readout — a conservative choice, since averaging partially cancels the oscillatory component that distinguishes HRF from LIF.

Energy is counted with a standard hardware cost model: multiply-accumulate operations for non-spiking models, synaptic operations (each one an accumulate) for spiking models, converted via CMOS estimates where an AC costs roughly 5× less than a MAC in 45 nm. The recurrent O(N²) term dominates, and because it is gated by spikes rather than executed densely, the savings are structural rather than task-specific.

Why This Matters

Impact on research. The paper reframes neuron choice in spiking reservoirs as a representational design decision rather than an implementation detail. It provides a controlled, single-variable demonstration that band-pass dynamics improve class discriminability, connecting classical observations about resonant subthreshold dynamics in liquid state machines to modern spiking reservoir practice — and doing so inside a network rather than at the single-neuron level.

Real-world applications:

• Edge vibration and industrial monitoring: univariate sensor streams like FordA/FordB, where class-relevant structure is periodic and HRF tuning concentrates around a dominant temporal scale.
• Neuromorphic audio: keyword spotting on spike-encoded audio (SHD), where frequency content is directly class-relevant.
• Event-based vision: gesture and digit recognition on event cameras (DVS Gesture, N-MNIST), with the caveat that spatial trajectory information currently favours LIF or ensemble approaches.
• Low-power sensing on neuromorphic hardware: deployment targets such as Intel Loihi, IBM TrueNorth, and SpiNNaker, where the spike-based readout variant enables an end-to-end event-driven pipeline.

Industry relevance. The 9–19× inference-energy reduction at 1.8–4.9 points of accuracy loss is the kind of tradeoff that matters for battery-constrained devices. The paper also reports a limitation that practitioners need: HRF-Res is more sensitive to input sampling-rate shifts than LIF-Res, which moves the input spectrum away from the neurons' preferred bands, even though it degrades more gracefully under additive noise.

Future Directions

• Learning the resonance parameters. Tuning γ and ε with surrogate gradients would trade some of the reservoir's training efficiency for task-adapted frequency selectivity, potentially closing gaps like DVS Gesture.

• Reintroducing dynamic synapses. The original LSM formulation included short-term synaptic plasticity, which the authors omitted to isolate neuron dynamics. Since synapses filter at the synapse and HRF neurons filter at the membrane, the two mechanisms may be complementary.

• Validating energy on real hardware. Current energy figures are model-based; on-chip measurement on neuromorphic platforms is needed, particularly to justify the spike-based readout pipeline end to end.

• Theoretical characterisation. Memory capacity, separation properties, and stability of HRF reservoirs remain unanalysed, as does robustness under distribution shift beyond the noise and sampling-rate probes in Appendix C.

Target Audience

Graduate students and researchers in neuromorphic computing, reservoir computing, and spiking neural networks who are interested in neuron-model design as a lever on representational quality. It is also useful for edge-ML engineers evaluating whether spiking or oscillatory reservoirs can meet a power budget, and for computational neuroscientists interested in how subthreshold resonance translates into network-level function. Readers without background in reservoir computing or spiking dynamics will need to consult the cited LSM and LIF literature first.

Authors’ abstract

Spiking reservoir computing combines the training efficiency of fixed recurrent reservoirs with the energy efficiency of event-driven inference, where computation is triggered by sparse spikes rather than executed densely at every timestep. However, existing spiking reservoirs almost exclusively rely on Leaky Integrate-and-Fire (LIF) neurons, whose low-pass dynamics limit their ability to distinguish information encoded at different temporal frequencies. We introduce HRF-Res, a liquid state machine in which LIF neurons are replaced by a heterogeneous population of Harmonic Resonate-and-Fire (HRF) neurons. These neurons exhibit intrinsic oscillatory dynamics and selectively respond to input frequencies, enabling the reservoir to capture frequency-specific temporal patterns. The result is more discriminative representations and higher classification accuracy than LIF-based reservoirs, achieved without sacrificing the energy efficiency of spiking computation, with up to 19x lower energy consumption than an equivalent non-spiking oscillatory reservoir. Evaluated on ten benchmarks spanning time series, sequential images, audio spike trains, and neuromorphic vision, HRF-Res achieves state-of-the-art accuracy among spiking reservoir methods. These results demonstrate that introducing frequency-selective dynamics into spiking reservoirs improves representational quality and classification performance while preserving their inherent energy efficiency.