acceptodds
Under review as a conference paper at ICLR 2027

Spoken-Digit Recognition Without Training: Geometry, Coupling, and Drive Effects in Oscillator Networks

Abstract

Coupled-oscillator networks are gaining interest as a machine-learning substrate, on the premise that oscillator physics turns signals into useful structure. Yet the ablations that would isolate what the oscillator dynamics contribute are rarely run. We run them for untrained coupled oscillator networks, using spoken-digit classification on AudioMNIST as the measure. A small untrained network is read by a closed-form linear readout and compared with five trained conventional networks (GRU, TCN, CNN, transformer and S4D) matched in parameters and with three controls: the same network with its coupling removed, a non-oscillating bank of leaky integrators with the same states and parameters, and the same readout fitted directly to the network's input. Every experimental arm's state is summarized by the same time-pooled statistics and spoken digits classified using the same number of features, so that readout capacity cannot mask dynamics. The ablations, eight experiments of 6,353 runs in all, vary six coupling functions, eight lattice geometries, and the natural frequencies, input pathway, input gain, readout width and training set size, on digit recognition and on a temporal order task that order-blind readouts cannot solve. The design separates three explanations for an untrained network's accuracy: its input representation, the width of its readout, and the oscillator dynamics and structure itself. Our ablation study results give future studies of trained oscillator networks an untrained baseline to compare against for speech recognition tasks, and insight into which factors contribute to an oscillator network's effectiveness on time series tasks.

Primary area:
learning on time series and dynamical systems
Keywords:
coupled oscillators, reservoir computing, Kuramoto, Stuart-Landau, oscillatory neural networks, controlled ablations, spoken digit classification, AudioMNIST

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GuWEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper gives an ablation analysis of untrained coupled oscillator networks, which are basically used as reservoirs for the task of spoken digit recognition on the AudioMNIST data. The reservoir is connected to a linear readout for the output computation, and the results are analyzed and compared in different setups, including: with the uncoupled network, with the input alone, with leaky integrators, and with small fully trainable baselines. The paper studies the performance by varying several aspects of the methodology, among which are the coupling function, natural frequencies, lattice geometry, input gain, and pathway. The main result is that the reservoir-based network is able to classify the digits almost as well as its own input, and that the untrained reservoir is mainly adding memory.

Critical strengths
  • I appreciated the design of the experimental analysis, which I could find pretty careful and transparent. In particular, the authors included controls that are often left out in most contributions in the area, such as the ablation vs the input alone, the uncoupled network, and the leaky integrators.
  • Systematic ablations of oscillatory networks are largely missing in the related literature, and this paper can be considered as a reference for the future works in this sense.
Critical weaknesses
  • Lack of model selection for the oscillatory networks risks to reduce the soundness of the reported comparisons. In fact, from my understainding, the reference configuration is fixed, the input gain is also set to 1 or 2 in the main experiments.

  • I was surprised to notice that some natural baselines are missing in the comparison, including standard echo state networks, and previously proposed architectures in the area of coupled oscillators networks (see also the questions section).

Questions for authors
  • How did you choose the reference configuration anf the input gains? Please report the results achieved after careful selection of the hyperparameters on a validation set.
  • Please include among the baselines to test also other relevant models from the reservoir computing literature, including the conventional echo state networks, and the reservoir oscillator network RON (Ceni et al., 2024), again after propser tuning of the hyperparameters through rrigorous model selection.
  • It would be interesting to add a further control to assess the effect of the dynamics against the effect of the dimensionality. For example, using the same readout, but applying it to a random nonlinear expansion of the input alone.
Reviewer_fvkpOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors perform a series of ablation experiments on networks of oscillators used to classify spoken digits in AudioMNIST. They find that oscillatory dynamics generally provide limited improvement over suitable controls. The oscillator networks retain temporal information, but simpler non-oscillating controls also provide memory.

Critical strengths

Mechanistic understanding of any class of networks is important.

Critical weaknesses

The paper is written in a way that is very hard to read. For instance, the abstract does not state the main findings. The experiments address a coherent question about the contribution of oscillator dynamics, but are presented largely as a list of comparisons, with insufficient explanation of how the results build toward a central conclusion. The discussion provides some of this, but at a very late stage of the manuscript.

Beyond these strong presentation issues, I am not convinced that the study establishes a significant scientific contribution. The parameter sweeps document performance differences within the chosen setup, but provide limited explanation of when or why oscillator dynamics help. It remains unclear which findings offer substantial new insight into oscillator networks, or how broadly the conclusions apply beyond the particular input representation, network configuration, and readout used here.

Questions for authors

See above

Reviewer_AAAqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This is an ablation study involving eight experiments and 6,353 runs for an untrained oscillator architecture, which performs both a speaker-disjoint AudioMNIST classification task and a temporal order task. It compares pools of oscillators (oscillator reservoirs) against pooled log-mel input features, uncoupled oscillating units, leaky-integrator reservoirs, and five different trained neural network architectures (GRU, a dilated residual TCN, CNN, transformer and S4D). The primary digit recognition task uses a ridge-type readout of 192 features, supplemented by experiments varying readout width and training sample size. The results indicate that the front-end and the readout account for much of the single-digit classification accuracy, that even non-oscillating integrated memory is valuable in this task, that free-amplitude Stuart-Landau performs better than the input baseline in the principal noisy comparisons and is less sensitive to increased input gain over part of the tested range, and that the choice of network structure and phase-coupling has a relatively small effect on performance.

Critical strengths

Strengths

• Broad experimental comparison. The study examines multiple coupling functions, geometries, input gains and readout widths within a common evaluation framework. Speaker-disjoint testing and noisy conditions strengthen its practical relevance. • Useful controls for interpreting reservoir performance. The input-only baseline, uncoupled reference network and state/signal-matched leaky-integrator banks help distinguish the benefits of the front end, temporal memory and oscillator interactions. The strong performance of leaky integrators on the order task is particularly informative. • Amplitude and gain ablations. Including both free and fixed amplitude Stuart Landau variants, together with an input-gain sweep, provides a useful investigation of when amplitude dynamics may matter. The readout width and training size experiments also reveal important limits of the primary comparison.

Critical weaknesses
  1. Incomplete specification of the Stuart-Landau comparison: Section 3.1 and Appendix B specify the general phase equation and the complex coupling term, but do not clearly and explicitly state the complete driven Stuart-Landau equations, the amplitude-clamping procedure, and the numerical updates. It is important that Appendix B states that at fixed amplitude the Stuart-Landau equations reduce to the Kuramoto equation, yet Figure 6 shows a large divergence at higher gains. The paper also does not specify the solver, the integration timestep, or the initial states. Finally, since Kuramoto is already a unit-amplitude phase model, it is essential to clarify how the fixed-amplitude Stuart-Landau ablation differs from the Kuramoto comparator already used in Zhang et al.'s earlier work.

  2. Dependence on the dimensionality-reduction procedure: At the primary width, oscillator features are reduced from 24,576 to 192 via a random Gaussian matrix, whereas the input baseline retains its 192 pooled features. The authors acknowledge this asymmetry and provide useful checks on the width and seeded projection. Still, because a random Gaussian matrix mixes signals across channels, rows, and columns, it remains unclear how sensitive the principal amplitude and coupling conclusions are to the reduction method itself. A training-set PCA, a lightweight shared spatial layer, or a clearly specified structured pooling control at the same final width could help assess this.

  3. Limited support for the proposed amplitude mechanism: The suggestion in Section 5 that the amplitude absorbs strong input is plausible, but tentative. Figure 11, top panel, shows higher mean F-statistics for Stuart-Landau in bands 8-15, which is an interesting observation, but the figure covers only one seed and operating condition. This does not establish that the pattern explains either the accuracy improvement or the robustness to increased gain. A clearer distinction could be drawn between the empirical amplitude advantage and its proposed mechanism.

Questions for authors
  1. Can the authors provide the exact update equations for both variants of Stuart-Landau, including all the terms for inputs, restoring pull, amplitude clamps, readouts, and the details of the solver, timestep, and initializations? This would help explain the divergence from the Kuramoto model in Figure 6 and clarify the differences between the fixed-amplitude variant and the Kuramoto comparator in Zhang et al.

  2. Apart from the reported width and seeded projection checks, how sensitive are the main amplitude and coupling conclusions to replacing the random Gaussian matrix with another dimensionality reduction method, like a training-set principal component analysis (PCA) or a specific structured pooling scheme, at the same final width?

  3. How do the authors interpret Stuart-Landau's higher mean F-statistics in bands 8-15 in Figure 11, and is there evidence connecting this observation to its noisy classification performance or its response to increased input gain? Additional feedback

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.