{"id":"7a56c0ee-d54d-415e-a3c5-3802a667d2bc","arxiv_id":"2607.24420","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Independent oscillatory E-I units selectively amplify input frequencies near their intrinsic frequencies, yielding an interpretable reservoir that matches random reservoirs and can be tuned for short-term chaotic prediction.","lead":"A reservoir computer built from independent oscillatory units, each tuned to a different frequency, stores input by selectively amplifying nearby frequencies and can match or beat random reservoirs on chaotic forecasting. The design makes the reservoir interpretable and tunable for short-term prediction using forced-oscillator ideas rather than black-box hyperparameter search.","discovery_kind":"new_method","skeptic_critique":null,"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript proposes a \"frequency-based reservoir computer\": a recurrent layer whose connectivity matrix is block-diagonal with 2×2 rotation-scaling blocks, so that the reservoir is an ensemble of independent damped oscillatory units (motivated by Wilson–Cowan E–I pairs), each with a prescribed intrinsic frequency ω_k drawn from a roster spanning [ω_min, ω_max]. Only a linear readout (minimum-norm least squares) is trained. The paper claims three things: (i) driven units selectively amplify input frequency components near their intrinsic frequency — shown via pwelch heat maps for synthetic periodic, Mackey-Glass, and Lorenz-63 inputs; (ii) this reservoir performs as well as or better than a random reservoir of matched size and connection count on Mackey-Glass, Lorenz-63, and Kuramoto-Sivashinsky closed-loop forecasting; (iii) the frequency roster can be reshaped to improve short-term Lorenz prediction, an optimization capability asserted to be absent in random reservoirs. The architecture is clean, sparse (4N−1 connections), and the empirical demonstrations are plausible, but the mechanistic claim is supported only qualitatively and the comparative claims rest on thin statistics and a possibly unfair baseline protocol.","tokens_in":18437,"tokens_out":4403,"duration_ms":153329,"significance":"If the claims hold up, the paper offers a genuinely interpretable reservoir design: an extremely sparse recurrent layer (4N−1 connections for N units, vs. O(N²) in standard ESNs) whose hyperparameters all carry physical meaning (leak α scales intrinsic frequencies, spectral radius sets ω_max, input scale ε controls the linear/nonlinear response regime). The uncoupled-oscillator construction is simple enough to be reproducible from the text, is directly relevant to neuromorphic implementations, and the short-term-forecast result in Fig. 7f with statistics over 500 initial conditions is the kind of falsifiable, quantified evidence the field values. The selective-amplification picture, if made quantitative, would be a useful handle on a long-standing interpretability problem. Impact is moderate: the architecture is close in spirit to existing oscillator reservoirs, and the practical performance advantage over tuned ESNs is not yet demonstrated.","major_comments":[{"comment":"§II, 'Selective frequency amplification as a mechanism for chaotic time series prediction' (Figs. 4–6): the central mechanistic claim rests entirely on qualitative pwelch heat maps. The load-bearing inference — 'when the reservoir is sufficiently large ... a simple linear combination W_out r(t) should be enough to recover all the information about the input' — is asserted, not demonstrated. For a broadband (continuous-spectrum) input like Mackey-Glass or Lorenz-63, amplification of components near intrinsic ω_k does not by itself imply the readout can reconstruct u(t+1); one needs a quantitative link between the units' frequency responses and prediction error. A concrete test would be: (i) measure each unit's effective transfer function under the actual driving input; (ii) correlate per-band readout weight magnitude |W_out| with amplification; and (iii) ablate units in a frequency band a","section":"§II, Figs. 4-6"},{"comment":"The comparative claim 'performs as well as or better than equivalent random reservoirs' is supported only by single-trajectory plots. Fig. 3 shows one Mackey-Glass and one Lorenz-63 closed-loop run with no error metric; only Fig. 7f gives statistics (mean±std over 500 ICs), and only for the short-term Lorenz task. Fig. 8 (Kuramoto-Sivashinsky) reports no random-reservoir baseline at all, despite the same-comparison framing used elsewhere. Standard quantitative metrics (valid time in Lyapunov units, NRMSE with dispersion over initial conditions and over reservoir/input-matrix realizations — note W_in is random even for the frequency-based reservoir) should be reported for all three systems, otherwise the headline comparison is not evaluable.","section":"Fig. 3, Fig. 8, §II"},{"comment":"Fig. 7 and the associated text: the claim that the frequency-based reservoir 'can be optimized to improve short-term prediction, a property that random reservoirs lack' is not supported by the protocol used. The random reservoir in Fig. 7f is a single default configuration (fixed α, ρ, ε, σ, degree q), while the frequency-based reservoir is optimized by reshaping the ω_k roster after observing the default's test performance. Random reservoirs are routinely optimized for short-term prediction by tuning exactly these hyperparameters (the manuscript itself cites Bayesian optimization and the noise-resonance work of Zhai et al. [32]); reshaping the roster has the same epistemic status as tuning ρ. The comparison needed is the optimized frequency-based reservoir against a comparably optimized random reservoir (e.g., grid/random search over ρ, α, ε, σ with the same evaluation budget), or a cle","section":"Fig. 7, 'Optimizing the Frequency-based reservoir'"},{"comment":"Methods, Lorenz-63 subsection: the input is stated as u = [x(t), y(t), x(t)]; the third component is presumably z(t) (consistent with the §II discussion of the z-component's dominant spectral peak and with Eqs. 12–14). If the duplication of x is a typo, it must be corrected; if the model was genuinely trained on [x, y, x], the Lorenz-63 results are not forecasting the full system and the attractor reconstruction in Fig. 3f is misleading. Please clarify which was used.","section":"Methods, 'Forecasting the Lorenz-63 chaotic system'"},{"comment":"Eq. (8) and surrounding text: the Jacobian is given as J = (1−α)I + αA, which is exact only if the derivative of tanh is replaced by 1, i.e., linearization at r = 0. Under driving, r(t) is generically nonzero and the actual instantaneous Jacobian is (1−α)I + α diag(sech²(...)) A, so the prescribed eigenvalues (and hence the intrinsic frequencies f_k of Eq. 9) are only nominal. Since the entire interpretability framework hangs on units oscillating at prescribed frequencies, the text should state this approximation explicitly and quantify how much the realized oscillation frequencies deviate from f_k under the chaotic inputs used (this can be read off the Fig. 5/6 heat maps, which would strengthen the paper). Relatedly, the eigenvalue formula below Eq. (8) reads '(1+α) + αν_k ± jαω_k'; the first term should be (1−α).","section":"Methods, Eq. (8)-(9)"}],"minor_comments":[{"comment":"Fig. 4 caption: panel labels are inconsistent — '(c) log(Power) for large amplitude values ε = 0.1' should presumably be ε = 0.5, and panels (g) and (h) are both labeled ε = 0.5 while the text describes (g) as intermediate. Please correct the ε values in the caption.","section":"Fig. 4 caption"},{"comment":"Unit/node bookkeeping is inconsistent throughout: §II defines N = 301 units (601 nodes), but Methods says 'N = 301 units, which corresponds to 2N−1 = 601 units' (should be nodes), and the Fig. 7 subsection says 'the network size was also reduced to N = 301' while the Fig. 7 caption gives N = 150 for panel (b). Please state units vs. nodes unambiguously at each occurrence.","section":"§II; Methods; Fig. 7"},{"comment":"Broken cross-references: the optimization section cites 'Fig.6-(f)' and '(Fig.6)' where Fig. 7(f) and Fig. 7(b/e) are meant; the KS section refers to 'Fig.3-(b)', 'Fig.3-(c)', 'Fig.3-(d)' where Fig. 8(b,c,d) is intended; and Methods says the Mackey-Glass attractor is 'Fig.8-(c)' where Fig. 3(c) is meant.","section":"§II (optimization, KS paragraphs); Methods"},{"comment":"Spelling: 'Kuramoto-Shivanski' (§II) and 'Kuramoto-Sivashinki' (Methods heading) should both be 'Kuramoto-Sivashinsky'. Also 'RESUL TS' heading, 'the magic relies on hyperparameter optimization' (informal), and 'a reservoir computing trains' in the abstract.","section":"Various"},{"comment":"Discussion: the claim '2N ≤ 2p, where N is the number of units and p the number of harmonics' reduces trivially to N ≤ p; presumably a sharper bound was intended. Also 'essential property for brain computation' and similar promotional phrasing could be toned down.","section":"§III Discussion"},{"comment":"The effective-noise discussion (Methods) notes the actual noise intensity is ϵσ scaled by a uniform random factor, i.e., heterogeneous per node; this deserves mention when the optimal-σ results are discussed, since σ spans 10^−7 to 10^−1 across tasks — a five-order-of-magnitude range that is itself a significant hyperparameter sensitivity worth reporting.","section":"Methods; 'Properties of the frequency-based reservoir'"},{"comment":"Related work: the reservoir here is an ensemble of uncoupled driven damped oscillators; the connection to filter-bank / random-feature views of echo-state networks and to classical tapped-delay embeddings would help position the novelty of the 'selective amplification' framing relative to known resonance behavior of driven leaky units.","section":"§I/§III"},{"comment":"References [36] and [52] are the same Strogatz book cited twice; [10] is an arXiv preprint — check whether a published version exists.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"Single-author manuscript with a fair amount of self-citation of the author's prior rhythm work ([34], [35], [49], [55]); the citations are topically relevant, so I do not see a problem, but the editor may wish to note it. Given the forced-oscillator framing and the qualitative-mechanism emphasis, a nonlinear-dynamics venue (e.g., Chaos, PRE) might be a more natural home than a stat.ML venue if the quantitative-comparison comments are not addressed."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a concrete, readable RC design—independent E-I units on a prescribed frequency roster—plus closed-loop forecasts that look competitive with sparse random reservoirs on Mackey–Glass, Lorenz-63, and KS. The useful bit is not “brain-inspired” branding; it is that the recurrent layer is block-diagonal and uncoupled, so you can actually inspect what each unit does and reshape the frequency list when short-term Lorenz prediction fails under the default roster.\n\nWhat is new is narrower than the abstract sells. Oscillatory and harmonic-oscillator reservoirs already exist (Ceni, Kramer, HORN/CoRNN, etc.), and the paper cites them. The contribution is the specific uncoupled frequency roster, the hierarchy-of-timescales framing, the selective-amplification heat maps on periodic and chaotic drives, and the hand reshape that lifts short-term Lorenz performance above the matched random baseline (Fig. 7f, 500 ICs). That last point is the practical hook random reservoirs lack.\n\nSoft spots, in proportion. Selective amplification is shown with pwelch maps, not a forced-oscillator derivation that predicts readout error. The jump from “units store nearby frequencies” to “therefore linear readout forecasts continuous-spectrum chaos” is plausible and consistent with the figures, but still observational. The short-term “optimization” is a one-shot roster reshape after the default failed—not a general design rule. Comparisons use shared hand-chosen α, ε, σ, ρ and matched connection count; that is fairer than nothing, but it does not fully isolate architecture from residual tuning. No code or artifacts. Hyperparameters still free (α, ρ/ω_max, ε, σ, N, noise).\n\nMath and citations look solid for a methods preprint: standard leaky tanh RC, lsqminnorm readout, Wilson–Cowan motivation without overclaiming equivalence. Not circular in the bad sense—targets are held-out closed-loop trajectories.\n\nWho it is for: people who build or interpret reservoirs and neuromorphic oscillators, not general ML. Worth a serious referee. I would bring it to reading group if we are in dynamical ML this month; I would not rush to cite unless I am already writing on interpretable RC. Send to peer review.","headline":"Clean block-diagonal oscillatory reservoir with a usable short-term knob; mechanism is spectral evidence, not a prediction theorem, and novelty sits inside an already busy oscillatory-RC lane.","tokens_in":19265,"tokens_out":572,"would_cite":false,"duration_ms":19161,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Reservoirs built from independent oscillators store input frequencies by selective amplification and can be tuned for short-term chaotic forecasting.","keywords":["reservoir computing","frequency-based reservoir","selective frequency amplification","forced nonlinear oscillators","chaotic time-series prediction","hierarchy of timescales","Wilson-Cowan units","spatiotemporal chaos"],"falsifier":"Build a frequency-based reservoir whose intrinsic frequencies are deliberately mistuned away from the input’s dominant bands and check whether closed-loop short- and long-term forecast error rises sharply relative to a matched random reservoir and to a correctly tuned frequency roster on the same Lorenz or Mackey–Glass task.","tokens_in":19471,"feed_emoji":"🌊","tokens_out":914,"duration_ms":17197,"temperature":0.7,"pith_summary":"Standard reservoir computers leave a random recurrent layer untrained and learn only a linear readout, but the random wiring makes it hard to see what the reservoir is doing or how to tune it. This paper replaces that random layer with an ensemble of independent oscillatory units whose intrinsic frequencies are prescribed, inspired by brain rhythms and timescale hierarchy. Each unit acts like a forced nonlinear oscillator and selectively amplifies input frequency components near its own frequency, storing that portion of the signal for the linear readout. Under matched size and hyperparameters the frequency-based design matches or beats random reservoirs on Mackey–Glass, Lorenz-63, and Kuramoto–Sivashinsky forecasting, and the roster of intrinsic frequencies can be reshaped to improve short-term prediction—something random reservoirs lack a clear handle for. The claim is that reservoir computing can be designed from oscillatory physics rather than black-box randomness.","feed_headline":"Oscillator banks beat random reservoirs at chaotic forecasts","feed_subtitle":"Each unit stores nearby input frequencies; reshaping the roster improves short-term prediction","key_machinery":"Selective frequency amplification: each two-node unit, treated as a nonlinear oscillator forced by the input, preferentially boosts spectral content near its intrinsic frequency (visible in power-spectrum heat maps), so the full reservoir stores the input as a bank of frequency-specific traces that a linear readout can recombine.","core_discovery":"A reservoir whose recurrent layer is an ensemble of independent oscillatory units with prescribed intrinsic frequencies selectively amplifies and stores input frequency components near each unit’s frequency; the stored content is sufficient for linear-readout prediction of chaotic time series, matches or exceeds equivalent random reservoirs, and can be optimized by reshaping the frequency roster to improve short-term forecasts.","pith_inferences":["If unit-level storage is sufficient, the minimal number of oscillators needed for a given chaotic attractor should scale with the number of dominant spectral bands rather than with embedding dimension, a testable scaling law the paper leaves open.","The same frequency-roster design could be dropped into other untrained recurrent architectures (beyond classical reservoirs) wherever interpretability and short-horizon tuning matter.","Noise’s dual role—sustaining fast decaying modes while breaking overly coherent slow modes—suggests an optimal noise schedule that tracks the target’s spectral tilt, not a single scalar intensity."],"forward_implications":["Hyperparameters (leaky rate, spectral radius, input gain, noise) acquire direct physical meaning as scalers of the intrinsic-frequency bank and of the forced-oscillator regime, so they can be chosen from spectral theory rather than black-box search.","Coupling among units is optional; computation occurs at the unit level, which simplifies hardware and neuromorphic realizations to banks of independent oscillators.","Short-term prediction can be improved by deliberately over-representing high intrinsic frequencies when the target has fast content—an optimization path random reservoirs do not expose.","The same construction extends to spatiotemporal chaos (Kuramoto–Sivashinsky), indicating the mechanism is not limited to low-dimensional scalar series."],"fun_headline_variants":["Oscillatory units store nearby frequencies to match random reservoirs","Frequency roster of independent oscillators predicts chaotic series","Reshaping oscillator frequencies improves short-term forecasts","Brain-like frequency banks equal or beat random reservoirs","Forced nonlinear oscillators amplify input frequencies for readout"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That the selective amplification seen in the power spectra is what actually makes the linear readout succeed on continuous-spectrum chaos, and that comparing architectures under the same hand-chosen hyperparameters fairly isolates the design rather than residual tuning.","fun_headline_variants_meta":{"raw":{"variants":["Oscillatory units store nearby frequencies to match random reservoirs","Frequency roster of independent oscillators predicts chaotic series","Reshaping oscillator frequencies improves short-term forecasts","Brain-like frequency banks equal or beat random reservoirs","Forced nonlinear oscillators amplify input frequencies for readout"]},"model":"grok-4.5","effort":"low","cost_usd":0.005654,"raw_usage":{"total_tokens":1514,"prompt_tokens":803,"num_sources_used":0,"completion_tokens":56,"cost_in_usd_ticks":56544000,"prompt_tokens_details":{"text_tokens":803,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":655,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":803,"tokens_out":56,"duration_ms":11440,"temperature":1.0,"reasoning_tokens":655,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T15:19:47.694600+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Build a frequency-based reservoir whose intrinsic frequencies are deliberately mistuned away from the input’s dominant bands and check whether closed-loop short- and long-term forecast error rises sharply relative to a matched random reservoir and to a correctly tuned frequency roster on the same Lorenz or Mackey–Glass task.","supporting_citations":[],"review_version":1}