acceptodds
Under review as a conference paper at ICLR 2027

Reliability-Conditioned Product-Tangent Learning for Degraded Acoustics

Abstract

Synchronized acoustic descriptors offer complementary views of one signal, but they do not fail uniformly: impulsive, co-channel, and additive interference perturb their coordinates, norms, and temporal structure in different ways. A corrupted high-energy descriptor can therefore dominate Euclidean relation learning even when another domain retains useful evidence. We study this failure as reliability-conditioned geometric representation learning. The proposed RMAL CRNN first estimates coordinate reliability, then maps every descriptor domain to its own learned spherical tangent space. Head-specific tangent-metric attention learns cross-domain relations only after this correction, and a factorized nonnegative sparse affinity refines the fused sequence over time. Controlled simulations isolate the response of these stages to individual and compound degradations. Across a public ten-class UAV benchmark and an independent eight-class Glasgow/ACSAC2022 acoustic-identity release, the complete model reaches 86.30% and 93.88% on the ten-class benchmark at −5 and 0 dB, and 81.18% at 0 dB on Glasgow versus 74.28% for CRNN. Multi-seed ablations and Holm corrected paired tests consistently favor the complete path. These results support reliability-before-relation learning for the evaluated descriptors, data, and corruption regimes; they do not imply universal robustness to unseen domains or degradations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Geometric representation, spherical tangent space, manifold attention, heterogeneous feature fusion, robust acoustic recognition

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UgYXOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RMAL-CRNN, a reliability-conditioned representation learning framework for acoustic classification under degraded conditions. The central idea is to estimate coordinate-level reliability before learning cross-domain relations, since different acoustic descriptors can be affected unequally by impulsive noise, co-channel interference, and additive noise. The model first applies a Student-t-inspired reliability weighting to six synchronized acoustic descriptor domains (HHT, LOFAR, FFT, WPT, HHSA, and MFCC), projects each domain into a separate learned spherical tangent space, and then performs cross-domain relation learning using tangent-metric attention. A factorized nonnegative sparse temporal-affinity module is subsequently applied to refine the fused sequence.
The paper evaluates the proposed approach through controlled degradation simulations and two public UAV acoustic-classification benchmarks. Proposed methods are compared against several baselines. The controlled experiments further provide diagnostics intended to isolate the behavior of each component under different degradation mechanisms.

Critical strengths
  1. Strongly structured analysis of the proposed mechanism
    The paper does more than report aggregate classification gains: it explicitly isolates the behavior of reliability weighting, tangent-metric attention, and sparse temporal affinity through controlled degradation experiments and structural ablations. The multi-seed paired tests with Holm–Bonferroni correction further strengthen the evidence that the complete model consistently outperforms its structural variants. This makes the proposed “reliability-before-relation” principle substantially more interpretable than a black-box architectural improvement.

  2. Consistent improvements across two independent acoustic benchmarks
    RMAL-CRNN consistently outperforms the listed baselines across a range of SNR conditions on both the Kielce and Glasgow/ACSAC 2022 benchmarks. In particular, the improvement remains substantial at low SNR, suggesting that the proposed reliability-conditioned fusion is useful for the intended degraded-acoustic classification setting rather than only improving performance under clean conditions.

Critical weaknesses
  1. The empirical validation is focused on a relatively specific acoustic setting
    The proposed reliability-before-relation framework is evaluated primarily with six predefined acoustic descriptor domains and UAV acoustic classification benchmarks under the considered degradation conditions. While the controlled experiments provide useful evidence for the proposed mechanism, it remains less clear how well the approach would transfer to substantially different acoustic representations, tasks, or unseen degradation conditions. This does not undermine the reported results, but additional cross-setting validation would strengthen the broader generality of the proposed framework.
Questions for authors

Q1. Contribution of the tangent-space formulation
The proposed framework combines reliability weighting with product-tangent representation and tangent-metric attention. Could the authors provide further analysis or discussion on how much of the performance improvement is specifically attributable to the tangent-space formulation, compared with applying a conventional Euclidean attention/fusion mechanism after the same reliability weighting? This would help clarify the unique contribution of the geometric formulation.

Q2. Dependence on the selected acoustic descriptors
The experiments use six complementary acoustic descriptor domains (HHT, LOFAR, FFT, WPT, HHSA, and MFCC). How sensitive is the proposed framework to the particular choice and number of descriptor domains? An analysis using subsets of the descriptors or alternative descriptor combinations would help clarify whether the proposed product-tangent fusion is broadly applicable to heterogeneous acoustic representations.

Reviewer_Q2PPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes RMAL-CRNN (Robust Manifold-Attention Low-Rank Convolutional Recurrent Network) for acoustic classification from K=6K = 6 synchronized 32-dimensional handcrafted descriptor domains (HHT, LOFAR, FFT, WPT, HHSA, MFCC) under acoustic degradation. The architecture orders three front-end modules before a Conv-BiLSTM classifier: (1) Geometric Student-tt Tangent-Subspace Projection (GST), which reweights per-domain coordinates using Student-tt-inspired gates and orthogonally projects aligned domain vectors onto learned spherical tangent hyperplanes TpkSd−1\mathcal{T}_{\mathbf{p}_k}\mathbb{S}^{d-1}; (2) Learned Manifold Attention (LMA), which projects normalized per-head queries, keys, and values onto head tangent planes and computes negative squared Euclidean distance attention logits; and (3) Factorized Nonnegative Sparse Affinity Refinement (FNSAR), which forms a rank-rℓr_\ell temporal affinity matrix followed by shifted ReLU thresholding and gated LayerNorm residual mixing. Stage-level diagnostics are shown on a synthetic rotating-source simulator, and classification accuracy is reported on two multirotor UAV audio datasets (10-class Kielce and 8-class Glasgow/ACSAC 2022) across additive Gaussian SNR levels.

Critical strengths
  1. Practically Important Problem & Well-Motivated Core Premise ("Reliability-Before-Relation Learning"): Acoustic recognition and individual device authentication in severely degraded outdoor environments (low SNR, impulsive transients, and co-channel interference) is an important, challenging problem. The paper's high-level conceptual premise—that when fusing heterogeneous, noncommensurate signal descriptors (HHT, LOFAR, FFT, WPT, HHSA, MFCC) that fail asymmetrically under noise, coordinate-level reliability gating should precede cross-domain attention so that high-magnitude corrupted descriptors are not mistaken for reliable features—is intuitive and well-motivated.
  2. Modular Stage-Wise Decomposition: Organizing the front end into three distinct functional stages (per-domain reliability gating via GST, bounded distance-based cross-domain attention via LMA, and temporal affinity refinement via FNSAR) paired with targeted stage diagnostics and multi-seed paired ablation tests (Tables 2 and 4) provides a clear conceptual structure.
Critical weaknesses
  1. Unacceptable Hallucinated References and Fabricated Author Lists in the Bibliography (Sufficient Grounds for Rejection by Itself): It is unacceptable to have hallucinated references and fabricated author lists in a scientific submission (especially when the AI Use Statement on page 9 explicitly claims that "Generative AI assisted with ... literature checks ... and the authors verified all assisted content"), and in my judgment this by itself implies rejection. Cross-checking the 30-entry bibliography reveals 7 references (23.3%) with LLM-fabricated/hallucinated author names—including a completely fabricated author list for an ICLR 2025 paper and 6 entries where author first initials were expanded into wrong given names—plus misattributed or missing baseline/dataset citations:
    • Wang et al. (2025) (TimeMixer++, ICLR 2025): Completely fabricated author list—cites "Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Mingsheng Long, and Jianmin Wang". The actual authors of TimeMixer++ (ICLR 2025, arXiv:2410.16032) are Shiyu Wang, Jiawei Li, Xiaoming Shi, Zhou Ye, Baichuan Mo, Wenze Lin, Shengtong Ju, Zhixuan Chu, and Ming Jin (zero overlap with the cited author list—the bibliography conflated the TimesNet/TimeMixer Tsinghua authors with TimeMixer++!).
    • Hou et al. (2024) (IEEE/ACM TASLP, 32:68–82, doi:10.1109/TASLP.2023.3323135): Hallucinates 3 first names—cites "Yuchen Hou, Boseong Kang, Audrius Mitchell..." instead of Yuanbo Hou, Bo Kang, Andrew Mitchell (expanded from Y. Hou, B. Kang, A. Mitchell).
    • Tan et al. (2024) (IEEE/ACM TASLP, 32:1286–1297, doi:10.1109/TASLP.2024.3353578): Hallucinates 3 first names—cites "Yizhen Tan, Haoyu Ai, Shanshan Li..." instead of Yizhou Tan, Hao Ai, Shengchen Li (expanded from Y. Tan, H. Ai, S. Li).
    • Yang & Li (2024) (IEEE/ACM TASLP, 32:4211–4225, doi:10.1109/TASLP.2024.3458811): Hallucinates both first names—cites "Bo Yang and Xuelong Li" instead of Bing Yang and Xiaofei Li (expanded from B. Yang and X. Li).
    • Fang et al. (2022) (JASA, 151(2):968–981, doi:10.1121/10.0009365): Hallucinates 2 first names and truncates the 4th author—cites "Jun Fang, Anthony Finn, Rick Wyber, et al." instead of Jian Fang, Anthony Finn, Ron Wyber, and Russell S. A. Brinkworth (expanded from J. Fang, A. Finn, R. Wyber).
    • Bresciani et al. (2023) (JASA, 154(2):1179–1190, doi:10.1121/10.0020666): Hallucinates the second author's first name—cites "Jean Maillard" instead of Julien Maillard (expanded from J. Maillard).
    • Chakraborty et al. (2022) (ManifoldNet, IEEE TPAMI, 44(2):799–810, doi:10.1109/TPAMI.2020.3003846): Hallucinates the second author's first name—cites "Javier Bouza" instead of Jose Bouza (expanded from J. Bouza).
    • Misattributed Baselines & Uncited Primary Dataset: In Table 1, MAtt (actually Pan et al., NeurIPS 2022) is wrongly cited to Chakraborty et al., 2022; Vaswani et al., 2017; MambaSL (actually Jung & Kim, ICLR 2025) is wrongly cited to Gu & Dao, 2023; and the primary Kielce 10-class UAV dataset has no bibliography entry at all.
  2. Explicit Admissions of an Unpreserved, Non-Reproducible Experimental Record: The manuscript explicitly admits across five sections that the authors do not possess basic experimental metadata required to verify or reproduce the study:
    • Reproducibility Statement (Page 9, lines 465–467): "Exact dataset counts, a persistent locator, the complete training schedule, and UMAP settings are unavailable and are not claimed; no public-code claim is made."
    • Appendix A.2 (Page 12, lines 599–600): "exact simulator sampling distributions and parameter ranges were not preserved with the experimental record and are not reconstructed here."
    • Appendix A.3 (Page 12, lines 611–614): "A formal bibliographic record, persistent access URL, exact split counts, and redistribution terms were not preserved with the experimental record."
    • Appendix D (Page 12, lines 645–647): "Primary citations or complete external configurations for the implementation labels MAtt, CNN-Transformer, and MambaSL were not preserved with the experimental record."
  3. Disconnect Between Compound-Corruption Claims (Q1) and SNR-Only Benchmarks (Q2–Q3), Plus Quantitative/Statistical Discrepancies:
    • Although GST and LMA are specifically motivated by sparse impulses (κ\kappa) and co-channel interference (SIR\text{SIR}), neither impulses nor co-channel interference are applied to the Kielce or Glasgow classification benchmarks (lines 300–302), and no classification metrics are reported on the Q1 simulator.
    • Table 1a vs. Table 2/4 Mismatch & Undisclosed Seed Count Change: At 0 dB0\text{ dB} on Kielce, CRNN is 87.15±0.3387.15 \pm 0.33 in Table 1a (n=4n = 4 seeds, disclosed only on page 13 line 689) vs. 86.93±0.3586.93 \pm 0.35 in Table 2 (RMAL-CRNN is 93.88±0.3693.88 \pm 0.36 vs. 93.89±0.3393.89 \pm 0.33), where t/dz=7=2.6458t / d_z = \sqrt{7} = 2.6458 in Table 4 reveals n=7n = 7 seeds without explanation.
    • 4×4\times Parameter Count Contradiction: Line 368 claims all ablations span 333.3k–339.4k333.3\text{k}\text{--}339.4\text{k} parameters (only 6.1k6.1\text{k} parameters for GST + LMA + FNSAR), whereas the layer dimensions specified in Section 3 and Appendix D (66 matrices Ak∈R32×32=6,144\mathbf{A}_k \in \mathbb{R}^{32\times 32} = 6{,}144; Wq,Wk,Wv,Wo∈R32×32=4,096\mathbf{W}_q, \mathbf{W}_k, \mathbf{W}_v, \mathbf{W}_o \in \mathbb{R}^{32\times 32} = 4{,}096; FNSAR 192×64192 \times 64 projection + Wu,Wv∈R8×64=13,312\mathbf{W}_u, \mathbf{W}_v \in \mathbb{R}^{8\times 64} = 13{,}312; plus biases/embeddings) require ∼24,750\sim 24{,}750 parameters (4×4\times larger).
    • Figure 4 Caption Mismatch: Figure 4's caption (line 421) states "Percentages in parentheses are the reported overall accuracies", but no percentages or descriptor names appear on subplots (a)–(h).
  4. Mathematical & Geometric Flaws in GST, LMA, and FNSAR (Section 3 & Appendices E–G):
    • Eq. (8) Erases Both Geodesic Distance and Reliability Scale: Proposition S.1 relates ut(k)=Πpkht(k)\mathbf{u}_t^{(k)} = \mathbf{\Pi}_{\mathbf{p}_k}\mathbf{h}_t^{(k)} to the spherical log map log⁡pk(υt(k))\log_{\mathbf{p}_k}(\boldsymbol{\upsilon}_t^{(k)}) (whose norm is the geodesic distance θt,k\theta_{t,k}). Yet Eq. (8) divides ut(k)\mathbf{u}_t^{(k)} by ∥ut(k)∥2+ϵ\|\mathbf{u}_t^{(k)}\|_2 + \epsilon—destroying both the geodesic angle θt,k\theta_{t,k} and any domain-level reliability attenuation from GST—and adds a bias Πpkbk\mathbf{\Pi}_{\mathbf{p}_k}\mathbf{b}_k that breaks collinearity with log⁡pk(υt(k))\log_{\mathbf{p}_k}(\boldsymbol{\upsilon}_t^{(k)}). Moreover, Eq. (4) lacks positivity constraints for σk,νk\boldsymbol{\sigma}_k, \nu_k and applies uncentered shrinkage gt(k)⊙xt(k)\mathbf{g}_t^{(k)} \odot \mathbf{x}_t^{(k)} (which scales clean coordinates at xt(k)=μk≠0\mathbf{x}_t^{(k)} = \boldsymbol{\mu}_k \neq \mathbf{0} by 1+1/νk>11 + 1/\nu_k > 1).
    • Immediate Exit from the Product Tangent Space: Section 3.4 adds an unconstrained embedding ek∈Rd\mathbf{e}_k \in \mathbb{R}^d (pk⊤ek≠0\mathbf{p}_k^\top \mathbf{e}_k \neq 0), leaving TpkSd−1\mathcal{T}_{\mathbf{p}_k}\mathbb{S}^{d-1}, and applies shared projection matrices Wqh,Wkh,Wvh\mathbf{W}_q^h, \mathbf{W}_k^h, \mathbf{W}_v^h before Euclidean LayerNorm.
    • Post-ReLU Rank Inflation in FNSAR: Entrywise C=ReLU⁡(Craw−ϑ11⊤)\mathbf{C} = \operatorname{ReLU}(\mathbf{C}_{\text{raw}} - \vartheta\mathbf{1}\mathbf{1}^\top) in Eq. (14) breaks the rank-rℓr_\ell bound of Proposition S.3.1, requires O(T2)O(T^2) complexity, can yield zero rows when max⁡sCraw,ts≤ϑ\max_s C_{\text{raw},ts} \le \vartheta, and does not bound rank⁡(Fout)≤rℓ=8\operatorname{rank}(\mathbf{F}^{\text{out}}) \le r_\ell = 8 in Eq. (15) (contradicting Fig. 7c–d).
Questions for authors
  1. Hallucinated Bibliography Entries (Clarification): How did 7 of the 30 references acquire fabricated author names—including attributing TimeMixer++ (ICLR 2025) to 6 completely unrelated authors (Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Mingsheng Long, Jianmin Wang instead of Shiyu Wang et al.) and expanding author initials into wrong given names across 6 other papers (Jean Maillard, Javier Bouza, Jun Fang / Rick Wyber, Yuchen Hou / Boseong Kang / Audrius Mitchell, Yizhen Tan / Haoyu Ai / Shanshan Li, Bo Yang / Xuelong Li)?
  2. Unpreserved Experimental Record (Clarification): Why are the dataset split counts, Kielce dataset citation, simulator parameter distributions, UMAP settings, training schedules, and baseline configurations/citations for MAtt, CNN-Transformer, and MambaSL described as "not preserved with the experimental record"?
  3. Parameter Count & Seed Discrepancies (Clarification): How can GST + LMA + FNSAR add only 6.1k6.1\text{k} parameters (333.3k–339.4k333.3\text{k}\text{--}339.4\text{k}, line 368) when the matrices defined in Section 3 and Appendix D total ∼24.8k\sim 24.8\text{k} parameters? And why does Table 1a use n=4n = 4 seeds (CRNN =87.15%= 87.15\%) while Tables 2 and 4 use n=7n = 7 seeds (CRNN =86.93%= 86.93\%) at 0 dB0\text{ dB} on Kielce?
Reviewer_NFmnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addressed the problem of acoustic classification from several synchronized handcrafted descriptors per frame. The authors argue that impulses, co-channel interference, and noise corrupt the descriptors unequally, so a damaged high-energy descriptor can dominate fusion in a shared Euclidean space. The proposed method consists of three stages. GST weights each coordinate with a Student-t-style reliability score and projects each descriptor onto its own learned spherical tangent space. LMA then computes attention across descriptors using negative squared distance in head-specific tangent planes, which keeps the logits bounded. FNSAR builds a low-rank, thresholded temporal affinity over frames before a CRNN classifier. Classification is evaluated on two public UAV benchmarks under additive noise at several SNRs.

Critical strengths

The motivation of the problem is sensible. Different descriptors of the same signal do fail at different rates under different corruptions, and estimating reliability before fusing them is a reasonable approach. Each stage is paired with a controlled diagnostic that checks the behavior it was designed for, for example, that the reliability weights drop for the descriptors most affected by impulses, rather than only reporting final accuracy.

Critical weaknesses

The evaluation does not currently test the problem the method is built for sufficiently. The method is motivated by unequal and compound corruption, such as impulses and co-channel interference, but all classification results use additive noise at fixed SNRs. The simulator that produces compound corruption is used only for diagnostics and visualizations, not for classification. In the ablations, components are removed but not replaced with standard alternatives. As a result, the main claim that reliability-before-relation improves classification under compound degradation is not sufficiently validated from the reviewer’s point of view.

Questions for authors
  1. Can the authors report classification accuracy under impulsive, co-channel, and compound corruption, at least in the controlled simulation, for RMAL-CRNN, CRNN, and the strongest baselines?
  2. Please provide the details of dataset sizes and split counts for both benchmarks.
  3. The abstract compares Glasgow results with CRNN. However, the strongest baseline in Table 1(b) is SEResNet1D (78.50 at 0 dB). Please report the margin against the strongest baseline, and explain why the baseline sets differ between the two benchmarks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.