RAES: Multifaceted Acronym in Research

• RAES is an acronym representing distinct concepts such as residual acoustic echo suppression, threshold-based graph processes, relation attribution explanations, reversible adversarial examples, and autoencoder models, adaptable by context.
• In speech processing, RAES acts as a neural post-filter that refines adaptive filtering by predicting phase-sensitive spectral masks to effectively suppress residual echoes, particularly in double-talk conditions.
• In distributed graph algorithms and generative modeling, RAES supports dynamic expander construction under node churn and enables reversible adversarial example generation and regularized deterministic autoencoding.

RAES is an acronym used for several technically unrelated constructs in contemporary research literature. In arXiv-indexed work, it denotes a real-time residual acoustic echo suppression module in speech communication, a threshold-based distributed graph process descended from “Request a link, then Accept if Enough Space,” and, in closely related spellings such as RAE or RAEs, relation attribution explanations in quantitative bipolar argumentation, reversible adversarial examples for dataset intellectual-property protection, representation autoencoders and regularized deterministic autoencoders in generative modeling, and the UK Research Assessment Exercises in scientometric analysis (Zhou et al., 2020, Angileri et al., 31 Jul 2025, Yin et al., 2024, Xing et al., 2023, Yu et al., 2 Apr 2026, Ghosh et al., 2019, Kousha et al., 2022).

1. Acronymic scope and terminological usage

The literature does not provide a single canonical expansion of RAES. Precise interpretation is domain-dependent, and several usages are only orthographically separated by the singular or plural form RAE/RAEs. In the dataset-protection literature, “RAE” is the paper’s consistent term and “RAES” refers to the same concept; in argumentation, “RAEs” denotes relation attribution explanations; in speech enhancement, “RAES” is explicitly residual acoustic echo suppression (Xing et al., 2023, Yin et al., 2024, Zhou et al., 2020).

Expansion	Domain	Core meaning
Residual Acoustic Echo Suppression	Speech processing	Neural post-filter after adaptive filtering
Request a link, then Accept if Enough Space	Distributed algorithms	Threshold-based bounded-degree expander extraction
Relation Attribution Explanations	Quantitative argumentation	Edge-level attribution of argument strength
Reversible Adversarial Examples	Dataset IP protection	Recoverable adversarially perturbed images
Representation Autoencoders	Generative modeling	Frozen representation encoder plus learned decoder
Regularized Deterministic Autoencoders	Generative modeling	Deterministic autoencoders with explicit regularization
Research Assessment Exercises	Scientometrics	UK peer-review research assessment exercises

This multiplicity creates recurrent ambiguities. A common source of confusion is that RAE/RAES in generative modeling has no connection to RAES in acoustic echo control or to RAES in distributed graph construction. Another is that relation attribution explanations in QBAFs are edge-attribution methods, not adversarial examples or autoencoders. The acronym is therefore best treated as a family of local domain abbreviations rather than a unified research program.

2. Residual acoustic echo suppression in hands-free communication

In acoustic echo control, RAES denotes a second-stage suppressor operating after a conventional adaptive filter. The microphone signal model is

$d(n)=s(n)+y(n),$

where $u(n)$ is the far-end playout signal, $s(n)$ is near-end speech, $y(n)$ is acoustic echo, and $d(n)$ is the microphone signal. A subband NLMS adaptive filter estimates the linear echo $\hat y(n)$, forms the error signal $e(n)=d(n)-\hat y(n)$, and RAES then suppresses the residual echo remaining in $e(n)$. The paper defines RAES as a neural post-filter that takes the AF output signal and far-end reference as input and predicts a phase-sensitive spectral mask to suppress the residual echo in $e(n)$ while preserving near-end speech $s(n)$, especially in double talk (Zhou et al., 2020).

The signal path is STFT-based, with a square-root Hann window, window size $u(n)$0, and 50% overlap. The RAES CNN consumes log-spectrum magnitudes of the AF output error and the far-end reference, stacks $u(n)$1 consecutive frames per signal, forms an original feature tensor of $u(n)$2, and reshapes it to $u(n)$3. The target is a phase-sensitive mask

$u(n)$4

clipped to $u(n)$5, and applied as

$u(n)$6

Architecturally, the backbone is a MobileNetV2-style CNN with inverted residual bottleneck blocks, depthwise convolutions, pointwise $u(n)$7 convolutions, residual connections, and two task-specific heads: a mask head for RAES and a 3-class double-talk detection head. The double-talk states are near-end only, far-end only, and double talk. The auxiliary DTD task is trained with focal loss, and the overall objective combines DTD loss with a mask-regression loss using uncertainty-based multi-task weighting. The paper’s distinctive contribution is the suppression loss, an asymmetric MSE-like criterion that penalizes under-suppression more strongly than over-suppression through a suppression ratio $u(n)$8, with $u(n)$9 and $s(n)$0 evaluated experimentally (Zhou et al., 2020).

This RAES should not be mistaken for a complete acoustic echo canceller. The paper is explicit that it is a post-filter placed after an existing AF, not a replacement for linear AEC. That distinction matters because the method assumes access to both the AF output error $s(n)$1 and the far-end reference $s(n)$2, and its evaluation is framed as improving the nonlinear echo processor stage rather than redesigning the adaptive filter.

The reported implementation is lightweight: 1.2M parameters, 6.9 MFLOPs, and real-time factor $s(n)$3 on a 2.5 GHz x86 CPU for 60 s audio. In single far-end talk, the proposed system reports average ERLE values of $s(n)$4 dB for speech and $s(n)$5 dB for speech+music at $s(n)$6, versus $s(n)$7 dB and $s(n)$8 dB for the AF+DNN baseline. In double talk, for speech-only mixtures at SER $s(n)$9, $y(n)$0, and $y(n)$1 dB, the $y(n)$2 model reports PESQ $y(n)$3, $y(n)$4, $y(n)$5 and STOI $y(n)$6, $y(n)$7, $y(n)$8. The DTD head achieves F1 $y(n)$9 on training and $d(n)$0 on validation (Zhou et al., 2020).

3. Threshold-based graph construction under streaming churn

In distributed graph algorithms, RAES abbreviates “Request a link, then Accept if Enough Space.” In its original static setting, each vertex of a dense $d(n)$1-vertex expander graph seeks exactly $d(n)$2 outgoing links, sends requests to random neighbors, and each recipient accepts all requests in a round if its current in-degree plus the round’s load is at most $d(n)$3; otherwise it rejects all of them. Becchetti et al. showed that, in the static case, RAES terminates in $d(n)$4 rounds with high probability and the final undirected graph is a bounded-degree expander with high probability. The 2025 paper extends this line to a streaming node-churn model and embeds the RAES threshold rule into a dynamic graph process $d(n)$5 (Angileri et al., 31 Jul 2025).

The streaming model is discrete-time. At each round, one new node joins and, after warm-up, the oldest node departs, so every node remains in the system for exactly $d(n)$6 rounds. Edge maintenance is continuous rather than restart-based: a node keeps the same $d(n)$7 request slots throughout its lifetime, and a slot is re-launched if its incident neighbor leaves or a request was rejected. Each pending request targets a uniformly random node in the current vertex set via a minimal link manager, and a node accepts $d(n)$8 requests in round $d(n)$9 iff its in-degree at the start of round $\hat y(n)$0 is at most $\hat y(n)$1. Consequently,

$\hat y(n)$2

This dynamic RAES is analyzed through conductance and vertex expansion. The main structural theorem states that for sufficiently large constants $\hat y(n)$3 and $\hat y(n)$4, and every $\hat y(n)$5, every snapshot $\hat y(n)$6 has constant-conductance behavior on all sets of size at least $\hat y(n)$7, and contains an induced expander $\hat y(n)$8 on $\hat y(n)$9 nodes with high probability. More specifically, Lemma 5.1 and Lemma 5.2 establish vertex expansion at least $e(n)=d(n)-\hat y(n)$0 for sets from $e(n)=d(n)-\hat y(n)$1 up to $e(n)=d(n)-\hat y(n)$2, and Lemma 5.3 gives a subset $e(n)=d(n)-\hat y(n)$3 with $e(n)=d(n)-\hat y(n)$4 such that every $e(n)=d(n)-\hat y(n)$5 with $e(n)=d(n)-\hat y(n)$6 satisfies

$e(n)=d(n)-\hat y(n)$7

A central technical device is the multiple-requests destination lemma, which upper-bounds the joint probability that many correlated requests land in a small set $e(n)=d(n)-\hat y(n)$8 by

$e(n)=d(n)-\hat y(n)$9

This controls the threshold-induced and churn-induced dependencies that make the dynamic analysis nontrivial (Angileri et al., 31 Jul 2025).

A second major result concerns rumor spreading. Under the snapshot-by-snapshot dynamics of $e(n)$0, PUSH and PULL both disseminate a rumor from a freshly joined source node to at least $e(n)$1 nodes in $e(n)$2 rounds with high probability. The leftover $e(n)$3 nodes reflect the tightness of the expansion theorem, since very young nodes can remain isolated for $e(n)$4 rounds with non-negligible probability. The paper also proves control of the pending-request queue: $e(n)$5 for $e(n)$6, implying $e(n)$7 messages per round with high probability and $e(n)$8 expected pending time per request (Angileri et al., 31 Jul 2025).

A persistent misconception is that dynamic RAES simply reruns the static algorithm after every topology change. The paper explicitly rejects that view: RAES is not restarted each time step, but instead runs as a continuous maintenance process over stable request slots and repeated threshold-based accept/reject decisions.

4. Relation attribution explanations in quantitative bipolar argumentation

In quantitative bipolar argumentation, RAEs denote relation attribution explanations: edge-level explanation methods for the final strength of a topic argument in a QBAF

$e(n)$9

This usage is formalized in two complementary strands. One paper studies AAEs and RAEs in Truth-Discovery QBAFs with cyclic support/attack structure, while another develops a general Shapley-value theory of RAEs, their properties, and an approximation algorithm (Yin et al., 2024, Yin et al., 2024).

Two main definitions appear. The removal-based RAE for edge $e(n)$0 and topic argument $e(n)$1 is

$e(n)$2

so it measures the change in $e(n)$3’s strength when only that relation is deleted. The Shapley-based RAE treats edges as players in a coalition game whose value is the topic strength under a restricted edge set: $e(n)$4 The Shapley construction yields efficiency, dummy, symmetry, and dominance properties. Under stability,

$e(n)$5

so the deviation from the base score is fully distributed across relations (Yin et al., 2024).

The theory distinguishes direct, indirect, and multifold edges with respect to a topic. For direct and indirect edges, and under monotonic semantics, RAEs satisfy sign correctness, counterfactuality, qualitative invariability, and quantitative variability. For multifold edges, these properties can fail because multiple paths with different attack/support parity can produce nonlocal interaction effects. This is not a defect in the formalism so much as a statement that edge contributions in cyclic, multiply connected QBAFs are genuinely non-additive (Yin et al., 2024).

The Truth-Discovery application instantiates these ideas in TD-QBAFs, where sources have base score $e(n)$6, claims have base score $e(n)$7, contradictory claims attack each other, and sources and their reported claims support each other bidirectionally. Using QE semantics, the cyclic case study with 17 arguments and 32 edges shows that removal-based and Shapley-based RAEs both identify $e(n)$8, $e(n)$9, $s(n)$0, and $s(n)$1 as the dominant positive contributors to topic $s(n)$2, while edges such as $s(n)$3 and $s(n)$4 contribute negatively. The paper also reports that an argument’s AAE is not necessarily equal to the sum of the RAEs of all its incident edges, directly contradicting a natural but false decomposition intuition (Yin et al., 2024).

Exact RAE computation is exponential in the number of relations. The technical report gives complexity

$s(n)$5

for exact evaluation and proposes a Monte Carlo estimator with complexity

$s(n)$6

where $s(n)$7, $s(n)$8, and $s(n)$9 is the number of samples. The same report demonstrates RAEs in fraud detection and in a small LLM-generated argument graph, where direct supports receive positive attribution, direct attacks negative attribution, and indirect path effects are quantified relation by relation (Yin et al., 2024).

5. Reversible adversarial examples for dataset protection

In dataset intellectual-property protection, the relevant term is RAE, sometimes written RAES. Here a reversible adversarial example is an adversarially perturbed image $u(n)$00 that degrades unauthorized model training or inference while remaining recoverable to an original-like image $u(n)$01 by an authorized party. The RAEDiff framework replaces earlier auxiliary-information schemes with self-generation and self-recovery using a DDPM trained with a Biased Gaussian Distribution and a backdoor trigger (Xing et al., 2023).

The core probabilistic modification is a biased forward diffusion step

$u(n)$02

which induces a marginal with a trigger-dependent mean shift. The same diffusion model is used both to generate protected samples and to recover them. This is a substantive conceptual break from prior RAE schemes based on reversible data hiding or reversible image transform, because reversibility no longer depends on embedded auxiliary metadata. The paper states that the same diffusion model that generated the adversarial perturbation is used to remove it.

RAEDiff introduces permission levels through slightly noised datasets $u(n)$03, protected datasets $u(n)$04, and recovered datasets $u(n)$05. For CIFAR-10, the protected dataset reports SSIM $u(n)$06, PSNR $u(n)$07, and LPIPS $u(n)$08 relative to the original, while the recovered dataset reports SSIM $u(n)$09, PSNR $u(n)$10, and LPIPS $u(n)$11. On CelebA, the protected dataset reports SSIM $u(n)$12 and PSNR $u(n)$13, while the recovered dataset reports SSIM $u(n)$14 and PSNR $u(n)$15 (Xing et al., 2023).

The training-degradation effect is large. For CIFAR-10 with ResNet-18, training on clean-like $u(n)$16 yields test accuracy $u(n)$17, training on protected $u(n)$18 drops to $u(n)$19, and training on recovered $u(n)$20 returns to $u(n)$21. For CelebA, the corresponding values are $u(n)$22, $u(n)$23, and $u(n)$24. In AIGC evaluation, the FID of $u(n)$25 is $u(n)$26 on CIFAR-10 and $u(n)$27 on CelebA, whereas the recovered datasets return to FID $u(n)$28 and $u(n)$29. The recommended setting is $u(n)$30 and $u(n)$31, balancing adversarial strength and recovery fidelity (Xing et al., 2023).

A common misunderstanding is that reversibility in this line of work necessarily requires side-channel information hidden in the image. RAEDiff explicitly positions itself against that assumption: its reversibility is learned in the generative prior, not embedded as auxiliary payload.

6. Autoencoder usages: representation autoencoders and regularized deterministic autoencoders

In generative modeling, RAE denotes two distinct constructs. One is the representation autoencoder of RAE-AR; the other is the regularized deterministic autoencoder introduced as an alternative to VAEs. The shared acronym conceals substantial architectural and theoretical differences (Yu et al., 2 Apr 2026, Ghosh et al., 2019).

RAE-AR defines a representation autoencoder as a frozen pretrained representation encoder $u(n)$32 such as DINOv2, SigLIP2, or MAE, coupled to a learned decoder $u(n)$33: $u(n)$34 The paper argues that direct use of these high-dimensional semantic latents in continuous autoregressive image models is difficult for two reasons: complex token-wise distribution modeling and high-dimensionality amplified exposure bias. It introduces per-token normalization,

$u(n)$35

and Gaussian noise injection during training,

$u(n)$36

to simplify token distributions and make AR prediction robust to training–inference mismatch. Empirically, these modifications close much of the gap to VAE-based AR modeling: DINOv2 improves from gFID $u(n)$37 to $u(n)$38, SigLIP2 from $u(n)$39 to $u(n)$40, and MAE from $u(n)$41 to $u(n)$42 under the combined RAE-AR recipe (Yu et al., 2 Apr 2026).

The 2019 RAE paper uses the acronym differently. A regularized deterministic autoencoder has a deterministic encoder $u(n)$43, a deterministic decoder $u(n)$44, and objective

$u(n)$45

The reconstruction term is paired with a latent norm penalty and explicit decoder regularization such as $u(n)$46, gradient penalty, or spectral normalization. Because the model is deterministic, it dispenses with variational sampling during training and recovers a generative mechanism by ex-post density estimation in latent space, typically a multivariate Gaussian or a $u(n)$47 GMM fit to training codes. The paper reports that these RAEs are able to generate samples comparable to, or better than, those of VAEs and more powerful alternatives on images and structured data, and that ex-post density estimation also improves existing VAEs and WAEs (Ghosh et al., 2019).

These two RAEs should not be conflated. Representation autoencoders rely on frozen semantically pretrained encoders and address AR token modeling; regularized deterministic autoencoders arise from reinterpreting Gaussian-VAE stochasticity as noise injection into a deterministic decoder, then replacing the prior-matching machinery with explicit regularization and ex-post density estimation.

7. Research Assessment Exercises and AI-supported evaluation

In scientometric and policy literature, RAEs are the UK Research Assessment Exercises, predecessors of REF. The review on AI technologies for research assessment does not redefine their institutional mechanics, but synthesizes evidence on how bibliometric indicators and machine learning align with peer-reviewed RAE and REF outcomes. The central conclusion is cautious: metrics may support panels, but should not supplant expert judgement (Kousha et al., 2022).

The evidence base begins with strong departmental-level correlations in some subjects. For the 1992 RAE in Library & Information Science, Oppenheim reported Spearman correlations of $u(n)$48 with total citations and $u(n)$49 with citations per member of staff. For Psychology in the 1996 RAE, Smith and Eysenck reported $u(n)$50 between departmental ratings and mean citations. For the 2001 RAE, Norris and Oppenheim found $u(n)$51 between archaeology rankings and average citations per staff, while Mahdi, D’Este, and Neely reported strong 2001 departmental correlations in Chemistry $u(n)$52, Earth Sciences $u(n)$53, and Physics $u(n)$54, but non-significant associations in many humanities fields (Kousha et al., 2022).

For RAE 2008 and REF 2014, the pattern remains heterogeneous. Butler and McAllister reported useful predictive value for citations in Chemistry and Political Science. Taylor reported correlations of $u(n)$55 between ABS journal quality scores and Business & Management RAE 2008 ratings, and $u(n)$56 for Economics & Econometrics. In REF 2014 article-level analysis, HEFCE found overall correlations with peer review of $u(n)$57 for SCImago Journal Rank, $u(n)$58 for SNIP, $u(n)$59 for field-weighted citation impact, and $u(n)$60 for citations per publication. These were much stronger in Clinical Medicine $u(n)$61, Chemistry $u(n)$62, Physics $u(n)$63, and Biological Sciences $u(n)$64 for older outputs, and much weaker in most social sciences and arts and humanities (Kousha et al., 2022).

The review also emphasizes negative evidence against simplistic automation. In HEFCE’s threshold experiment for predicting whether a 2008 output was 4*, the baseline accuracy from predicting no article as 4* was $u(n)$65, while raw Scopus citations achieved $u(n)$66, FWCI $u(n)$67, and SNIP $u(n)$68. Thus, across all disciplines, no simple global threshold outperformed the trivial baseline. The review accordingly states that no previous published studies had used machine learning to predict the quality scores of individual articles, although machine learning had been used to predict long-term citation counts and statistical methods had been used to predict quality profiles for sets of articles (Kousha et al., 2022).

The broader AI agenda in this literature is operational rather than replacementist. Promising functions include reviewer matching, field classification, plagiarism and statistical checking, and decision-support systems that flag anomalies for panel attention. The review also discusses explainability, transparency, field bias, language bias, prestige effects, and the danger that AI may inherit both bibliometric and peer-review biases. Its policy stance is therefore measured: bibliometric indicators, altmetrics, and AI can assist research assessment, but high-stakes quality judgement remains peer-review led (Kousha et al., 2022).