{"id":"cd54d225-7549-4b70-afb6-3ff6d686ee65","arxiv_id":"2411.14489","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GhostRNN compresses RNN hidden states by generating ghost states from a small set of intrinsic states with cheap linear operations, cutting parameters by about 40% with similar accuracy.","lead":"GhostRNN is a new RNN architecture that shrinks the hidden state by generating some of its dimensions with cheap operations instead of full matrix multiplications. On keyword spotting and speech enhancement benchmarks it keeps accuracy roughly unchanged while cutting parameters and computation by about 40%.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of 'similar performance' is supported only by point estimates with no reported variance; the 0.1% accuracy and 0.1 dB margins are within typical seed noise.","rationale":"I read the paper as an empirical proposal: GhostRNN is a parameter-efficient RNN architecture whose value is demonstrated by matched-budget comparisons. The reader's verdict (CONDITIONAL) already flags missing error bars, but I consider this the single most load-bearing concern because it gates the central claim directly. The regeneration assumption (Eq. 9) is theoretically important, but even if it is only partially true, the experiments could still show similar performance; conversely, without variance information, no conclusion can be drawn from the current numbers. The paper's own statement that KWS experiments were repeated three times suggests variance data exists and should be reported. A concrete re-run with confidence intervals would settle whether the reported margins are meaningful. I partially agree with the reader's weakest_assumption: the regeneration concern is real, but the empirical reliability is the prerequisite for evaluating it.","tokens_in":8161,"tokens_out":15375,"duration_ms":163932,"concrete_test":"Run the KWS experiment comparing GRU-498K and GhostRNN-292K, and the SE experiment comparing DCCRN-GRU128 and DCCRN-Ghost128, with at least 10 random seeds each under identical hyperparameters. Compute the per-seed paired differences and report the mean, standard deviation, and 95% confidence interval for each metric (accuracy for KWS; SDR, Si-SDR for SE). If the confidence interval for the difference (GhostRNN minus baseline) excludes a drop larger than 0.5% accuracy or 0.2 dB SDR, the 'similar performance' claim is supported; otherwise it is not.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that GhostRNN reduces parameters by ~40% 'while keeping performance similar' (abstract, conclusions). The supporting evidence is Table 1 (KWS: 94.79% for GhostRNN-292K vs 94.68% for GRU-498K, a +0.11 pt difference) and Table 2 (SE: GhostRNN128 within 0.06 dB SDR of GRU128; GhostRNN512-TasNet +0.12 dB over GRU512-TasNet; GhostRNN192 +0.13 dB over GRU136-TasNet). No standard deviations, confidence intervals, or significance tests are reported anywhere. The KWS section states that 'each model was trained with the same configuration for three times and the average experiment results are reported here,' but the individual runs and their spread are omitted; the SE results appear to be single runs. For speech tasks, a 0.1% accuracy or 0.1 dB difference is well within run-to-run variance (e.g., seed changes, data shuffling). Thus the observed differences are statistically indistinguishable from zero, and the actual performance of GhostRNN could be worse than the baseline by a margin that contradicts 'similar performance.' Because the entire contribution of the paper is an empirical efficiency claim, this missing variance reporting is the load-bearing weak point: if the point estimates move by one standard error, the central claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"GhostRNN proposes to compress RNNs by exploiting redundancy in hidden states. The method keeps a small set of 'intrinsic' states computed by a vanilla GRU/LSTM recurrence, then generates 'ghost' states via a cheap operation (a linear layer plus activation, Eq. 9), concatenating intrinsic and ghost states as the input to the next time step. The parameter-count analysis (Eqs. 10-12) shows a reduction by roughly the ratio r of the total state size to the intrinsic state size. Experiments on keyword spotting (Google Speech Commands) and speech enhancement (LibriMix, with DCCRN and GRU-TasNet baselines) report that GhostRNN matches or slightly exceeds baseline accuracy/SDR at roughly 40% fewer parameters and MACs. The paper concludes that GhostRNN reduces memory and computation while keeping performance similar.","tokens_in":8436,"tokens_out":12803,"duration_ms":122667,"significance":"If the empirical claims hold, GhostRNN is a simple and broadly applicable plug-in compression module for RNN-based speech models, and the extension of the cheap-operation principle from CNNs (GhostNet [16]) to recurrent hidden states is a reasonable and non-obvious step. The parameter-count derivation in Sec. 2.3 is correct and checkable by hand, and the compression is by construction; the matched-parameter comparisons in Tables 1-2 are an appropriate experimental design; and the redundancy analysis in Sec. 2.2 provides a concrete, falsifiable motivation. The contribution is currently limited, however, by the absence of any variance estimate for the headline margins (~0.1% accuracy, ~0.1 dB SDR), by the lack of measured memory/latency to back the abstract's 'memory usage' phrasing, and by the single-model basis of the redundancy observation. With the requested statistical and ablation support, the paper would be a useful contribution to efficient speech modeling.","major_comments":[{"comment":"The paper's central claim is that GhostRNN 'keeps performance similar' while compressing parameters by ~40%, but the supporting evidence is a set of point estimates with no variance: the KWS section states that each model was trained three times and averages are reported (Sec. 3.2), yet the individual runs and their spread are never shown, and the SE results in Table 2 appear to be single runs. All reported margins (accuracy +0.11 percentage points over the 498K GRU; SDR differences of +0.12, +0.13, and -0.06 dB across Table 2) are of a size that typical seed-to-seed variation in speech tasks can reverse. Because the entire contribution is an empirical efficiency claim, please report per-seed results, standard deviations or confidence intervals, and ideally a paired significance test for the KWS and SE comparisons; without this, the headline conclusion is statistically unsupported.","section":"Tables 1-2; Sec. 3.2, 3.4-3.5"},{"comment":"The proposed architecture is under-specified: the value of the ratio r (or, equivalently, the intrinsic-state dimension) and the form of the cheap operation phi used in the experiments are never stated. The reader cannot determine how GhostRNN-292K, DCCRN Ghost128, or GhostRNN512-TasNet were configured, which prevents reproduction and independent verification. Please report r, the intrinsic and ghost dimensions, and the exact phi used for each experiment, and add a sensitivity analysis over r (e.g., r = 2, 4, 8) so that the trade-off between compression and performance is quantified.","section":"Sec. 3; Eqs. (9)-(12)"},{"comment":"The abstract states that GhostRNN 'significantly reduces the memory usage (~40%) and computation cost,' but Tables 1-2 report only parameter counts and MACs; no actual memory footprint, peak activation memory, latency, or energy is measured. Since the stated motivation is deployment on low-resource devices, parameter-count-based inference is a legitimate but incomplete proxy: the RNN hidden state itself is tiny relative to model weights, and MACs do not capture operational memory traffic. Please either measure actual memory and runtime on the target hardware, or rephrase the claims to refer explicitly to parameter count and MACs.","section":"Abstract; Sec. 3.4-3.5"},{"comment":"The key architectural assumption — that the redundant dimensions of a hidden state can be regenerated from the intrinsic states by a learned linear transform plus activation without losing task-relevant information — is motivated by a PCA and cosine-similarity analysis of a single trained model (one DCCRN GRU with 128 units, Sec. 2.2). The paper provides no evidence that this regeneration assumption holds across architectures, layer depths, or scales, which creates a circularity risk because the observation and the validation use the same task family. Please add either (i) reconstruction-error or downstream-task measurements of the ghost operation on different RNN variants (e.g., LSTM, different GRU widths), or (ii) ablations comparing the proposed phi against alternative cheap operations and against a matched-parameter plain GRU, to establish generality.","section":"Sec. 2.2; Eq. (9)"}],"minor_comments":[{"comment":"The candidate equation contains a parenthesis imbalance (\"rt * (Whch(t-1) + bhc) + Wgcgt-1 + bgc)\"); please rewrite it so the scope of the tanh is unambiguous.","section":"Eq. (7)"},{"comment":"The initialization of the ghost state g0 at the first time step is never defined; please specify g0 (e.g., zeros).","section":"Sec. 2.2"},{"comment":"The text states that the total parameter count 'will be compressed by the factor r'; this is only an upper bound, since Eqs. (11)-(12) give a compression ratio strictly less than r for finite feature dimension. Please rephrase to 'approximately r' or 'up to r' for consistency with the formulas.","section":"Sec. 2.3"},{"comment":"Typographical errors: 'vanila' (Sec. 2.3), 'vanllia' (Conclusions), 'roubustness' (Sec. 3.1), and 'Simliar' (Introduction); also, the notation 'statei' in Sec. 2.2 is inconsistently formatted.","section":"Various"},{"comment":"The claim of a '0.1% accuracy improvement' is better stated as +0.11 percentage points (about 0.12% relative); also clarify whether the reported parameter counts include the classification head and feature extractor, since the baseline GRU sizes (498K/295K) do not obviously match a pure 10-dimensional-input GRU.","section":"Table 1"},{"comment":"The text refers to the 'librimix1 dataset' while Sec. 3.1 describes LibriMix; please use the dataset name and the train-360 subset designation consistently.","section":"Sec. 3.5"},{"comment":"The paper says GhostRNN 'can be applicable to other RNNs, e.g., LSTM,' but no LSTM experiment appears; either add an LSTM result or soften the claim, since the conclusions also list LSTM extension as future work.","section":"Sec. 2, Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is formatted as a short conference paper (five pages plus references), and its empirical depth is at conference level; for a journal venue, the requested variance reporting and measured efficiency metrics are essential. The heavy reliance on the authors' own GhostNet [16] is legitimate and disclosed, but the novelty should be positioned as the transfer of cheap operations to recurrent hidden-state compression rather than as a new compression principle. No concerns about citation misconduct; the self-citation is relevant to the method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable, clearly-written paper that does exactly what it says — adapts the GhostNet cheap-operation trick to RNN hidden states. The parameter-count math is right, and the matched-budget comparisons are the correct way to evaluate a compression method. The new bit is the empirical observation that trained GRU hidden states have a lot of cosine similarity and that about half the singular values carry 99% of the PCA energy. That justifies the design.\n\nWhere it gets soft is the evidence for the headline claim. The KWS table shows GhostRNN 292K at 94.79% vs a 498K GRU at 94.68% — about a 0.1 point gap, which is well inside the kind of seed-to-seed variation anyone working on Speech Commands expects. The text says three runs were averaged, but the individual runs and spread are not shown. The SE tables look like single runs, and the differences are again around 0.1 dB. With no standard deviations, confidence intervals, or significance tests, the observed differences are statistically indistinguishable from zero. That does not mean the method fails; it means the central claim of 'similar performance' is not actually supported. The memory savings are also inferred from parameter counts, not measured, which matters for edge deployment.\n\nThe self-citation to GhostNet is fine — the idea is directly inherited, so citing it is appropriate. The novelty is modest but real: applying the cheap-operation pattern inside the GRU recurrence is a reasonable extension, and the empirical characterization of RNN redundancy is a useful data point.\n\nWho gets value: researchers working on efficient RNNs for edge devices, particularly speech tasks. If the authors release code and seed-level results showing the variance is small, the method becomes a solid plug-in. Without that, it's a plausible idea with an under-supported empirical claim. I'd send it to peer review, but the review should focus on the statistical rigor. Conditional acceptance, not a desk reject.","headline":"A clean, plausible adaptation of GhostNet to RNN hidden states, but the empirical case rests on point estimates with no variance; needs error bars before I'd trust the 'similar performance' claim.","tokens_in":8980,"tokens_out":2404,"would_cite":false,"duration_ms":23808,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that an RNN's hidden state is partly redundant, and that the redundant part can be regenerated from a smaller intrinsic set by cheap operations, giving ~40% compression with no accuracy loss.","keywords":["GhostRNN","RNN compression","hidden state redundancy","cheap operations","ghost states","keyword spotting","speech enhancement","GRU"],"falsifier":"Train the same keyword spotting and speech enhancement models on tasks whose hidden states are known to be irreducibly high-dimensional, or measure the PCA contribution rate across a variety of RNN architectures: if a well-trained RNN shows near-zero cosine similarity among most dimensions and needs close to 100% of singular values for 99% energy, then generating ghost states from a half-sized intrinsic set should noticeably degrade accuracy or SDR, and the roughly 40% compression would fail to match the baseline.","tokens_in":7934,"feed_emoji":"🎙️","tokens_out":4080,"duration_ms":41568,"temperature":0.7,"pith_summary":"This paper tries to show that the hidden states of recurrent neural networks contain redundant dimensions that can be exploited for model compression on speech tasks. It proposes GhostRNN, in which a small number of intrinsic states are computed by the usual RNN equations, and the remaining ghost states are produced by cheap learned linear transformations plus an activation. On keyword spotting and speech enhancement, the resulting models keep accuracy and signal quality essentially unchanged while using about 40% fewer parameters and fewer FLOPs. If the redundancy is as generic as the paper suggests, this gives a simple way to fit competitive RNNs into low-resource devices.","feed_headline":"RNN state redundancy is exploitable: GhostRNN cuts memory ~40%","feed_subtitle":"A small intrinsic state plus learned linear transforms reproduces the rest, keeping accuracy on speech tasks.","key_machinery":"The load-bearing object is the ghost-state generation equation $g_t = \\phi(h_t)$, where $\\phi$ is a cheap operation composed of a learned linear layer and an activation function. This is what converts a smaller intrinsic hidden vector into the concatenated full state $[h_t; g_t]$ carried to the next time step. The machinery also includes the parameter-count identity $\\text{Param}_\\phi = (D/r)(D - D/r)$, which keeps the extra cost quadratic in the intrinsic dimension but divided by $r^2$ relative to a full GRU matrix. This identity is what guarantees the roughly $r$-fold reduction in parameters and MACs.","core_discovery":"The central discovery is empirical: in a trained GRU hidden-state trajectory, roughly half of the singular values account for 99% of the PCA energy, and several hidden-state components have cosine similarity near 1 while others are near 0. The paper reads this as partial redundancy in the state space. GhostRNN formalizes the exploitation by computing an intrinsic state $h_t$ of dimension $D/r$ with the standard GRU equations, then generating a ghost state $g_t = \\phi(h_t)$ through a learned linear map followed by an activation, and feeding the concatenation $[h_{t-1}, g_{t-1}]$ into the next time step. This replaces a fraction of the recurrent matrix multiplications with much cheaper operations. Across GRU-style keyword spotting and speech enhancement models, a model with roughly 40% fewer parameters matches or slightly beats its larger baseline, including about 0.1% higher accuracy on Google Speech Commands and about 0.1 dB higher SDR and Si-SDR than matched-parameter baselines.","pith_inferences":["Editorial inference: the PCA and cosine-similarity redundancy signature is measured on one DCCRN GRU with 128 hidden units, so the 2:1 intrinsic-to-ghost split may not be optimal for other sizes; the ratio $r$ could be treated as a tunable hyperparameter.","Editorial inference: if the redundancy is a property of gradient-trained recurrent dynamics rather than of speech specifically, the same compression should transfer to RNNs for automatic speech recognition, acoustic echo cancellation, and other sequence tasks, which would be a direct test.","Editorial inference: a sharper comparison would pit GhostRNN against simply halving the hidden size with a wider input projection, isolating whether the ghost-state structure itself, rather than the parameter reduction, drives the preserved performance."],"forward_implications":["At roughly 40% parameter compression, GhostRNN matches or slightly exceeds the 500K GRU on Google Speech Commands (94.79% vs 94.68% accuracy).","Because ghost states are produced once per time step and reused in all subsequent recurrences, per-step FLOPs drop approximately in proportion to the parameter reduction.","The construction is not tied to GRU: the same intrinsic-plus-cheap-ghost pattern is claimed to apply to LSTM and other RNN variants.","When the cheap operation is parameter-free, the compression ratio can reach the full factor $r$; with a learned linear layer the saving is slightly less but still large."],"supporting_citations":[{"why":"Supplies the cheap-operations idea of generating ghost features from intrinsic features, which GhostRNN adapts to RNN hidden states.","marker":"[16]"},{"why":"Provides the DCCRN speech enhancement baseline whose GRU hidden states are analyzed for redundancy and then compressed.","marker":"[5]"},{"why":"Provides the GRU-TasNet architecture used as the second speech enhancement baseline.","marker":"[24]"},{"why":"Provides the keyword spotting dataset setup, training recipe, and the GRU baselines used in Table 1.","marker":"[3]"},{"why":"Defines the Google Speech Commands dataset used for the keyword spotting experiments.","marker":"[18]"},{"why":"Defines the LibriMix dataset used to build the noisy speech enhancement evaluation set.","marker":"[20]"},{"why":"Supplies the Li-GRU compressed RNN baseline that GhostRNN is compared against.","marker":"[12]"},{"why":"Supplies the SITGRU compressed RNN baseline that GhostRNN is compared against.","marker":"[14]"}],"fun_headline_variants":["GhostRNN cuts RNN memory 40% by reusing hidden states","RNN redundancy exploited: GhostRNN shrinks state, keeps accuracy","Cheap ops turn few RNN states into many, saving 40% memory","GhostRNN: less redundancy, same speech performance, 40% smaller"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest premise is that the redundant dimensions of a trained RNN hidden state can always be regenerated from a smaller intrinsic set by a simple learned linear-plus-activation map without losing task-relevant information, a claim the paper supports with PCA and cosine-similarity evidence from a single GRU model rather than across architectures.","fun_headline_variants_meta":{"raw":{"variants":["GhostRNN cuts RNN memory 40% by reusing hidden states","RNN redundancy exploited: GhostRNN shrinks state, keeps accuracy","Cheap ops turn few RNN states into many, saving 40% memory","GhostRNN: less redundancy, same speech performance, 40% smaller"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00035,"raw_usage":{"total_tokens":1897,"prompt_tokens":916,"completion_tokens":981,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":898}},"tokens_in":532,"tokens_out":981,"duration_ms":7132,"temperature":1.0,"reasoning_tokens":898,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:41:10.546000+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same keyword spotting and speech enhancement models on tasks whose hidden states are known to be irreducibly high-dimensional, or measure the PCA contribution rate across a variety of RNN architectures: if a well-trained RNN shows near-zero cosine similarity among most dimensions and needs close to 100% of singular values for 99% energy, then generating ghost states from a half-sized intrinsic set should noticeably degrade accuracy or SDR, and the roughly 40% compression would fail to match the baseline.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the keyword spotting dataset setup, training recipe, and the GRU baselines used in Table 1."},{"cited_title":"• GRU-TasNet","cited_arxiv_id":null,"evidence_quote":"Provides the DCCRN speech enhancement baseline whose GRU hidden states are analyzed for redundancy and then compressed."},{"cited_title":"Learning long-term de- pendencies with gradient descent is difficult,","cited_arxiv_id":null,"evidence_quote":"Provides the GRU-TasNet architecture used as the second speech enhancement baseline."},{"cited_title":"Gate-variants of gated recurrent unit (gru) neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines the Google Speech Commands dataset used for the keyword spotting experiments."},{"cited_title":"An optimized recurrent unit for ultra- low-power keyword spotting,","cited_arxiv_id":null,"evidence_quote":"Defines the LibriMix dataset used to build the noisy speech enhancement evaluation set."}],"review_version":1}