{"id":"b2dfcb9d-d4e4-4d33-a474-786debbb6e58","arxiv_id":"2608.05464","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A local, unsupervised pruning rule based on activity fluctuations preserves task performance in trained recurrent networks and outperforms magnitude and second-order pruning baselines, with optimal rescaling weaker than theory predicts.","lead":"This paper tests a biologically inspired pruning rule, noise-prune, on recurrent neural networks trained to perform cognitive tasks. It finds that pruning based on noisy activity fluctuations preserves task performance far better than pruning by connection strength alone, and that moderating the strength boost given to surviving connections works best.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: a magnitude-based sampling-and-rescaling baseline is never tested, so the covariance term in Eq. (7) is not isolated as the cause of noise-prune's advantage.","rationale":"The reader's weakest assumption is that the covariance of neural activity under injected noise is a valid proxy for connection importance in nonlinear task-trained networks. I agree, and the missing magnitude+rescale control is the sharpest, most testable form of that concern. Fig. 2 convincingly shows that stochastic sample-and-rescale beats deterministic thresholding for both L-NP and S-NP, and that deterministic S-NP scores beat deterministic magnitude. But those comparisons do not test whether the covariance term contributes once sampling and rescaling are used. If magnitude+rescale is as good, the title's 'noisy fluctuations' would be misleading. This is a feasible experiment, not a demand for theory. The cap-quantile result (Fig. 5) is a secondary concern because the optimal q is selected on the same evaluation data and the gains are modest; a held-out cap selection would be needed before relying on the specific q=50/60 recommendation. However, even if cap overfitting were confirmed, the main empirical finding would survive; the missing magnitude+rescale control threatens the central mechanistic claim. The overall verdict should remain conditional: the algorithm works as a package, but the paper should add this control (and ideally release code and artifacts) before claiming that covariance-based importance is what drives pruning.","tokens_in":18393,"tokens_out":9986,"duration_ms":93911,"concrete_test":"Add a magnitude-probability control: for each of the 24 task-trained networks, define p_ij = K|w_ij|, choose K to match each target density in expectation, sample every edge independently, rescale retained edges by 1/p_ij, apply the same final top-k density correction as S-NP, and evaluate task retention at 50/60/70/80% sparsity with the same three pruning seeds. Plot this curve against L-NP and S-NP in Figs. 1–2 and run the same paired Wilcoxon/sign tests. If magnitude+rescale matches noise-prune, the covariance term is not the driver; if it is significantly worse, the mechanism is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that noise-prune—using the covariance term in Eq. (1)/(7)—preserves task performance far better than magnitude-based pruning, and that stochastic sampling plus rescaling is essential. But the experiments vary two factors at once. Noise-prune differs from the magnitude baseline in both the importance score (|w| times the covariance expression vs. |w| alone) and in the pruning operation (probabilistic sample-and-rescale vs. deterministic top-k). The paper includes a deterministic noise-prune variant (Fig. 2) to isolate the score without rescaling, but it never includes the complementary control: a probabilistic magnitude-based rule with 1/p rescaling. If such a control performs as well as L-NP/S-NP, the advantage attributed to 'noisy fluctuations' would instead come from the sample-and-rescale recipe, and the distinctive covariance term would be unnecessary. This matters because the covariance term is the biologically motivated, theoretically grounded component of the rule; its validity in trained tanh networks is an empirical assumption, not a consequence of the linear spectral-sparsification theory. The current 8-task, 24-network data set supports the algorithmic package but does not identify the active ingredient.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript evaluates noise-prune, an unsupervised local pruning rule for recurrent networks, on task-trained continuous-time RNNs (8 Mod-Cog tasks, 24 networks, H=512). It compares two noise-prune variants (simulation-based S-NP and linearized L-NP) against magnitude pruning, random pruning, a recurrent adaptation of optimal brain surgeon, and deterministic versions of noise-prune, at 50--80% sparsity. The main findings are that noise-prune preserves task performance substantially better than magnitude pruning, is comparable to or better than recurrent-OBS at high sparsity, that the probabilistic sample-and-rescale operation is essential, and that capping the rescale factor at an intermediate quantile improves performance relative to the uncapped rule.","tokens_in":18649,"tokens_out":4671,"duration_ms":43995,"significance":"If the claims hold, the paper makes a useful contribution: it provides the first evaluation of noise-prune on functional, task-trained recurrent networks, with a broad suite of tasks, a strong comparison baseline in recurrent-OBS, and careful statistical practices including Holm-corrected paired tests, sign-test robustness checks, fixed evaluation batches, and reproducibility details. The work also identifies a practically important non-monotonicity in rescaling that refines the original theoretical prescription. The main caveats are that the comparison to magnitude pruning does not isolate the covariance-based importance score from the sampling-and-rescaling recipe, and that the optimal cap quantile is selected on the same data used to report significance, so the reported p-values are optimistic for the selection procedure.","major_comments":[{"comment":"The experimental design never includes a magnitude-based importance score combined with the same sample-and-rescale operation that noise-prune uses. As a result, the comparison between noise-prune and magnitude pruning varies two factors at once: the importance score (|w| times the covariance expression versus |w| alone) and the pruning operation (probabilistic sample-and-rescale versus deterministic top-k). The paper explicitly calls the covariance term 'the key contribution of noise-prune' in the Results, but the current data do not rule out the possibility that a probabilistic magnitude-based rule with 1/p rescaling performs as well as S-NP or L-NP. A control that applies the same sampling and rescaling to the magnitude score is needed to support the claim that noisy fluctuations, rather than the sample-and-rescale recipe alone, drive the observed advantage.","section":"Results, 'Optimal rescaling is lower than predicted by theory' and Methods, 'Variance-capped rescaling'"},{"comment":"The cap quantile q is selected by sweeping q on the same evaluation data used to report the peak-versus-uncapped significance (L-NP q=60, P=6.53e-5; S-NP q=50, P=8.34e-6). The reported p-values therefore do not account for the fact that the best q was chosen after looking at the same data. Holm correction across the 18 pairwise comparisons is a step in the right direction, but it does not control for the maximum-selection implicit in choosing q. The claim that optimal rescaling is lower than the theoretical prediction should be validated on held-out data or with an analysis that properly adjusts for selection (for example, a split-half or cross-validated choice of q).","section":"Results, 'Noise-prune preserves function...'"},{"comment":"The statement that 'both variants of noise-prune significantly outperform magnitude pruning (p<10^-4 for all noise-prune to magnitude comparisons)' is not supported by the supplementary statistics. At 50% sparsity, the Holm-corrected p-value for L-NP rescale versus magnitude is 1.79e-4, which is not below 1e-4, and the deterministic L-NP variant is not significantly different from magnitude at 50% (p=0.0885) or 60% (p=1.000). The claim should be corrected or qualified to refer specifically to the rescale variants at sparsity levels of 60% and above.","section":"Results, 'Noise-prune preserves function...'"}],"minor_comments":[{"comment":"The procedure for choosing K is described only as 'chosen to match the desired connection density in expectation'; please specify the algorithm, for example a binary search over K, and state any bounds or tolerances used.","section":"Methods, noise-prune"},{"comment":"The text should state unambiguously whether the covariance matrix C is computed from firing rates r(t) or from voltages v(t), since Eq. (9) adds noise to rates while the calibration in Eq. (10) uses voltage variability.","section":"Methods, simulation noise-prune"},{"comment":"Several rows in Supplementary Table S2 show an asterisk in only one of the two significance columns, but the meaning of a single asterisk is not explained; please clarify which test each asterisk refers to.","section":"Supplementary Table S2"},{"comment":"The evaluation of capped rescaling averages task retention across the 50--80% sparsity levels; it would be informative to also show the cap effect separately at each sparsity level, since the optimal cap may depend on sparsity.","section":"Methods, variance-capped rescaling"},{"comment":"The manuscript would benefit from a brief statement about the computational cost of S-NP versus L-NP, since the near-equivalence of the two variants is a useful practical finding but the runtime implications are not quantified.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The two main issues are the missing magnitude-plus-rescaling control and the data-dependent selection of the cap quantile. Both are fixable within the scope of the manuscript: adding one control experiment and either validating the cap choice on held-out data or explicitly framing the cap analysis as exploratory would address them. The statistical overstatement in the first results paragraph should also be corrected. The manuscript is otherwise solid and the empirical contribution is valuable for the journal's audience."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth your time, but with one gap. The paper gives the first test of noise-prune on task-trained nonlinear RNNs (8 Mod-Cog tasks, 24 networks), and the main comparative claims are credible: the full noise-prune package beats magnitude pruning and is competitive with recurrent-OBS, especially at high sparsity. The finding that stochastic sampling plus rescaling is essential, and that capping the rescale factor improves retention, is useful and new.\n\nThe experimental design is better than usual for this area: paired Wilcoxon tests with Holm correction, sign-test robustness checks, and a strong OBS baseline. I believe the headline numbers.\n\nThe soft spots are both real. First, the paper never runs a magnitude-based sample-and-rescale control. Noise-prune differs from magnitude in both the importance score and the pruning operation; the deterministic noise-prune variants isolate the score, but no one checks whether a stochastic, rescaled magnitude rule does just as well. If it does, the biological story about covariance probing is overstated. The deterministic comparison suggests the score contributes, but it is not a complete dissociation. A referee should ask for this control. Second, the capping result is modest (S-NP 0.730 vs 0.676 uncapped) and the optimal q is picked on the same evaluation data used for the significance test, so those p-values are inflated. The effect is plausible, but report it as exploratory or use held-out data. Reproducibility is also hampered by no code or data artifact.\n\nAudience: people working on biologically plausible pruning or sparse RNN deployment. I'd bring it to a reading group, and I'd cite it for the empirical result, but I'd want the missing control before building on the mechanistic claim. My recommendation: send to peer review; the core evidence is solid and the gaps are fixable.","headline":"A solid empirical validation of noise-prune on task-trained RNNs, but the missing magnitude-based sample-and-rescale control leaves the active ingredient unproven.","tokens_in":19177,"tokens_out":4510,"would_cite":true,"duration_ms":41272,"reading_group":"yes","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 shows that noise-prune, a local pruning rule using noise-driven activity covariance, preserves task performance in task-trained recurrent neural networks, beating magnitude-based pruning and matching or exceeding a non-local…","keywords":["noise-prune","synaptic pruning","recurrent neural networks","biologically-plausible learning","covariance-based importance","task-trained RNNs","cognitive tasks","stochastic sampling"],"falsifier":"Run the same pruning comparison on task-trained recurrent networks with a different activation function (e.g., ReLU or gated units) or on a task with long temporal dependencies; if noise-prune's retention advantage over magnitude pruning at 70–80% sparsity disappears, the covariance-based importance criterion does not generalize beyond the tanh networks tested here.","tokens_in":18195,"feed_emoji":"🧠","tokens_out":12530,"duration_ms":91553,"temperature":0.7,"pith_summary":"The paper tests noise-prune, a pruning rule for recurrent networks that decides which synapses to keep using the covariance of neural activity under injected noise rather than connection strength alone. On recurrent networks trained to perform eight cognitive tasks, the rule preserves task accuracy far better than magnitude-based pruning, and matches or exceeds a non-local second-order method (recurrent OBS) at high sparsities. The authors show the rule's stochastic sampling step—retaining connections with probability proportional to their importance and strengthening survivors by the inverse probability—is what makes it work; deterministic versions of the same importance scores perform much worse. They also find the optimal amount of strengthening is lower than the theoretical 1/p rescaling predicts: capping amplification factors at modest levels improves both task retention and dynamical stability. The result matters because it offers a biologically-plausible account of how synaptic pruning could work in the brain, using only local information.","feed_headline":"Noise-based pruning beats weight-size pruning in trained RNNs","feed_subtitle":"A local, unsupervised rule keeps more task accuracy through 80% pruning, rivaling costlier second-order methods.","key_machinery":"The central machinery is a covariance-based importance score combined with stochastic sample-and-rescale pruning. The retention probability for a connection from neuron j to neuron i is p_ij = K w_ij (C_ii + C_jj − 2C_ij) for excitatory weights and K|w_ij|(C_ii + C_jj + 2C_ij) for inhibitory weights, where C is the covariance matrix of activity under injected noise; the covariance term C_ii + C_jj − 2C_ij penalizes connections between neurons whose activity is already highly correlated, marking the direct connection as redundant. The rule independently retains each edge with probability p_ij and, if retained, multiplies the weight by 1/p_ij, preserving the expected strength of the edge while replacing a redundant set of connections by a few strengthened survivors. The paper obtains C in two ways: empirically from noisy simulation (S-NP) and analytically from the Lyapunov equation applied to the linearized network (L-NP).","core_discovery":"The central claim is that noise-prune preserves function in task-trained recurrent neural networks. For each connection, the rule computes a retention probability from the connection weight and the covariance of the two neurons' noisy activity—p_ij = K w_ij (C_ii + C_jj − 2C_ij) for excitatory connections, with the sign of the covariance term flipped for inhibitory connections—then samples whether to keep the connection with that probability and, if kept, strengthens it by 1/p_ij. Across 24 networks trained on eight cognitive tasks requiring working memory, context-dependent selection, interval estimation, and sequence production, both the simulation-based and the linearized Lyapunov-based variants of noise-prune outperform magnitude pruning at all tested sparsities (50–80%), match a one-shot recurrent version of the Optimal Brain Surgeon at 50–60%, and exceed it at 70–80%. The paper further establishes that the sampling-and-rescaling mechanism, not just the covariance importance scores, is essential: deterministic thresholding of the same scores performs much worse. Finally, the authors show that the theoretically-derived rescaling is too aggressive; capping the amplification at the 50th–60th percentile of the candidate distribution improves task retention and keeps the spectral abscissa lower, indicating a stability benefit.","pith_inferences":["If the covariance importance score generalizes beyond tanh networks, noise-prune could be paired with retraining or continual learning to prune recurrent networks iteratively, pushing sparsity well below the one-shot limits reported here.","The stability-accuracy tradeoff from capping suggests a testable biophysical prediction: homeostatic synaptic scaling should be bounded, and post-pruning synapse strength distributions should show a finite cutoff rather than the unbounded heavy tail predicted by exact expectation-preserving rescaling.","Because the linearized variant works almost as well, even coarse covariance estimates—for example from subsampled or noisy recordings—might drive effective pruning, which could be tested in hybrid models that estimate C from partial observations.","The same covariance-based score could be reinterpreted as a Hebbian/anti-Hebbian rewiring rule operating during learning, not just a one-shot pruning event, connecting noise-prune to theories of developmental synapse elimination."],"forward_implications":["A biologically-plausible local pruning rule—using only the weight of a connection and the activity of the two neurons it connects—can preserve function in recurrent networks that perform cognitive tasks.","The linearized variant performs nearly as well as the full simulation variant, so a closed-form covariance estimate suffices when the network can be linearized.","The stochastic sampling-and-rescaling step is the load-bearing component; deterministic thresholding of the same importance scores loses most of the benefit.","Capping rescaling at modest factors (roughly 5–10x) improves both task retention and dynamical stability, so the exact expectation-preserving rescaling suggested by theory is not optimal in nonlinear networks.","At high sparsity (70–80%), noise-prune exceeds one-shot recurrent OBS, indicating that a local rule can beat a non-local second-order method when both are applied in a single shot."],"supporting_citations":[{"why":"Introduces the noise-prune algorithm, its covariance-based retention probabilities, and the theoretical spectral-sparsification guarantees that the present study extends to task-trained networks.","marker":"[28]"},{"why":"Provides the battery of eight cognitive tasks used to train the recurrent networks.","marker":"[30]"},{"why":"Provides the multi-task training framework from which the cognitive task battery is derived.","marker":"[29]"},{"why":"The Optimal Brain Surgeon method that the paper adapts into the recurrent-OBS comparison baseline.","marker":"[14]"},{"why":"The Optimal Brain Damage precursor whose Hessian-based sensitivity approach motivates the second-order comparison.","marker":"[12]"},{"why":"The layer-wise OBS variant whose Hessian approximation is adapted for the recurrent layer in recurrent-OBS.","marker":"[16]"},{"why":"Supplies the effective-resistance graph sparsification theory that underlies noise-prune's original guarantees.","marker":"[35]"}],"fun_headline_variants":["Noise-prune preserves RNN tasks over weight-size pruning","Local unsupervised rule rivals costly second-order methods in RNNs","Sampling and rescaling key to noise-based RNN pruning","Noise fluctuations outdo connection strength in RNN pruning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The rule assumes that the covariance of two neurons' activity under noise, or its linear approximation, correctly identifies which direct connections are redundant in a nonlinear task-trained network, and this is validated only empirically on eight tasks.","fun_headline_variants_meta":{"raw":{"variants":["Noise-prune preserves RNN tasks over weight-size pruning","Local unsupervised rule rivals costly second-order methods in RNNs","Sampling and rescaling key to noise-based RNN pruning","Noise fluctuations outdo connection strength in RNN pruning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001914,"raw_usage":{"total_tokens":7527,"prompt_tokens":1005,"completion_tokens":6522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":6452}},"tokens_in":621,"tokens_out":6522,"duration_ms":46047,"temperature":1.0,"reasoning_tokens":6452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T12:34:12.133152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pruning comparison on task-trained recurrent networks with a different activation function (e.g., ReLU or gated units) or on a task with long temporal dependencies; if noise-prune's retention advantage over magnitude pruning at 70–80% sparsity disappears, the covariance-based importance criterion does not generalize beyond the tanh networks tested here.","supporting_citations":[{"cited_title":"& Chaudhuri, R","cited_arxiv_id":null,"evidence_quote":"Introduces the noise-prune algorithm, its covariance-based retention probabilities, and the theoretical spectral-sparsification guarantees that the present study extends to task-trained networks."},{"cited_title":"R., Joglekar, M","cited_arxiv_id":null,"evidence_quote":"Provides the multi-task training framework from which the cognitive task battery is derived."},{"cited_title":"& Stork, D","cited_arxiv_id":null,"evidence_quote":"The Optimal Brain Surgeon method that the paper adapts into the recurrent-OBS comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Optimal Brain Damage precursor whose Hessian-based sensitivity approach motivates the second-order comparison."},{"cited_title":"& Pan, S","cited_arxiv_id":null,"evidence_quote":"The layer-wise OBS variant whose Hessian approximation is adapted for the recurrent layer in recurrent-OBS."}],"review_version":1}