{"id":"38b6fac8-ab31-49d0-82c4-01a922679087","arxiv_id":"2501.06962","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Post-hoc signal-to-noise and signal-plus-noise pruning with a short resampling run can shrink MCMC-trained Bayesian neural networks by 75% with modest accuracy loss, though uncertainty retention is not measured.","lead":"This paper prunes Bayesian neural networks that are trained by Markov Chain Monte Carlo sampling, removing up to 75% of their weights while keeping test accuracy roughly intact, and then briefly re-trains the survivors. The authors test this on small benchmark datasets and on coral reef drill-core classification, where the compact model still trails a random forest.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-pruning resampling is the load-bearing novelty, but 1000 no-burn-in Langevin iterations are not shown to yield a valid posterior for the compact BNN, and no uncertainty metric is measured.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the post-pruning resampling must produce a faithful posterior for the surviving weights, but the paper offers no direct evidence of posterior validity. The abstract's uncertainty-retention claim is the strongest part of the central claim, and Section 4.2 never measures uncertainty quality. The paper's own R-hat results at high pruning rates contradict the convergence narrative, and the rebuttal's trace plots select one parameter per dataset. This is not a minor reporting gap; it is the methodological core of the claimed novelty. The missing 0%-pruning baseline in Tables 2 and 3 is also a real weakness, because 'retaining generalisation performance' needs a comparison to the unpruned model, but that is a fixable reporting issue rather than the deepest threat to the argument. The direction of the empirical results is plausible and honestly presented, especially the reef-core comparison where the compact BNN trails Random Forest while offering predictive distributions. Given the paper's current form, the appropriate verdict is the reader's CONDITIONAL: the central claim is not yet established, but the required evidence is concrete and obtainable. I therefore recommend keeping the verdict unchanged rather than escalating to rejection.","tokens_in":28164,"tokens_out":4034,"duration_ms":44321,"concrete_test":"On one dataset (e.g., Exp 325) at 75% pruning, rerun Stage 4 with 10,000 Langevin iterations and discard the first 2,000 as burn-in, keeping the same proposal and step-size settings. Compute per-parameter R-hat, expected calibration error, and test log-loss for the full predictive distribution, and compare against the paper's 1000-iteration version and the unpruned BNN. If longer resampling materially changes accuracy, log-loss, or calibration, or yields R-hat > 1.1 for a non-negligible fraction of parameters, the compact posterior is not converged as claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on Stage 4 of Algorithm 1: after zeroing pruned weights, the surviving weights are resampled for 'an additional 1000 samples without burn-in' (Section 4.1) and this resampled posterior is what the abstract says lets the compact BNN 'retain its ability to estimate uncertainty.' Section 3.1 explicitly identifies this resampling as the novel contribution relative to [54, 99]. But the paper's own convergence evidence is contradictory: Section 4.5 reports 'post-pruning resampling shows slightly poor convergence with higher (R-hat) values at high pruning rates,' and the rebuttal relies on trace plots of one manually selected parameter per dataset (Figure 6). Worse, Section 4.2 evaluates only RMSE, accuracy, and AUC; no calibration, coverage, or predictive-variance metric is reported, so the claim that uncertainty is retained is not directly tested. If the 1000-iteration chain has not mixed, the pruning scores (computed from posterior moments of the pre-pruning chain) and the uncertainty statements after resampling are both built on unreliable samples. The empirical pattern is plausible, but the mechanism claimed as novel is precisely the one whose validity is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pruning framework for Bayesian neural networks trained with Langevin MCMC. After a full MCMC run, weights and biases are ranked by signal-to-noise (STN) or signal-plus-noise (SPN) criteria, and those below a threshold are zeroed; the surviving parameters are then resampled with a short Langevin MCMC run. The approach is evaluated on three regression and five classification datasets, including two coral reef drill-core lithology datasets, with comparisons to random pruning. The paper claims that the resulting compact BNNs retain generalization performance and uncertainty estimation, and reports experiments at 25%, 50%, and 75% pruning levels.","tokens_in":28244,"tokens_out":7922,"duration_ms":67824,"significance":"If the reported results hold, the paper would provide a practical way to obtain compact MCMC-trained BNNs with uncertainty estimates, potentially useful for real-world applications such as drill-core classification. The empirical comparison of structured versus random pruning across multiple datasets and 30 independent runs is a genuine strength, as is the release of code. However, the paper's most distinctive claim—that post-pruning resampling preserves the posterior's uncertainty estimates—is not directly tested, and the convergence evidence for the resampled chains is weak. The structured-pruning-over-random-pruning pattern is visible in Tables 2 and 3, but several load-bearing presentation choices and missing baselines prevent the central claims from being fully verified.","major_comments":[{"comment":"The abstract claims that the compact BNN retains its ability to estimate uncertainty via the posterior distribution, but Section 4.2 reports only RMSE, classification accuracy, and AUC. No calibration, coverage, predictive log-likelihood, or predictive-variance metric is measured, so the uncertainty-preservation claim is not directly tested; this is a central claim of the paper and should be verified or removed.","section":"Section 4.2"},{"comment":"The validity of the post-pruning resampling step is not established. The text reports that the resampled chains show higher R-hat values at high pruning rates, and the rebuttal relies on trace plots of one parameter per dataset in Figure 6. The reported sampling budget is also inconsistent: Section 4.1 says 50,000 samples with an additional 1000 samples without burn-in, while the Figure 6 caption refers to 25,000 post-burn-in samples and 900 post-burnin resampling samples. Please provide proper convergence diagnostics for the resampled posterior (e.g., R-hat for all parameters, effective sample size, or a longer resampling run) and align the reported numbers.","section":"Section 4.5 and Algorithm 1 Stage 4"},{"comment":"No unpruned baseline is reported in the tables. The columns labeled 'Resampling No' refer to the pruned network without resampling, not to the full BNN, so the reader cannot verify the claim that performance is retained relative to the full network at any pruning level. The paper should report the full-network (0% pruning) RMSE and accuracy alongside the pruned values, so the 'retaining generalisation performance' claim can be assessed.","section":"Tables 2 and 3"},{"comment":"The statement that 'STN consistently outperforms both RND and SPN across all classification datasets and pruning levels' is not supported by the table. For example, at 25% pruning with resampling, Ionosphere SPN achieves 92.73 versus STN 92.55, and Expedition 310 SPN achieves 37.11 versus STN 36.78; at 50% pruning with resampling on Abalone, SPN achieves 78.37 versus STN 78.27. Please qualify this claim or provide a paired statistical comparison.","section":"Section 4.4 and Table 3"},{"comment":"The Metropolis-Hastings acceptance probability is incorrectly specified. The expression α = min(1, P(θ')q(θ_i|θ) / P(θ_i)q(θ''|θ_i)) uses an undefined θ'' and does not state the proposal density q for the Langevin proposal in Equation (7). Without the correct proposal-ratio term, the sampler as written is not a valid MH algorithm and cannot be reproduced. Please define q and give the correct acceptance ratio (or state that the implementation uses an alternative valid scheme).","section":"Algorithm 1, Step 3"},{"comment":"The abstract claims 'over 75% reduction in network size,' but the experiments evaluate pruning levels of 25%, 50%, and 75% only. The largest tested reduction is exactly 75%; no result supports 'over 75%.' Either test higher pruning rates or revise the abstract to 'up to 75%.'","section":"Abstract and Sections 4.3-4.4"}],"minor_comments":[{"comment":"The affiliation line contains a typo: 'Asutralia' should be 'Australia.'","section":"Affiliations"},{"comment":"'German-Rubin' should be 'Gelman-Rubin' in the text, the figure caption, and the conclusion.","section":"Section 4.5 and Figure 7"},{"comment":"'datsets' is a typo for 'datasets.'","section":"Section 3.1.3"},{"comment":"The condition 'Pruning Ratio<λ' is unclear; the manuscript should specify that the pruning criterion is the STN or SPN value for each weight, not a generic ratio.","section":"Algorithm 1, Stage 3"}],"recommendation":"major_revision","confidential_remarks":"The empirical core appears genuine and the code is released, but the central uncertainty-preservation claim is under-tested and the convergence evidence is weak. The issues are addressable with additional experiments and clarifications, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful engineering recipe, not a methodological breakthrough, and the paper becomes acceptable after a revision round that fixes three overclaims. The genuinely new bit is applying STN/SPN pruning to MCMC-trained BNNs plus a short post-pruning resampling run. The empirical core is visible in Tables 2 and 3: structured pruning beats random pruning at high rates, and resampling recovers most of the accuracy loss. Thirty runs, mean/std reporting, and a public GitHub repository make the results reasonably reproducible. The reef-core application is reported honestly—the compact BNN trails the Random Forest baseline on both expeditions—and that is credit-worthy.\n\nNow the soft spots, in proportion. The biggest is the uncertainty claim. The abstract says the compact BNN retains its ability to estimate uncertainty, and Section 3.1 sells the resampling stage as the novel contribution, but Section 4.2 reports only RMSE, accuracy, and AUC. No calibration, coverage, or predictive-variance metric is measured, so the central claim is untested. The stress-test note hits the right target: 1000 Langevin iterations without burn-in is a thin basis for a faithful posterior for the surviving weights, and the paper's own R-hat diagnostics show convergence worsening at high pruning rates. The rebuttal—trace plots of one selected parameter per dataset—does not settle that.\n\nSecond, the claim that STN consistently outperforms SPN across all classification datasets and pruning levels is contradicted by the authors' own Table 3: at 75% pruning on Ionosphere with resampling, SPN scores 78.58 versus STN's 77.38, and most method differences sit within one standard deviation. The honest summary is that structured pruning beats random pruning, with STN and SPN roughly comparable and dataset-dependent ordering.\n\nThird, smaller issues: Tables 2 and 3 lack 0% pruning columns, the only baseline is random pruning rather than a simple magnitude-pruning baseline, and numeric hyperparameters such as Langevin step size and prior variances are not stated. All are fixable.\n\nThe convergence narrative also contradicts itself: Section 4.5 reports slightly poor convergence after aggressive pruning, Section 6 reports better convergence. That needs to be reconciled.\n\nVerdict: worth sending to peer review. A serious referee can push on the uncertainty claim and the resampling-convergence evidence, and the recipe itself is honest and likely useful for small MCMC-trained BNNs. I would not bring it to a general reading group unless someone is actively working on BNN pruning or MCMC-based uncertainty; otherwise it is a solid, unglamorous empirical paper.","headline":"Useful MCMC-BNN pruning recipe with honest reef application, but the uncertainty claim is untested and the convergence evidence is thin.","tokens_in":28993,"tokens_out":3604,"would_cite":true,"duration_ms":29789,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Pruning a Bayesian neural network by each weight's signal-to-noise ratio after MCMC training, then briefly resampling the surviving weights, cuts the network to under a quarter of its size while retaining accuracy and uncertainty estimates.","keywords":["Bayesian neural networks","MCMC","network pruning","signal-to-noise ratio","Langevin dynamics","uncertainty quantification","lithology classification","model compression"],"falsifier":"Run at least three resampling chains of 10,000 iterations or more on the surviving weights at 75% pruning and compute per-weight Gelman-Rubin values; if many weights exceed an R-hat of 1.1, the short resampling phase has not converged and the uncertainty estimates are not trustworthy. A complementary check is to compare predictive-interval calibration of the compact model on held-out data against the full model.","tokens_in":27765,"feed_emoji":"🧠","tokens_out":5786,"duration_ms":50126,"temperature":0.7,"pith_summary":"Bayesian neural networks (BNNs) quantify prediction uncertainty by placing a distribution over every weight, but they are slow to train and costly to deploy. This paper claims that most of those weights are redundant and can be removed after MCMC training, provided the remaining weights are given a short resampling run; with that step, networks pruned to 75% smaller retain their accuracy and their uncertainty estimates on the benchmarks and reef drill-core datasets tested. The pruning is guided by two posterior statistics of each weight—signal-to-noise and signal-plus-noise ratios—and is compared against random pruning, which degrades performance sharply at high pruning rates. The authors argue this makes uncertainty-aware models practical in resource-constrained applications such as drill-core lithology classification, underwater robotics, and remote sensing.","feed_headline":"MCMC-trained nets keep accuracy after 75% pruning","feed_subtitle":"A short resampling run lets compact Bayesian networks keep their uncertainty estimates.","key_machinery":"The load-bearing machinery is the post-pruning resampling stage (Stage 4 of Algorithm 1), applied to weights selected by two pruning ratios computed from MCMC posterior samples: the signal-to-noise ratio $|\\mu_i|/\\sigma_i$ and the signal-plus-noise ratio $|\\mu_i| + \\sigma_i$, where $\\mu_i$ and $\\sigma_i$ are the posterior mean and standard deviation of weight $i$. Weights whose score falls below the user-defined threshold $\\lambda$ are set to zero, and the surviving weights are then resampled by Langevin MCMC—the step the paper calls novel for these criteria—so the compact model's posterior can absorb information lost from the pruned weights. Convergence of the resampled chains is checked with the Gelman-Rubin potential scale reduction factor.","core_discovery":"Stated on the paper's own terms, the discovery is that a Bayesian neural network trained with Langevin MCMC can be made compact without sacrificing predictive or probabilistic performance: after sampling the posterior for 50,000 iterations, the authors sort weights by either $|\\mu_i|/\\sigma_i$ (STN) or $|\\mu_i|+\\sigma_i$ (SPN), zero out those below a threshold $\\lambda$, and then resample the surviving weights for 1000 Langevin iterations without burn-in. Across three regression and five classification datasets, including two real-world coral reef lithology sets, the pruned-and-resampled models retained accuracy at 75% pruning, SPN proved best for regression and STN for classification, and resampling improved results for every method. The paper's headline quantitative claim is over 75% network-size reduction with retained generalisation performance.","pith_inferences":["A longer resampling phase with a proper burn-in might reduce the elevated R-hat values the paper reports; the 1000-iteration choice looks like a computational economy rather than a convergence guarantee.","Because the pruning scores are posterior statistics of the full chain, the method inherits any convergence bias of the initial MCMC run; using tempered or parallel-tempered MCMC could shift which weights are pruned.","The same STN/SPN scores could be computed from a variational posterior, so a natural testable extension is whether the prune-then-resample recipe transfers to variational Bayesian networks.","The comparison with Random Forest suggests the practical case for these compact BNNs rests on uncertainty information rather than raw accuracy; a calibration study on the reef-core classes would test that case directly."],"forward_implications":["At 75% pruning, structured STN/SPN pruning with resampling preserves benchmark accuracy where random pruning loses it.","Resampling is an essential post-pruning step: it improved accuracy across all datasets and recovered up to 25% of performance at high pruning levels.","The criterion choice matters: SPN gives the most precise regression models, while STN gives the most accurate classifiers.","Compact MCMC-trained BNNs become a viable option for resource-limited deployments that need uncertainty estimates, such as drill-core analysis and marine robotics.","The pruned-resampled BNNs approach but do not match Random Forest AUC on the reef-core datasets, while adding the uncertainty information Random Forest lacks."],"supporting_citations":[{"why":"Graves 2011, the signal-to-noise pruning heuristic the paper adapts to MCMC-trained BNNs.","marker":"[54]"},{"why":"Nalisnick 2018, source of the signal-plus-noise pruning criterion.","marker":"[99]"},{"why":"Chandra et al. 2019, the Langevin-gradient parallel tempering MCMC approach used to sample the posterior.","marker":"[3]"},{"why":"Chandra and Simmons 2024, the BNN-via-MCMC model, priors, and likelihood setup the experiments follow.","marker":"[2]"},{"why":"Chandra et al. 2017, Langevin-dynamics BNN training for chaotic time series and benchmark datasets reused here.","marker":"[64]"},{"why":"Gelman and Rubin 1992, the convergence diagnostic used to claim the sampled chains have converged.","marker":"[100]"},{"why":"Deo et al. 2024, the reef-core segmentation and classification work motivating the real-world datasets.","marker":"[111]"},{"why":"Expedition 325 proceedings, one of the two drill-core lithology datasets used for real-world evaluation.","marker":"[106]"},{"why":"Expedition 310 proceedings, the second reef drill-core dataset used for real-world evaluation.","marker":"[107]"}],"fun_headline_variants":["Pruned Bayesian nets keep uncertainty with 75% less size","MCMC pruning shrinks Bayesian nets 75% without losing accuracy","Compact Bayesian nets: prune weights, resample, keep certainty","Langevin MCMC + pruning gives 75% slimmer Bayesian models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that after pruning, 1000 Langevin MCMC iterations without burn-in are enough for the surviving weights to converge to a faithful posterior, so the compact model's predictions and uncertainties remain valid; the paper itself reports higher R-hat values at high pruning rates.","fun_headline_variants_meta":{"raw":{"variants":["Pruned Bayesian nets keep uncertainty with 75% less size","MCMC pruning shrinks Bayesian nets 75% without losing accuracy","Compact Bayesian nets: prune weights, resample, keep certainty","Langevin MCMC + pruning gives 75% slimmer Bayesian models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000821,"raw_usage":{"total_tokens":3614,"prompt_tokens":990,"completion_tokens":2624,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2549}},"tokens_in":606,"tokens_out":2624,"duration_ms":16125,"temperature":1.0,"reasoning_tokens":2549,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:51:27.607193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run at least three resampling chains of 10,000 iterations or more on the surviving weights at 75% pruning and compute per-weight Gelman-Rubin values; if many weights exceed an R-hat of 1.1, the short resampling phase has not converged and the uncertainty estimates are not trustworthy. A complementary check is to compare predictive-interval calibration of the compact model on held-out data against the full model.","supporting_citations":[{"cited_title":"Chandra, K","cited_arxiv_id":null,"evidence_quote":"Chandra et al. 2019, the Langevin-gradient parallel tempering MCMC approach used to sample the posterior."},{"cited_title":"Chandra, J","cited_arxiv_id":null,"evidence_quote":"Chandra and Simmons 2024, the BNN-via-MCMC model, priors, and likelihood setup the experiments follow."},{"cited_title":"Chandra, L","cited_arxiv_id":null,"evidence_quote":"Chandra et al. 2017, Langevin-dynamics BNN training for chaotic time series and benchmark datasets reused here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Deo et al. 2024, the reef-core segmentation and classification work motivating the real-world datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Expedition 325 proceedings, one of the two drill-core lithology datasets used for real-world evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Expedition 310 proceedings, the second reef drill-core dataset used for real-world evaluation."}],"review_version":1}