{"id":"7dd66d05-64f6-4324-91ca-b24612683a99","arxiv_id":"2504.15328","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Using the previous posterior as the next prior in federated SGLD training cut iterations to 85% accuracy by about 50% over three days of radar data, with improved calibration.","lead":"This paper tests a Bayesian federated learning method that reuses the previous day's posterior as the prior for the next day, applied to radar-based human sensing. It reports faster convergence to 85% accuracy and better calibration than retraining from scratch, but the evaluation omits key numbers and does not test forgetting.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'improves final performance and reliability' claim is unverified: Table I omits accuracy and ECE for P-CL and Retr, and the convergence counts are ambiguous (day-2 count below the stated burn-in).","rationale":"I read the paper as proposing a sequential Bayesian update (previous posterior as prior) for federated SGLD and claiming better speed and reliability than retraining. The method is standard and plausible; the assumption of non-abrupt shifts is explicitly stated in Section III and the conclusion, so I do not treat that scope condition as hidden. The weakest point is that the quantitative evidence in Table I does not report the quantities the claim is about. The reader's weakest_assumption was the smooth-shift condition; my concern is the missing direct comparison, which is more immediately load-bearing because even in the declared regime the claim is unverified. The proposed test is feasible with the existing simulation and would settle whether P-CL actually delivers equal or better accuracy and calibration at the claimed speed. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":7135,"tokens_out":6809,"duration_ms":66944,"concrete_test":"Re-run the same three-day protocol with at least five seeds and produce a complete results table per day and setup (TL, Retr, P-CL): mean and standard deviation of accuracy, ECE, and the full accuracy-vs-iteration curve, with the iteration counter explicitly defined as total or post-burn-in. Also compute iterations to 80%, 85%, and 90% accuracy from a common initialization. If P-CL's accuracy and ECE are not at least statistically indistinguishable from Retr's at matched final accuracy, or if the convergence advantage disappears under a warm-start control, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing problem is evidentiary: Table I does not contain the numbers that the central claim is about. It reports one accuracy row for Transfer Learning (95.5/79.9/72.7) and iterations-to-85% for Retr and P-CL, but no accuracy for Retr or P-CL and no ECE for any setup. Section V claims P-CL achieves 'similar performance' to Retr with lower ECE, while the conclusion says P-CL 'improves the final performance' and 'enhances reliability'; the stated evidence cannot verify either claim, and the two statements are not even mutually consistent. The convergence comparison is also ambiguous: with T=100 and Tb=50, P-CL's day-2 count of 14 iterations is below the burn-in length, so the reader cannot tell whether the counter includes burn-in or refers to post-burn-in samples. The 85% threshold is chosen post hoc from unshown accuracy curves, and starting the day from the previous posterior mean would alone explain earlier threshold crossings, so the speed advantage does not isolate the Bayesian-prior mechanism. None of this disproves the method, but it means the central claim is not currently supported by the reported evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Posterior-aided Continual Learning (P-CL), a Bayesian Federated Learning (BFL) method based on Stochastic Gradient Langevin Dynamics (SGLD) that reuses the posterior distribution from the previous day as the prior for the next day. The method is evaluated on a simulated human-sensing scenario with mmWave radar data collected over three days, comparing P-CL against transfer learning and per-day model retraining. The authors report accuracy for transfer learning, iterations needed to reach an 85% accuracy threshold, and expected calibration error (ECE), claiming that P-CL achieves similar accuracy to retraining with better calibration and reduced communication overhead.","tokens_in":7393,"tokens_out":5362,"duration_ms":44406,"significance":"The proposed approach addresses a real gap in BFL—continual learning under dynamic data distributions—and the idea of propagating the posterior as a prior is simple and communication-efficient. If the empirical claims were fully supported, the work would be of practical interest for federated sensing applications. However, the paper as submitted does not provide the numbers needed to verify these claims: accuracy and ECE for the two main methods are missing, the convergence-speed metric is ambiguous and apparently thresholded after the fact, and no forgetting metric is reported. The contribution is therefore not yet established; the method is plausible but the evidence is incomplete. The paper also does not ship code or machine-checked proofs; no such artifacts are claimed.","major_comments":[{"comment":"Table I reports accuracy only for Transfer Learning and iteration counts for Retr. and P-CL, but no accuracy or ECE values for Retr. and P-CL. The text in Section V claims P-CL 'achieves similar performance' to Retr. with lower ECE, and Section VI concludes P-CL 'improves the final performance' and 'enhances reliability'; without the underlying numbers, neither claim can be checked, and the two claims are not mutually consistent.","section":"Section V, Table I"},{"comment":"The convergence metric is ill-defined. With T=100 and Tb=50, the reported day-2 count of 14 iterations for P-CL is below the burn-in of 50, so the reader cannot determine whether the iteration count includes burn-in, refers to post-burn-in samples, or is measured differently. Moreover, the 85% threshold is said to be set 'based on the based on the accuracies from Fig. 3', but Fig. 3 displays only confidence histograms and reliability diagrams, not learning curves; the threshold therefore appears to be chosen post hoc, which weakens the claimed iteration reductions of 40%, 34%, and 50%.","section":"Section V, convergence speed"},{"comment":"The paper claims that P-CL preserves knowledge in a continual learning setting, but it never evaluates forgetting or performance on previous days' data after adaptation. The experiments report only current-day accuracy and ECE, so backward transfer and any notion of knowledge preservation are not measured; the continual-learning claim is therefore unsupported by the presented metrics.","section":"Section III and abstract"},{"comment":"The method relies on the assumption that the data distribution changes slowly between days (the paper repeatedly states 'no significant changes in the system happened between d and d-1'), but no experiment tests this assumption or varies the magnitude of distribution shift. Because a stale posterior prior could slow adaptation or lock the model onto outdated beliefs under abrupt shifts, the authors should either constrain their claims to the smooth-shift regime or include a sensitivity analysis; as written, the generality of the method is unverified.","section":"Section III and Section VI"}],"minor_comments":[{"comment":"The sentence 'The threshold was set based on the based on the accuracies from Fig. 3' contains a duplicated phrase and refers to Fig. 3, which does not show accuracy curves; please correct and clarify.","section":"Section V, final paragraph"},{"comment":"The word 'assimption' should be 'assumption'.","section":"Section VI, conclusion"},{"comment":"'Respected day's data' should be 'respective day's data'.","section":"Section IV"},{"comment":"The pointer 'as shown in Fig. 2' for the accuracy/ECE comparison is likely a typo for 'Fig. 3'.","section":"Section V"},{"comment":"Equation (4) introduces \\xi_{k+1,n} while Equation (2) uses \\xi_{k+1}; please align the notation.","section":"Equation (4) and Equation (2)"},{"comment":"The diagonal covariance approximation in Equation (6) is not discussed; a sentence justifying this choice would be helpful.","section":"Equation (6)"},{"comment":"The comparison set includes only transfer learning and retraining; adding a standard continual learning baseline (e.g., EWC or a variational continual learning method) would make the quantitative claims more convincing.","section":"Section V, comparison baselines"},{"comment":"The paper does not state whether the code or dataset will be released; a reproducibility statement would be welcome.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is extremely brief and the experiments are preliminary; the missing data in Table I would likely be caught by any careful reviewer. In my view, the novelty is modest—the method is essentially an application of standard Bayesian updating—so the acceptance decision should hinge on whether the empirical evidence can be strengthened substantially. The paper may be better suited to a workshop or a venue that accepts short empirical papers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's central claims are not supported by the numbers it actually reports. The idea—reusing the previous posterior as the prior for the next day's SGLD updates—is standard Bayesian online learning, but applying it to federated SGLD for radar-based human sensing is a legitimate extension, and the authors are upfront about the smooth-shift assumption. The experimental setup is concrete and the baselines are sensible.\n\nWhat the paper does well: it identifies a genuine gap (most BFL work ignores distribution shift over time), uses a real mmWave radar dataset, and at least attempts to measure calibration. The confidence histograms and reliability diagrams are the right diagnostic tool.\n\nThe soft spots are significant. Table I gives accuracy for transfer learning only, and iteration counts for Retr and P-CL, but no accuracy for those setups and no ECE anywhere. So 'improves final performance' and 'enhances reliability' are simply not checkable. The narrative is also inconsistent: the text says P-CL achieves 'similar performance' to Retr, while the conclusion claims it 'improves' performance. The convergence counts are ambiguous because with T=100 and burn-in 50, P-CL's day-2 count of 14 falls below the burn-in, so the reader cannot tell what is being counted. The 85% threshold was set after looking at the results, which softens the reported 40/34/50% gains. And a warm start from the previous posterior mean would cross any threshold faster than retraining from scratch, so the speed comparison does not isolate the Bayesian prior mechanism. There is also no forgetting metric, despite the knowledge-preservation claim.\n\nNone of this disproves the method, but it means the paper currently establishes a much narrower point: under smooth shifts, an SGLD federated model can reach a fixed accuracy in fewer iterations when initialized from the previous posterior. That is worth investigating, but the evaluation needs a full accuracy/ECE table, error bars, a pre-specified threshold or full convergence curves, and a control that separates warm-start from the Bayesian prior.\n\nI'd send this to review with major revisions rather than desk-reject, because the flaws are fixable and the question is relevant to industrial IoT and safety-critical HR collaboration. But I wouldn't cite it in its current form.","headline":"Plausible idea, but the reported evidence can't back the key claims about accuracy, calibration, and knowledge preservation.","tokens_in":7881,"tokens_out":3591,"would_cite":false,"duration_ms":30758,"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":"A Bayesian federated learning system for continually shifting radar data claims that reusing the previous day's posterior as the next prior reaches target accuracy faster, with better calibration, than retraining from scratch.","keywords":["Bayesian Federated Learning","Continual Learning","Stochastic Gradient Langevin Dynamics","Posterior-aided Continual Learning","Expected Calibration Error","radar human sensing","parameter server federated learning","mmWave FMCW radar"],"falsifier":"Train P-CL and retraining on the same radar benchmark but with a deliberately abrupt change on day 2—for example, swapping the ROI layout so the old posterior's high-probability regions point at the wrong locations. If P-CL's day-2 iterations-to-85% exceed retraining's, or its ECE rises above retraining's, the paper's claim that P-CL improves speed and reliability would fail for that regime.","tokens_in":6973,"feed_emoji":"📡","tokens_out":4394,"duration_ms":37913,"temperature":0.7,"pith_summary":"The paper argues that in a federated setting where sensor data drift slowly from day to day, a Bayesian model should not be retrained from scratch and should not be frozen. It proposes Posterior-aided Continual Learning (P-CL), in which the posterior distribution estimated on day d-1 becomes the Gaussian prior for day d, with samples drawn by Stochastic Gradient Langevin Dynamics. On a simulated human-sensing task with radar range-azimuth maps over three days, P-CL reaches the 85% accuracy threshold in 14 iterations on day 2 and 16 on day 3, compared with 36 and 47 for daily retraining, and reports lower expected calibration error. If correct, the approach cuts communication rounds by roughly half while keeping the model better calibrated than either transfer learning or retraining.","feed_headline":"Posterior-as-prior halves federated retraining cost","feed_subtitle":"A Bayesian federated model that reuses yesterday's posterior reaches 85% accuracy in 14 day-2 iterations instead of 36.","key_machinery":"The load-bearing mechanism is the prior-to-posterior chain. After day d-1, the federated SGLD sampler produces posterior samples; the paper computes their per-parameter mean and diagonal covariance and sets the day-d prior $p(\\theta^d)\\sim\\mathcal{N}(\\mu_{d-1},\\Sigma_{d-1})$. This replaces the standard zero-mean isotropic prior used in retraining. The work that this does is to carry the knowledge of the previous solution into the next optimization, so the sampler starts near the region of parameter space that already explains the sensor geometry, shortening burn-in and improving calibration.","core_discovery":"The central claim is that continual Bayesian updates—feeding the previous posterior back in as the prior—make continual federated learning both faster and more reliable. The paper shows this in a three-day mmWave radar human-sensing benchmark where the target Regions of Interest shift slightly between days. Using SGLD to sample the posterior, P-CL maintains accuracy comparable to full daily retraining but requires fewer iterations to reach 85% (14 vs 36 on day 2; 16 vs 47 on day 3), lowers ECE relative to retraining, and reduces communication overhead by about 50% over the simulation. The result is framed as evidence that Bayesian FL is a natural fit for continual learning in environments whose dynamics are not abrupt.","pith_inferences":["The P-CL principle is not tied to Gaussian priors; using the sample mean and diagonal covariance is a convenience. A natural testable extension is to replace it with a mixture, a low-rank covariance, or a nonparametric posterior summary and check whether the convergence and calibration gains persist.","The paper's own assumption that shifts are non-abrupt suggests an immediate stress test: simulate a day-to-day change that moves the Regions of Interest substantially. If P-CL then needs more iterations than retraining or locks onto the old geometry, the benefit is bounded by the smoothness of the drift.","One could also apply the same prior-recycling idea to variational Bayesian FL, where the prior is a Gaussian from the previous round; the mechanism is agnostic to the sampler and would compare directly with the SGLD version.","The communication saving is quantified in iterations, not bytes; a fuller accounting should include the cost of transmitting posterior covariance statistics or samples from server to nodes at each day boundary."],"forward_implications":["If the P-CL claim holds, federated systems that retrain daily can expect to reach a fixed accuracy target in roughly one-third to one-half the communication rounds on day 2 and day 3.","Using the previous posterior as prior makes calibration better (lower ECE) than retraining, so confidence estimates become more trustworthy for safety-oriented applications like human-robot coexistence.","Transfer learning is not enough in these settings: the fixed day-1 model degrades from 95.5% to 79.9% and 72.7% accuracy on days 2 and 3, showing the need for continual updates.","The same SGLD machinery with unchanged hyperparameters can serve repeated daily adaptation, so the approach is a drop-in modification to existing MCMC-based Bayesian FL loops."],"supporting_citations":[{"why":"Supplies the Stochastic Gradient Langevin Dynamics update rule and burn-in procedure that generate the posterior samples.","marker":"[12]"},{"why":"Establishes the Bayesian Federated Learning formulation and the posterior expression used as the starting point.","marker":"[7]"},{"why":"Defines the federated parameter-server synchronization and aggregation setup that P-CL builds on.","marker":"[1]"},{"why":"Provides the mmWave radar range-azimuth dataset and Region-of-Interest definitions used in all experiments.","marker":"[24]"},{"why":"Defines Expected Calibration Error, the reliability metric that distinguishes P-CL from retraining.","marker":"[26]"},{"why":"Supplies the MCMC-based decentralized Bayesian FL framework that the continual update extends.","marker":"[21]"},{"why":"Provides the LeNet architecture used as the classification model for the radar maps.","marker":"[25]"}],"fun_headline_variants":["Reusing yesterday's posterior speeds continual federated learning","Posterior replay: continual Bayesian FL with fewer iterations","Bayesian FL for shifting data: past posterior as prior","Faster continual federated learning via Bayesian posterior reuse","Continual Bayesian updates cut federated retraining iterations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole gain rests on the assumption that between consecutive days the data distribution only drifts mildly; if the workplace geometry changes abruptly, the previous posterior becomes a misleading prior and P-CL could be slower or overconfident.","fun_headline_variants_meta":{"raw":{"variants":["Reusing yesterday's posterior speeds continual federated learning","Posterior replay: continual Bayesian FL with fewer iterations","Bayesian FL for shifting data: past posterior as prior","Faster continual federated learning via Bayesian posterior reuse","Continual Bayesian updates cut federated retraining iterations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000952,"raw_usage":{"total_tokens":3996,"prompt_tokens":816,"completion_tokens":3180,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":3103}},"tokens_in":432,"tokens_out":3180,"duration_ms":20811,"temperature":1.0,"reasoning_tokens":3103,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:31:58.591896+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train P-CL and retraining on the same radar benchmark but with a deliberately abrupt change on day 2—for example, swapping the ROI layout so the old posterior's high-probability regions point at the wrong locations. If P-CL's day-2 iterations-to-85% exceed retraining's, or its ECE rises above retraining's, the paper's claim that P-CL improves speed and reliability would fail for that regime.","supporting_citations":[{"cited_title":"Bayesian learning via stochastic gradient langevin dynamics,","cited_arxiv_id":null,"evidence_quote":"Supplies the Stochastic Gradient Langevin Dynamics update rule and burn-in procedure that generate the posterior samples."},{"cited_title":"Federated learning: mmwave mimo radar dataset for testing,","cited_arxiv_id":null,"evidence_quote":"Provides the mmWave radar range-azimuth dataset and Region-of-Interest definitions used in all experiments."},{"cited_title":"On the impact of model compression for bayesian federated learning: An analysis on healthcare data,","cited_arxiv_id":null,"evidence_quote":"Supplies the MCMC-based decentralized Bayesian FL framework that the continual update extends."}],"review_version":1}