{"id":"38847884-1e13-4726-9f9a-7d615f538eaf","arxiv_id":"2505.00318","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"FedEMA, a server-side exponential moving average plus local negative entropy regularizer, reports higher mIoU on Cityscapes and CamVid than several federated baselines.","lead":"The paper proposes FedEMA, a federated learning method that combines server-side exponential moving averaging with local negative entropy regularization to improve street-scene semantic segmentation in autonomous driving. It reports accuracy gains over several federated baselines on Cityscapes and CamVid, but its central claim about preventing temporal catastrophic forgetting is not tested by experiments that simulate changing environments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed forgetting mitigation is never tested: all experiments are static, so the 7.12% mIoU gain does not support the central temporal-catastrophic-forgetting claim.","rationale":"The reader's weakest assumption already identifies the missing temporal experiment, and I agree that this is the load-bearing gap: the paper's motivation is temporal catastrophic forgetting, but Section IV only evaluates static datasets and never measures forgetting. My independent reading of Sections III-C and IV confirms that the convergence theorem is also internally inconsistent: Theorem 1's bound includes λ²γ² as a penalty while the surrounding text claims larger λ speeds convergence, and the 'Negative Entropy regularity' assumption is not proved. These issues reinforce the reader's high correctness risk but do not shift the verdict, because the reader already marked the paper CONDITIONAL pending a temporal benchmark and a complete proof. A temporal-domain re-evaluation test would settle the main empirical claim, so I recommend no change to the conditional verdict.","tokens_in":11845,"tokens_out":6272,"duration_ms":67669,"concrete_test":"Construct a temporally ordered benchmark from Cityscapes or CamVid: partition training images into sequential domain blocks, e.g., by coarse scene attributes, capture sequence, or CamVid's ordered video frames, with early blocks as 'past' and later blocks as 'current'. Train FedEMA and FedProx for equal rounds while shifting each round to the next block; after training on later blocks, measure mIoU on held-out early-block images before and after adaptation. If FedEMA's mIoU drop on the early block is not significantly smaller than FedProx's, the central forgetting-mitigation claim fails. This single check directly tests whether historical fitting capability is preserved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A describes distributing static Cityscapes/CamVid samples across vehicles and running standard Federated Learning rounds; there is no temporal ordering of data, no distribution shift between rounds, and no measurement of old-scene performance after later training. Therefore the reported 7.12% mIoU advantage over FedProx (Table IV) is evidence of generalization on a fixed data mixture, not of preserving historical fitting capability or mitigating temporal catastrophic forgetting, which is the paper's stated motivation (Abstract; Section I). The proof in Section III-C is not a reliable substitute: Theorem 1's bound contains a positive +λ²γ² term, while the text claims larger λ 'suppresses heterogeneity' and 'leads to faster convergence'; the displayed bound worsens as λ grows, and the 'Negative Entropy regularity' assumption is stated without proof. The central claim is thus supported neither by the experiments nor by the theoretical argument as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedEMA, a federated learning method for street-scene semantic segmentation in autonomous driving, combining a server-side exponential moving average of round-wise aggregated models with a vehicle-side entropy regularizer. The authors argue that the EMA preserves historical model fitting capability and the regularizer prevents overfitting to temporal patterns, thereby mitigating temporal catastrophic forgetting in dynamically evolving environments. They provide a convergence analysis (Section III-C) claiming O(1/√R) convergence and conduct experiments on Cityscapes and CamVid with two segmentation backbones, reporting that FedEMA outperforms several FL baselines, including a 7.12% relative mIoU improvement over FedProx(0.005) on Cityscapes with DeepLabv3+.","tokens_in":12030,"tokens_out":6413,"duration_ms":65867,"significance":"If the temporal-catastrophic-forgetting claim were actually demonstrated, the contribution would be a simple and potentially useful addition to federated semantic segmentation: a server-side EMA is easy to implement and the entropy regularizer is a standard regularizer, so the method could be adopted by practitioners. The paper also includes the useful discipline of comparing against multiple FL baselines on two datasets and two architectures, and it reports ablations over the two hyperparameters λ and N. However, the significance is currently conditional. The experiments are entirely static: data are distributed once over vehicles, there is no temporal ordering or distribution shift, and no forgetting metric is reported. The convergence analysis is a proof sketch whose displayed bound does not support the stated insights, and one of its assumptions is unproved. Thus the central advertised contribution, mitigating temporal catastrophic forgetting, is supported neither by the experiments nor by the theory as written.","major_comments":[{"comment":"The paper's central claim is that FedEMA mitigates temporal catastrophic forgetting in dynamically evolving AD environments, but the experimental protocol contains no temporal component. The Cityscapes and CamVid training sets are described as being distributed among vehicles once, and all reported results in Tables IV and V compare methods on the same fixed test sets. There is no task ordering, no distribution shift between rounds, no re-evaluation of previously seen scene types after later training, and no forgetting metric such as backward transfer or old-task accuracy. The 7.12% mIoU advantage on Cityscapes with DeepLabv3+ is therefore evidence of better accuracy on a static data mixture, not of preserving historical fitting capability, and it does not support the abstract's or Section I's catastrophic-forgetting motivation. A temporal continual-FL experiment, such as sequentially introducing scene conditions and measuring old-condition performance, is needed before the central claim can be evaluated.","section":"IV-A, IV-B"},{"comment":"The convergence analysis as written does not imply the stated insights. Theorem 1's bound in Eqs. (9) and (13) contains a positive +λ²γ² term in the numerator; as λ grows, the displayed upper bound becomes looser, and there is no term in which λ multiplies σ². The text's insight (III) that 'larger λ suppresses heterogeneity σ², which leads to faster convergence' is therefore not a consequence of the displayed theorem. In addition, the 'Negative Entropy regularity' assumption, ∥∇H(ω)∥≤γ E_x∥∇log p_k(x;ω)∥, is introduced without proof or citation, and the proof sketch in Steps 2–4 omits the derivation of the client-drift bound in Eq. (11) and the descent lemma in Eq. (12); the step from Eq. (12) to the final bound is not shown. At minimum, the theorem statement and insights must be reconciled with the displayed bound, or the theoretical claims should be withdrawn.","section":"III-C, Eqs. (9), (11), (13)"},{"comment":"There is an unresolved sign inconsistency in the problem definition. Equation (5) and Algorithm 1 line 7 define the vehicle objective as min Lc(ωc)−λH(ωc), which is equivalent to maximizing H(ωc), but Eq. (6) states the constraint ωc = arg min [Lc(ω)+λH(ω)], which minimizes H(ω). These two formulations have opposite effects on the prediction distribution: the latter would sharpen predictions, contradicting the stated goal of preventing overconfident predictions. The formal optimization problem must be corrected and aligned with the implemented update before the method is well defined.","section":"Eq. (6), Algorithm 1"},{"comment":"The empirical superiority statement is broader than the data support. The headline 7.12% improvement is computed only for Cityscapes with DeepLabv3+ against FedProx(0.005). In Table V (TopFormer on CamVid), FedEMA's mIoU is 41.79±0.36, numerically below FedProx(0.005)'s 42.10±0.31, and on several other rows FedEMA is not the best, such as CamVid mPrecision with DeepLabv3+ in Table IV. The claim of achieving 'almost the best performance' should be qualified by the configurations where the proposed method is not statistically distinguishable from or is worse than a baseline, and the abstract's unconditional '7.12% higher mIoU' should be tied to the specific configuration.","section":"Table V"}],"minor_comments":[{"comment":"The terminology for H(ω) is confusing: the paper calls H(ω)=E_x[Σ_k p_k(x;ω) log p_k(x;ω)] the 'negative entropy', but this quantity is the negative of the standard Shannon entropy. Since the algorithm maximizes H, it is actually encouraging flatter, less confident predictions. Please clarify the sign convention so that the text, equations, and algorithm are mutually consistent.","section":"III-B, Eq. (5)"},{"comment":"The experimental setup is under-specified for reproduction: the number of vehicles, the data partitioning strategy (IID versus non-IID, number of classes per vehicle, overlap), the total number of communication rounds, and the exact values of λ and N used for the main results in Tables IV and V are not reported.","section":"IV-A"},{"comment":"The captions of Figures 4 and 5 do not state which dataset and model are used, whether the curves are averaged over multiple seeds, or how 'faster convergence' is quantified. The claim in Section IV-C that a higher λ results in quicker convergence appears to rest on a visual impression; please provide a quantitative convergence criterion, especially because the theoretical bound in Eq. (9) suggests the opposite dependence on λ.","section":"IV-C, Figs. 4 and 5"},{"comment":"MOON with TopFormer achieves near-zero mIoU (0.20 on Cityscapes and 1.03 on CamVid), which likely indicates training instability rather than a meaningful algorithmic comparison. Please add a brief explanation or report adjusted hyperparameters for that configuration.","section":"Table V"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: FedEMA is a simple, easy-to-implement method—server keeps an EMA of aggregated models and vehicles add a negative entropy regularizer to their local loss—and on the static Cityscapes/CamVid splits they report it beats FedProx, FedDyn, MOON, and FedIR by a useful margin (up to 7.12% mIoU with DeepLabv3+ on Cityscapes). That part is credible on its face: the comparison set is reasonable, they test two backbones, report error bars, and include ablations for λ and window size. The method is incremental—EMA and entropy regularization are both established tools—but the specific combination for federated street-scene segmentation is not something I have seen, and the empirical story is coherent.\n\nThe soft spot is large. The paper's motivation, stated in the abstract and introduction, is mitigating temporal catastrophic forgetting in dynamically evolving environments. The experiments never test that. Section IV-A distributes static Cityscapes and CamVid samples across vehicles and runs ordinary FL rounds. There is no temporal ordering, no distribution shift across rounds, and no re-evaluation of old scenes after training on new ones. So the 7.12% mIoU advantage is evidence about generalization on a fixed mixture, not about preserving historical fitting capability. The paper effectively claims A but measures B. That is the main thing I would tell any referee.\n\nThe theoretical section does not rescue it. Theorem 1's bound contains a positive λ²γ² term, so the displayed bound worsens as λ increases, while the text claims larger λ suppresses heterogeneity and leads to faster convergence. The negative-entropy regularity assumption is given without proof, and the proof sketch (Steps 1–4) is not a complete derivation. This is a real internal inconsistency, though it is secondary to the missing temporal evaluation.\n\nMinor points: the \"novel framework\" language overstates; the pieces are known. Self-citations are frequent but they are on-topic and not inherently a problem.\n\nRecommendation: I would not desk reject this. The algorithm is concrete, the baseline comparison is useful, and the flaw is fixable with a temporal benchmark. Send it to review, but the referee should require either (a) an experiment that actually measures forgetting—e.g., task/domain-incremental FL where old scenes are re-evaluated—or (b) a revised scope that drops the temporal-catastrophic-forgetting claim and presents FedEMA as a calibration/regularization method for static federated segmentation. The convergence theorem should be completed or removed.","headline":"FedEMA combines server-side EMA and negative entropy regularization into a simple federated segmentation method; the static benchmark gains look real, but the paper's headline claim about temporal catastrophic forgetting is never actually tested.","tokens_in":12505,"tokens_out":2505,"would_cite":false,"duration_ms":25728,"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":"FedEMA claims temporal catastrophic forgetting in federated autonomous-driving perception is mitigated by a server-side exponential moving average plus a vehicle-side negative entropy regularizer, with a relative 7.12% mIoU gain on…","keywords":["federated learning","autonomous driving","semantic segmentation","catastrophic forgetting","exponential moving average","negative entropy regularization","DeepLabv3+","TopFormer"],"falsifier":"Run FedEMA on a temporally ordered stream — train on scene group A, adapt on scene group B, then re-measure mIoU on A — and compare the drop with FedAvg's; if the EMA model forgets A at the same rate, the historical-preservation claim fails. Separately, compute $\\|\\nabla H(\\omega)\\|$ and $\\mathbb{E}_x\\|\\nabla \\log p_k(x;\\omega)\\|$ on the trained segmenters to check the assumed 'negative entropy regularity' inequality; a clean violation would leave the convergence bound unsupported.","tokens_in":11659,"feed_emoji":"🚗","tokens_out":14929,"duration_ms":128683,"temperature":0.7,"pith_summary":"This paper argues that temporal catastrophic forgetting in federated autonomous-driving perception can be countered without replay buffers or task-specific memory. Its recipe is two changes to ordinary federated training: the server broadcasts an exponential moving average of past global models instead of the latest aggregate, and each vehicle adds a negative entropy regularizer to the segmentation loss to suppress overconfident, temporally overfit predictions. The combined scheme, FedEMA, reaches 60.62% mean Intersection-over-Union (mIoU) on Cityscapes with DeepLabv3+, a relative 7.12% gain over the best comparison baseline, and it is the top result among the federated baselines on that benchmark. The paper also proves an $O(1/\\sqrt{R})$ convergence rate that matches standard federated optimization, and the ablations show the EMA window size and regularizer coefficient are two tunable controls. A sympathetic reader would care because the recipe is simple, communication-friendly, and targets the round-by-round overwriting of historical knowledge that standard federated averaging does not address.","feed_headline":"Federated street-scene segmentation up 7.12% mIoU via EMA and entropy","feed_subtitle":"The recipe counteracts round-by-round overwriting that erases old scenes as driving conditions change.","key_machinery":"The load-bearing object is the server-side EMA model $\\omega^r_{EMA}$, updated as $\\omega^r_{EMA} = \\beta\\omega^{r-1}_{EMA} + (1-\\beta)\\omega$ with $\\beta = 2/(N+1)$, together with the vehicle-side negative entropy term $-\\lambda H(\\omega_c)$ in the local objective. The EMA acts as a slow-moving historical model that is broadcast to vehicles every round, so the current aggregate only contributes a fraction $1-\\beta$ of the next starting point; the entropy term, defined on per-pixel softmax probabilities, keeps local models from becoming overconfident on the current temporal slice. The convergence proof treats the EMA update as momentum, bounds client drift using the 'negative entropy regularity' inequality $\\|\\nabla H(\\omega)\\| \\le \\gamma\\mathbb{E}_x\\|\\nabla \\log p_k(x;\\omega)\\|$, and then applies L-smoothness to obtain the $O(1/\\sqrt{R})$ bound. In the ablations, window size $N$ governs how much history survives, and $\\lambda$ controls how strongly complexity is penalized, giving the method its two tuning knobs.","core_discovery":"At its center, FedEMA claims that the overwriting update of federated averaging is the mechanism of temporal forgetting, and that a recursive server-side exponential moving average is sufficient to preserve historical fitting capability: $\\omega^r_{EMA} = \\beta\\omega^{r-1}_{EMA} + (1-\\beta)\\omega$, with $\\beta = 2/(N+1)$, where $N$ is the EMA window. Because the EMA model, not the freshly aggregated model, is distributed to vehicles, every local training round starts from a blend of old and new knowledge. The vehicle-side counterweight is the negative entropy regularizer $-\\lambda H(\\omega_c)$ with $H(\\omega_c) = \\mathbb{E}_x[\\sum_{k=1}^K p_k(x;\\omega_c)\\log p_k(x;\\omega_c)]$, which the paper interprets as penalizing model complexity and preventing the EMA-introduced temporal patterns from being overfit. The two mechanisms define the dual objective in Eq. (6), and the convergence analysis reformulates the EMA as momentum to show client drift is bounded and the rate is $O(1/\\sqrt{R})$. On the paper's own terms, FedEMA is a demonstration that forgetting can be mitigated by changing only how models are averaged and regularized, without changing the underlying segmentation architecture or data sharing.","pith_inferences":["Because the paper validates on static splits rather than a temporal stream, the direct test of its forgetting claim would be to order Cityscapes or CamVid images by recording time or season, train in consecutive chunks, and re-measure old-chunk mIoU after adaptation; if FedEMA's old-scene accuracy declines as fast as or faster than FedAvg's, the memory mechanism is not doing the claimed work.","The EMA-as-momentum reformulation invites combining FedEMA with heterogeneity-aware client regularizers such as proximal or contrastive terms, addressing device skew and temporal drift simultaneously rather than separately.","A practical extension is to make the EMA window $N$ and the entropy coefficient $\\lambda$ adaptive: shrink the window when a distribution-drift detector fires, and grow $\\lambda$ when predictions become overconfident, removing the need to grid-search these two knobs.","If the mechanism is as general as the paper suggests, the same two changes should transfer to other federated dense-prediction tasks in autonomous driving, including depth estimation and panoptic segmentation, which share the overwriting failure mode."],"forward_implications":["On Cityscapes with DeepLabv3+, FedEMA reports 60.62% mIoU versus 56.59% for the best baseline FedProx, a relative 7.12% improvement, with mF1, mPrecision, and mRecall also higher.","The gains transfer at least partially across architectures and datasets: FedEMA reaches 74.73% mIoU on CamVid with DeepLabv3+, the best result in that table, and it is the best Cityscapes result with the TopFormer transformer backbone, though on CamVid with TopFormer its 41.79% trails FedProx's 42.10%.","Ablations show smaller EMA window sizes improve both convergence speed and final performance, while larger $\\lambda$ accelerates convergence with little effect on final mIoU, giving two hyperparameters that trade adaptability against stability.","The proved $O(1/\\sqrt{R})$ rate means adding the entropy regularizer does not degrade the asymptotic convergence order of federated optimization, and the analysis ties the regularizer to reduced client drift."],"supporting_citations":[{"why":"Defines FedAvg, the round-by-round overwriting aggregation scheme that FedEMA modifies with server-side EMA.","marker":"[9]"},{"why":"Cityscapes is the primary benchmark for the headline 7.12% mIoU comparison.","marker":"[15]"},{"why":"CamVid provides the second, simpler benchmark used to test transfer of the gains.","marker":"[16]"},{"why":"FedProx is the best-performing baseline on Cityscapes whose 56.59% mIoU defines the 7.12% margin.","marker":"[28]"},{"why":"FedDyn is a dynamic-regularization federated baseline included in the comparison tables.","marker":"[29]"},{"why":"DeepLabv3+ is the CNN segmentation model on which the main reported results are obtained.","marker":"[36]"},{"why":"TopFormer is the transformer segmentation model used to show the method is not tied to one architecture.","marker":"[37]"},{"why":"MOON is a model-contrastive federated baseline whose divergence on TopFormer shows architecture-dependent behavior in the comparison.","marker":"[39]"}],"fun_headline_variants":["FedEMA: EMA + entropy counters temporal forgetting, mIoU up 7.12%","Anti-forgetting FL: EMA blend + entropy penalty raise mIoU 7.12%","Temporal forgetting? EMA + entropy regularizer lifts mIoU 7.12%","Federated EMA + negative entropy: +7.12% mIoU on street scenes","EMA + entropy counter overfitting: mIoU +7.12% in federated AD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that splitting static Cityscapes and CamVid samples across vehicles and running standard FL rounds reproduces temporal catastrophic forgetting — but the experiments introduce no temporal shift and re-test no old scenes, and the convergence proof depends on an unverified 'negative entropy regularity' inequality.","fun_headline_variants_meta":{"raw":{"variants":["FedEMA: EMA + entropy counters temporal forgetting, mIoU up 7.12%","Anti-forgetting FL: EMA blend + entropy penalty raise mIoU 7.12%","Temporal forgetting? EMA + entropy regularizer lifts mIoU 7.12%","Federated EMA + negative entropy: +7.12% mIoU on street scenes","EMA + entropy counter overfitting: mIoU +7.12% in federated AD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001588,"raw_usage":{"total_tokens":6378,"prompt_tokens":1040,"completion_tokens":5338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":5218}},"tokens_in":656,"tokens_out":5338,"duration_ms":39206,"temperature":1.0,"reasoning_tokens":5218,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:45:17.918223+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FedEMA on a temporally ordered stream — train on scene group A, adapt on scene group B, then re-measure mIoU on A — and compare the drop with FedAvg's; if the EMA model forgets A at the same rate, the historical-preservation claim fails. Separately, compute $\\|\\nabla H(\\omega)\\|$ and $\\mathbb{E}_x\\|\\nabla \\log p_k(x;\\omega)\\|$ on the trained segmenters to check the assumed 'negative entropy regularity' inequality; a clean violation would leave the convergence bound unsupported.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Defines FedAvg, the round-by-round overwriting aggregation scheme that FedEMA modifies with server-side EMA."},{"cited_title":"Segmentation and recognition using structure from motion point clouds,","cited_arxiv_id":null,"evidence_quote":"CamVid provides the second, simpler benchmark used to test transfer of the gains."},{"cited_title":"Federated learning based on dynamic regularization,","cited_arxiv_id":null,"evidence_quote":"FedDyn is a dynamic-regularization federated baseline included in the comparison tables."},{"cited_title":"Topformer: Token pyramid transformer for mobile semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"TopFormer is the transformer segmentation model used to show the method is not tied to one architecture."}],"review_version":1}