{"id":"ee5e65dd-8ba8-4164-b5ce-dd81ecc40a79","arxiv_id":"2507.07130","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Ampere trains the device and server halves of a split model sequentially with a single activation transfer and consolidated server activations, cutting communication and computation while improving accuracy on non-IID data.","lead":"Ampere is a split federated learning system that trains the device and server halves of a model one after another, exchanging activations only once instead of in every training step. It reports large cuts in communication and computation, plus higher accuracy on non-IID image benchmarks, which matters for training on phones and edge devices without sharing raw data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 99.1% communication reduction relies on Eq. (27)'s once-per-epoch model exchange, but Algorithm 1 lines 6-7 send and receive models after every mini-batch; without fixing this inconsistency or specifying H, the headline number is unsupported.","rationale":"The reader's weakest_assumption identifies the same point, and I agree it is the most load-bearing concern. The abstract and Section 5 headline 'reduces device-server communication overhead by up to 99.1%' is a central, quantitative claim, and it is derived from Eq. (27) by counting model exchanges once per epoch. Algorithm 1, however, shows the device sending its model to the server and receiving the aggregated model inside the per-mini-batch loop (lines 6-7). Unless H=1 or the pseudocode is inaccurate, the real communication volume is H times larger than the formula, so the claimed reduction is not established by the written method. The paper does not state H or the actual aggregation interval used in the experiments, so a reader cannot resolve the contradiction from the text. This is an internal inconsistency rather than a disagreement with an external consensus, and it directly affects the paper's primary quantitative contribution. Other weaknesses, such as the lack of error bars, absence of code, and reliance on imported convergence analysis, are secondary and addressable; the communication accounting is the main claimed result. The system idea itself remains plausible, and the ablation study provides some evidence for the activation-consolidation benefit. Therefore, the appropriate verdict remains CONDITIONAL, pending clarification of the communication schedule and alignment of the algorithm with the cost model.","tokens_in":20640,"tokens_out":7578,"duration_ms":82187,"concrete_test":"Re-run the communication accounting with the actual per-device batch count H used in the experiments, or inspect the released code to determine whether the Send/Receive in Algorithm 1 executes once per epoch or once per mini-batch. If it executes once per mini-batch, recompute Table 5 using C = 2NH(s(d)+s(aux)) + s(act) and compare the resulting percentage reductions against the claimed 99.1%; a drop of more than a few points would require revising the headline communication claim and Eq. (27).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the 'up to 99.1%' communication reduction over SFL. That figure is computed from Eq. (27), C = 2N(s(d)+s(aux)) + s(act), which assumes one device-block model round-trip per epoch. However, Algorithm 1's DeviceTraining loop places the Send and Receive operations inside the inner mini-batch loop (lines 6-7), after each of the H iterations within an epoch. If H > 1, the actual model-exchange volume is 2NH(s(d)+s(aux)), a factor of H larger than Eq. (27). The paper never states H or the aggregation frequency used in the experiments; Section 5.1 only says each device trains on 10,000 local samples before transferring its updated model, which conflicts with the per-batch transfer in the algorithm. This is an internal inconsistency, not a disagreement with an outside baseline. If the implementation follows the algorithm, the communication reductions in Table 5 shrink materially: for the MobileNet-L CIFAR-10 row, Ampere's 0.80 GB becomes roughly 0.80H GB, and even H=20 would drop the reduction versus the best SFL baseline from about 99.1% to the mid-80s. Because the paper's headline claim depends directly on Eq. (27), this inconsistency must be resolved before the primary contribution is established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Ampere, a split federated learning system that trains the device block and server block sequentially rather than by end-to-end backpropagation. A lightweight auxiliary network lets devices train the device block locally; activations are then transferred once to the server, where they are consolidated into a single activation set on which the server block is trained. The authors claim that, compared with SFL baselines, Ampere reduces device-server communication by up to 99.1%, reduces on-device computation by up to 93.13%, reduces training time by up to 94.6%, and improves accuracy by up to 13.26%, with lower variance under non-IID data. The evaluation uses four models (MobileNet-L, VGG-11, Swin-T, ViT-S) on CIFAR-10 and Tiny ImageNet, on a 120-device Jetson testbed, against SplitFed, PiPar, SplitFed+SCAFFOLD, and SplitGP baselines, plus an ablation of activation consolidation.","tokens_in":20883,"tokens_out":9122,"duration_ms":95356,"significance":"If the claims hold, Ampere would be a meaningful advance: it eliminates the per-iteration activation/gradient exchange that dominates SFL communication, reduces the split-point trade-off between on-device computation and communication, and mitigates non-IID degradation by training the server block on consolidated activations. Strengths of the paper include the breadth of the evaluation (four architectures, two datasets, four SFL baselines), the use of a physical heterogeneous testbed rather than pure simulation, an ablation isolating activation consolidation, and a convergence analysis that builds on existing FedAvg and decoupled-learning results. The main weaknesses are an internal inconsistency between the communication-cost formula and the training algorithm, and the absence of statistical anchoring for the empirical claims. Neither weakness appears impossible to fix, but both are load-bearing for the headline results.","major_comments":[{"comment":"The communication-cost formula in Eq. (27), C = 2N(s(d)+s(aux)) + s(act), counts exactly one device-block model exchange per epoch. However, Algorithm 1 places the Send and Receive operations for {θ(d)_k, θ̃(d)_k} inside the inner loop over h (lines 6–7), i.e., once per mini-batch rather than once per epoch. Under the algorithm as written, the model-exchange volume is 2NH(s(d)+s(aux)) with H the number of mini-batches per epoch, not 2N(s(d)+s(aux)). Section 5.1 states that each selected device trains on 10,000 local samples before transferring its updated model, which also does not match a per-mini-batch transfer. Since Table 5 and the headline 99.1% communication reduction are computed from Eq. (27), this inconsistency must be resolved: either move the aggregation outside the h loop in Algorithm 1, or define H and the communication interval explicitly, or revise Eq. (27) and all subsequent communication comparisons. This is a load-bearing issue for the paper's primary quantitative claim.","section":"§4.2, Eq. (27), Algorithm 1"},{"comment":"The empirical accuracy, training-time, communication, and computation results are reported as single values with no indication of the number of runs, no error bars, and no per-setting hyperparameters (optimizer, learning rate, batch size, early-stopping threshold details). The claim that Ampere 'reduces standard deviation of accuracy by 53.39%' in Observation 3 is computed across non-IID degrees α, not across repeated training runs, so it does not address run-to-run variance. Because the central claim is 'high accuracy' (up to 13.26% improvement), the evaluation should include multiple independent runs with means and standard deviations, or at minimum a statement of how many seeds were used and why single runs are representative. Without this, the accuracy and training-time improvements are not statistically anchored.","section":"§5, Figures 8–11, Tables 4–5"},{"comment":"The auxiliary-network dimension ratio 0.5 is selected from a single accuracy/computation trade-off curve measured for MobileNet-L on CIFAR-10. The paper states that the ratio is user-adjustable, but it is a free parameter used in all subsequent experiments, and Figure 7 does not show whether the choice transfers to other architectures or datasets. Since the auxiliary network is a core component of the method, please include a sensitivity analysis for at least one additional model (e.g., a transformer) or one additional dataset, or explicitly state that the sensitivity has not been evaluated outside the single configuration shown.","section":"§3.2.2, Figure 7"}],"minor_comments":[{"comment":"The label 'CF edGrail' appears to be a typographical artifact; it should be 'C_Ampere'. Also, the sentence 's(a) ≪ s(s)' should use s(aux) consistently.","section":"§4.2, Eq. (27)"},{"comment":"Algorithm 1 aggregates models from all K devices at the server, whereas Section 5.1 says 12 devices are randomly selected to participate in each training round. Please clarify how partial participation is represented in the algorithm and how the aggregation loop over [K] should be read in that setting.","section":"Algorithm 1 and §5.1"},{"comment":"The temporary accuracy drop when switching from device-block training to server-block training is described qualitatively, but the reason for the recovery is not analyzed. A short explanation of the dynamics (e.g., the server block starts from random initialization on frozen activations) would help the reader interpret the plots.","section":"§5.2.1, Figure 8"},{"comment":"The column header of Table 2 appears corrupted or misaligned in the manuscript, and the relationship between 'Model', 's(act)', 's(d)', 's(aux)', and 's(s)' is unclear from the table alone. Please re-typeset the table so that each column is explicitly labeled and the units are unambiguous.","section":"Table 2"},{"comment":"The symbol N is used in Eq. (5) as total training epochs and in Table 4 the epochs are split into N(d) and N(s). Please state explicitly which N appears in the communication-cost formulas and whether the same N is used for both FL/SFL and Ampere when comparing communication volume.","section":"Eqs. (5), (27)–(31)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper proposes a genuinely different SFL variant. Train the device block locally with a lightweight auxiliary head, transfer activations exactly once, then train a single server block on the consolidated activation set. That is a clean, plausible way to kill the per-iteration activation traffic, and the idea is distinct from FedGKT, SplitGP, EcoFed, and related work. The empirical setup is real and broad: 120 heterogeneous Jetson Nanos, four models, two datasets, four SFL baselines. The ablation showing activation consolidation adds 6–13% accuracy is a good, clean demonstration that the mechanism matters.\n\nBut the headline number does not survive close reading. Equation (27) says total communication is 2N(s(d)+s(aux)) + s(act), i.e., one model exchange per epoch. Algorithm 1 lines 6–7 send and receive models after every mini-batch inside the inner loop. If H > 1, the model-exchange volume is H times larger than the formula. The paper never states H, and Section 5.1 says each device trains on 10,000 local samples before transferring its model—which hints at once-per-epoch aggregation, contradicting the algorithm. The 99.1% reduction is computed from Eq. (27), so this internal inconsistency is load-bearing. It is fixable if the implementation really aggregates per epoch, but as written, the proof does not match the pseudocode.\n\nOther soft spots are less severe but still real. The accuracy numbers come from single runs, with no error bars or code. The convergence theorems are imported from prior work under strong-convexity assumptions that the evaluated CNNs and transformers do not satisfy, so the formal section is mostly decoration. The abstract and Figure 2 imply Ampere beats classic FL accuracy, but FL accuracy is never measured, only its communication is estimated. These are addressable, not fatal.\n\nOn the credit side, the idea is coherent and the practical results, if reproducible, would matter for bandwidth-limited edge training. The communication-efficiency claim is the main unresolved piece; the rest is mostly a matter of added rigor.\n\nWho gets value: researchers working on split federated learning, edge ML systems, and communication-efficient distributed training. A serious referee should see it—the core idea deserves review and likely a major revision. My recommendation: send to peer review, with the expectation that the authors fix the algorithm/equation mismatch, add statistical anchors and reproducibility details, and soften the FL-accuracy overclaim.","headline":"Useful SFL variant with a clean core idea, but the headline communication claim is undermined by an algorithm/proof mismatch that needs fixing before acceptance.","tokens_in":21470,"tokens_out":4556,"would_cite":true,"duration_ms":47282,"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":"Ampere replaces split federated learning's per-step activation and gradient exchanges with a one-shot activation transfer, claiming up to 99.1% less device-server communication and up to 13.26% higher accuracy.","keywords":["split federated learning","communication efficiency","on-device computation","unidirectional inter-block training","auxiliary network","activation consolidation","non-IID data","edge computing"],"falsifier":"Run Ampere with more than one mini-batch per epoch, count every byte exchanged on the wire, and compare the total with Equation (27); if the measured volume scales with the number of mini-batches rather than matching the formula, the claimed 99.1% communication reduction does not hold for that configuration.","tokens_in":20352,"feed_emoji":"⚡","tokens_out":7218,"duration_ms":58332,"temperature":0.7,"pith_summary":"Ampere is a split federated learning system that removes the central cost of split learning: the per-iteration exchange of activations and gradients between devices and the server. Rather than training the device and server halves of a network jointly by end-to-end backpropagation, Ampere first trains the device block on-device with a small auxiliary network, sends the resulting activations to the server once, and then trains the server block on the consolidated activation set. The paper reports that this reduces device-server communication by up to 99.1% and on-device computation by up to 93.13% compared with existing split federated learning baselines, while improving model accuracy by up to 13.26% and reducing the standard deviation of accuracy by 53.39% under non-IID data. If correct, this would make split federated learning practical on bandwidth-limited edge devices rather than only on well-connected infrastructure.","feed_headline":"One activation upload per device cuts split-FL traffic by 99%","feed_subtitle":"Sequential block training, no per-step gradient exchange, cuts traffic up to 99% and lifts accuracy up to 13%","key_machinery":"The load-bearing mechanism is the lightweight auxiliary network $\\tilde{\\theta}^{(d)}$ attached to the device block. It is a two-layer network whose first layer mirrors the server block's first layer at half dimension and whose second layer is a fully connected layer with the same loss function as the server block. Its job is to decouple device and server training: it lets the device compute gradients for its own block locally, so the iterative activation-gradient loop of split federated learning can be replaced by one-shot activation transfer. The auxiliary network also shapes what features the device block extracts, since shallow, generalizable representations transfer better to the downstream server block than task-specific ones.","core_discovery":"The paper's central claim is that the interdependence of device and server blocks in split federated learning is unnecessary, and that breaking it by sequential training improves everything at once: communication, computation, accuracy, and robustness to data heterogeneity. The mechanism is unidirectional inter-block training: a lightweight auxiliary network lets each device compute a local loss and update its own block without gradients from the server; once the device block converges, activations are uploaded to the server exactly once; the server then trains a single consolidated server block on merged activations from all devices. The paper argues that this activation consolidation turns skewed per-device activations into a more homogeneous dataset, which is why accuracy improves on non-IID data. It also argues that because the one-time activation transfer cost is negligible compared with model exchanges over many epochs, communication becomes dominated by the small device-block exchange, which is less than the full model exchange of classic federated learning.","pith_inferences":["If Ampere's one-shot transfer holds up, it suggests a general design pattern for split learning: any method that can cheaply predict the downstream task's needs at the split boundary can replace iterative backpropagation with a single forward pass, at the price of a gap between local-loss features and global-task features.","The auxiliary network dimension ratio (fixed at 0.5 in the paper) is a tunable knob; the paper's own ablation suggests accuracy saturates around 0.5 while computation grows linearly, so the optimal ratio may shift with model scale or dataset difficulty.","A natural stress test would be to run Ampere with more than one local mini-batch per epoch and count actual bytes sent, to see whether the communication formula in Equation (27) correctly predicts measured traffic when device blocks are exchanged more than once per epoch."],"forward_implications":["Split federated learning can be made communication-competitive with classic federated learning, since per-step activation and gradient transfers are replaced by a single activation upload per device.","Devices can exit training much earlier; the paper reports device participation time reductions of 44.3% to 96.3%, freeing devices for other work.","Server-side training on consolidated activations yields accuracy gains of up to 13.26% and a 53.39% reduction in accuracy standard deviation across non-IID settings.","The split point no longer requires a trade-off between on-device computation and communication; Ampere claims the same split point minimizes both.","The approach applies to both CNNs and vision transformers, and to CIFAR-10 and Tiny ImageNet, suggesting it generalizes across architectures."],"supporting_citations":[{"why":"Defines the split federated learning baseline whose per-iteration activation and gradient exchanges Ampere replaces.","marker":"[5]"},{"why":"Supplies the FedAvg aggregation rule used for device blocks and the classic federated learning communication baseline.","marker":"[1]"},{"why":"Provides the convergence theorem and assumptions used to prove device block convergence in Theorem 1.","marker":"[24]"},{"why":"Provides the decoupled learning bound used to prove server block convergence in Theorem 2.","marker":"[25]"},{"why":"SCAFFOLD, combined with SplitFed, is the non-IID robustness baseline Ampere compares against.","marker":"[16]"},{"why":"PiPar, a pipeline-parallelism baseline, represents communication-computation overlap that Ampere claims to surpass.","marker":"[11]"},{"why":"SplitGP, a local-plus-global loss baseline, is used as a comparison for accuracy and communication.","marker":"[12]"}],"fun_headline_variants":["Single upload: cut split-FL comms 99% and boost accuracy 13%","Unidirectional split-FL: one upload, 99% less traffic, better accuracy","Ampere: sequential split training slashes traffic 99%, lifts accuracy 13%","Lose the gradient loop: split-FL comms down 99%, accuracy up 13%","One-shot activations make split-FL 99% leaner and 13% smarter"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The communication cost formula in Equation (27) counts one model exchange per device per epoch, but Algorithm 1 has devices upload their models after every mini-batch; if an epoch has more than one mini-batch, the actual communication volume is proportionally larger.","fun_headline_variants_meta":{"raw":{"variants":["Single upload: cut split-FL comms 99% and boost accuracy 13%","Unidirectional split-FL: one upload, 99% less traffic, better accuracy","Ampere: sequential split training slashes traffic 99%, lifts accuracy 13%","Lose the gradient loop: split-FL comms down 99%, accuracy up 13%","One-shot activations make split-FL 99% leaner and 13% smarter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1671,"prompt_tokens":1051,"completion_tokens":620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":504}},"tokens_in":667,"tokens_out":620,"duration_ms":5667,"temperature":1.0,"reasoning_tokens":504,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:07:21.460868+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Ampere with more than one mini-batch per epoch, count every byte exchanged on the wire, and compare the total with Equation (27); if the measured volume scales with the number of mini-batches rather than matching the formula, the claimed 99.1% communication reduction does not hold for that configuration.","supporting_citations":[{"cited_title":"SplitFed: When Federated Learning Meets Split Learning,","cited_arxiv_id":null,"evidence_quote":"Defines the split federated learning baseline whose per-iteration activation and gradient exchanges Ampere replaces."},{"cited_title":"Communication-Efficient Learning of Deep Networks from Decentralized Data,","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg aggregation rule used for device blocks and the classic federated learning communication baseline."},{"cited_title":"On the Convergence of FedAvg on Non-IID Data,","cited_arxiv_id":null,"evidence_quote":"Provides the convergence theorem and assumptions used to prove device block convergence in Theorem 1."},{"cited_title":"Decoupled Greedy Learning of CNNs,","cited_arxiv_id":null,"evidence_quote":"Provides the decoupled learning bound used to prove server block convergence in Theorem 2."},{"cited_title":"SCAFFOLD: Stochastic Controlled Averaging for Feder- ated Learning,","cited_arxiv_id":null,"evidence_quote":"SCAFFOLD, combined with SplitFed, is the non-IID robustness baseline Ampere compares against."},{"cited_title":"PiPar: Pipeline Parallelism for Collaborative Machine Learning,","cited_arxiv_id":null,"evidence_quote":"PiPar, a pipeline-parallelism baseline, represents communication-computation overlap that Ampere claims to surpass."},{"cited_title":"SplitGP: Achieving Both Generalization and Personalization in Federated Learning,","cited_arxiv_id":null,"evidence_quote":"SplitGP, a local-plus-global loss baseline, is used as a comparison for accuracy and communication."}],"review_version":1}