{"id":"aabc4d02-c223-4667-8903-9fd5c8b314c5","arxiv_id":"2411.15481","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CAMA combines FedZero's renewable-energy-aware client selection with HeteroFL's ordered dropout to adapt model sizes dynamically, reportedly achieving faster convergence and lower energy use in simulated federated learning.","lead":"CAMA is a federated learning system that lets clients train smaller or larger versions of the same model depending on how much renewable energy and computing power they have. The paper reports that CAMA trains faster and uses less energy than the FedZero baseline in simulations on CIFAR-10 and MNIST.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy claim rests on Eq. 3's unvalidated linear scaling; ordered dropout's channel shrinkage implies FLOPs scale ~mr^2, and fixed overheads may erode savings.","rationale":"The reader's weakest_assumption correctly identifies Eq. 3 as the load-bearing point for the energy-efficiency claim. I agree with that identification, and I sharpen it in two ways. First, the paper defines model rates in terms of channel shrinkage (Section 2.2), so the linear scaling of energy with mr is inconsistent with the quadratic FLOP scaling that ordered dropout implies for compute-bound layers. Second, the absence of fixed-overhead terms makes the linear model optimistic in the opposite direction: real devices have constant costs per batch that do not shrink with model size. Both readings point to the same conclusion: Table 2 and Figure 3 are not energy measurements but outputs of an unvalidated algebraic model. I considered other candidate concerns: the lack of a reported fairness metric undermines the 'equitable participation' claim, the convergence comparison rests on a small number of rounds with high variance, and there is no ablation separating dynamic model-size selection from the client-selection changes. These are real weaknesses, but they are secondary to the energy model because the paper's core contribution and headline savings are defined by Eq. 3. If that equation is materially wrong, the central quantitative result collapses; if it is right, the other issues can be repaired with additional experiments. The proposed hardware benchmark directly tests the assumption and settles whether the concern lands. The reader's CONDITIONAL verdict remains appropriate: the method is plausible and the code is available, but the energy evidence must be re-derived from measured data before the central claim can be accepted.","tokens_in":6121,"tokens_out":7382,"duration_ms":67314,"concrete_test":"On a T4 or similar GPU, run the CAMA implementation with ResNet18/CIFAR-10 and measure wall-clock energy (via nvidia-smi or an external power meter) for model rates mr ∈ {1.0, 0.5, 0.25, 0.125, 0.0625} using ordered dropout, holding batch size and epoch count fixed. Compare the measured energy ratios to the linear prediction mr. If, for example, mr=0.5 yields a measured energy ratio above 0.6, or mr=0.25 above 0.35, then Eq. 3 overstates the energy benefit. Recompute Table 2 with the measured ratios and check whether CAMA still achieves the reported margin over FedZero; if the margin disappears or inverts, the central claim fails without a revised energy model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central energy-efficiency claim is computed, not measured: Table 2 and Figure 3 derive every energy value from Eq. 3, Ec,i = ep × bc × mr, which postulates that per-batch energy decreases linearly with model rate mr. This proportionality is never validated on real hardware, and it conflicts with the paper's own definition of mr. Section 2.2 sets complexity levels via a hidden-channel shrinkage ratio of 0.5, so a client at mr=0.5 trains a network whose input and output channels (and hence matrix-multiply work) are both halved; per-batch FLOPs and energy for compute-bound layers scale roughly as mr^2, not mr. If real devices instead exhibit less-than-linear energy scaling due to memory-bandwidth limits or, more importantly, have fixed per-batch overheads such as data loading, kernel launch, and communication, then the reported savings of 0.31 kWh over FedZero in Table 2 could shrink or even reverse for small model rates. Because the paper presents no direct power measurements and the simulation uses this unverified algebraic relation, the headline 'energy-efficient' result stands on a load-bearing assumption rather than on measured behavior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CAMA, a carbon-aware federated learning framework that extends FedZero by dynamically reducing the model size assigned to clients based on their available renewable energy and spare computing capacity. The server estimates each client's batch-processing capability, assigns a model rate using an ordered-dropout partitioning scheme from HeteroFL, and aggregates updates from models of different sizes. Experiments on CIFAR-10 and MNIST with ResNet-18 and a CNN compare CAMA against FedZero over 15 rounds, reporting faster convergence, lower cumulative energy usage, and equitable client participation. The source code is released on GitHub.","tokens_in":6345,"tokens_out":5496,"duration_ms":49767,"significance":"If the central claims hold, CAMA is a practically relevant extension of FedZero: it allows clients with scarce green energy to contribute smaller models rather than being excluded, and the reported accuracy improvements over FedZero are measured directly from training curves and are independent of the energy model. The paper also provides public code, uses real Solcast solar traces, and reports results for two datasets and two model architectures. However, the headline energy savings are computed from an assumed linear energy model rather than measured on hardware, the fairness and scalability claims are never quantified, and there are structural inconsistencies in the definition of model rate. The learning/convergence result appears plausible, but the energy-efficiency claim requires validation before the paper's central message can be accepted.","major_comments":[{"comment":"The energy-efficiency claim rests entirely on the unvalidated linear relation Ec,i = ep × bc × mr. The authors never report direct power or energy measurements on any hardware; Table 2 and Figure 3 are computed from this equation. The paper's own model construction in Section 2.2 shrinks both input and output channels by a factor of 0.5 per complexity step, so for convolutional and linear layers the FLOP count, and hence compute-bound energy, scales approximately as the square of the per-dimension model rate, not linearly. This is compounded by fixed per-batch overheads such as data loading, kernel launches, and communication, which Eq. (3) omits. Because Algorithm 2 chooses the smallest model rate precisely to minimize ep×bc×mr, the reported 0.31 kWh saving over FedZero in Table 2 is partly an artifact of the assumed model rather than an empirical result. The authors should either measure device energy at the model-rate levels they use or re-derive the savings from a validated scaling law and quantify the impact of fixed overheads.","section":"Section 3, Eq. (3)"},{"comment":"The definition of model rate is internally inconsistent. The text lists complexity levels with model rates {1, 0.5, 0.25, 0.125, 0.625}, but Algorithm 2 halves mr at each step, so the fourth halving returns 0.0625, not 0.625; the default value µ = 0.0625 later in the section confirms that 0.625 is likely a typo. Moreover, Section 2.2 says class b trains on '50% of the global model's parameters', but ordered dropout sends a subnetwork with layer dimensions (dg×m) × (kg×m), so m = 0.5 corresponds to 25% of the parameters in that layer, not 50%. This ambiguity is not merely cosmetic: if mr denotes a parameter fraction, Eq. (3) is inconsistent with the channel-shrinkage implementation; if mr denotes the channel ratio, the linear energy scaling in Eq. (3) is still unjustified. The manuscript must fix the rate list and define mr precisely.","section":"Section 2.2 and Algorithm 2"},{"comment":"The claims of equitable client participation and fair selection are never evaluated. Section 2.1 and the abstract assert that CAMA 'ensures equitable client participation', but no experiment reports participation counts, selection frequencies, or any fairness metric such as Jain's index or the Gini coefficient. The only reported outcomes are accuracy and energy. Either add a quantitative fairness evaluation with the actual selection distributions or remove the fairness claim from the abstract and contributions.","section":"Section 2.1 and Section 3"},{"comment":"The scalability claim is unsupported by the experiments. The abstract and conclusion state that CAMA 'scales efficiently to handle large numbers of clients', but the experiments fix the number of clients at 100 with a maximum selection fraction of 0.1, i.e., about 10 selected clients per round. There is no study varying the client count, no measurement of server-side aggregation cost, and no wall-clock time or communication-volume comparison. Please add a scaling experiment or weaken the claim accordingly.","section":"Section 3 and Section 4"},{"comment":"Algorithm 1 is not reproducible as written. The termination condition 'until |clients| > nand count1 > 2' appears to be missing the operator 'and' (or is malformed), and line 7 contains the unparsed expression 'model size(Pd t=0 min(mspare c,t , rp,t δc ), c)' with undefined notation. The algorithm also refers to C′ without defining it. These issues make it impossible to reconstruct the client-selection procedure from the paper alone and should be corrected before publication.","section":"Algorithm 1"}],"minor_comments":[{"comment":"There are language errors: 'Despite of' should be 'Despite', and 'integratged' should be 'integrated'.","section":"Abstract and Section 1"},{"comment":"The cumulative energy for CAMA (BN True) at round 15 in Table 2 is 1.9226 kWh, while Table 3 reports a total energy usage of 1.85 kWh for the same configuration; please reconcile these numbers and state whether Table 2 is a single run or an average across iterations.","section":"Table 2 and Table 3"},{"comment":"The selection probability P(c) and the statistical utility σc use p(c) inconsistently; p(c) is never defined, and the condition p(c) ≥ 1 in Eq. (2) is unclear.","section":"Equations (1) and (2)"},{"comment":"Figure 1 is referenced before it is formally introduced, and its caption does not explain the meaning of the power-domain and client icons; consider labeling the panels.","section":"Figure 1"},{"comment":"The table entry 'labels per user 2(for balanced non-iid [1])' is awkwardly formatted; it should read '2 (balanced non-IID, following [1])'.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a credible engineering extension of FedZero, but the energy-efficiency claim is currently computed rather than measured, and the fairness and scalability claims are asserted without evidence. I would like the authors to either add direct power measurements or clearly present Eq. (3) as a simplifying assumption with a sensitivity analysis. The accuracy comparison appears to be the strongest part of the paper and should be retained as the main empirical contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CAMA is a sensible combination of FedZero's carbon-aware client selection and HeteroFL's ordered dropout, plus a new capacity-based method for assigning model sizes (Algorithm 2). It ships code and runs real training on MNIST and CIFAR-10, so the accuracy comparison against FedZero is a measured result. The convergence claim has some support: with BN enabled, CAMA reaches about 69.6% final accuracy on CIFAR-10 versus FedZero's 67.0%, and similar on MNIST. The new allocation algorithm is a legitimate, modest extension rather than a new paradigm. Credit is also due for releasing the code and comparing against a strong baseline.\n\nThe soft spot is the energy claim, which is the paper's headline. Equation 3 postulates Ec,i = ep × bc × mr, a linear scaling of per-batch energy with model rate. But the paper defines model rates through HeteroFL's hidden channel shrinkage ratio of 0.5, which means a client at mr=0.5 halves both input and output channels. For compute-bound convolutional layers, FLOPs (and therefore energy) scale closer to mr² than mr. Add fixed per-batch overheads like data loading, kernel launches, and communication, and the reported savings of 0.31 kWh over FedZero by round 15 become brittle. The authors never measure real power; every energy number in Table 2 and Figure 3 is arithmetic from Eq. 3. If the true scaling is sublinear, the savings could shrink or even reverse for the smallest model rates. This is load-bearing for the title's promise. Also minor but real: Algorithm 2's model rate sequence has a typo — it should end at 0.0625, not 0.625. The fairness claim is asserted but never quantified; no participation variance or Gini-style metric appears. Table 2 lacks error bars despite the text saying results were averaged over five iterations.\n\nThis paper is for people working on green or energy-aware federated learning systems. The combination is plausible and worth discussing, but the empirical evidence is not yet convincing. It deserves a serious referee — not a desk reject — because the system is genuinely new, buildable, and the accuracy results are real. A referee should push for hardware measurements or at least a validated per-device energy model, and for an explicit fairness metric. If I were the editor, I'd send it to review with expectation of major revision.","headline":"CAMA is a plausible new combination of FedZero and HeteroFL with a capacity-based allocation algorithm, but its headline energy savings rest on an unvalidated linear energy model while the accuracy results are genuinely measured.","tokens_in":6880,"tokens_out":2900,"would_cite":false,"duration_ms":26130,"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 federated learning framework that sizes each client's model to its available renewable energy can converge faster and use less energy than fixed-size green scheduling.","keywords":["federated learning","carbon awareness","dynamic model size","ordered dropout","renewable energy","client selection","energy efficiency","non-IID data"],"falsifier":"Take a GPU class such as the one the paper calls 'small' hardware, train the same model at model rates 1, 0.5, 0.25, and 0.125 on the same batch count, and measure wall-power draw over the run. If measured energy per batch does not scale linearly with $m_r$, the central energy-efficiency result fails.","tokens_in":5918,"feed_emoji":"🌱","tokens_out":8212,"duration_ms":69866,"temperature":0.7,"pith_summary":"The paper introduces CAMA, a carbon-aware federated learning strategy that decides how much of the global model each client trains based on how much excess renewable energy and spare computing capacity that client currently has. Its central claim is that replacing the all-or-nothing green-energy scheduling of the baseline it extends with a dynamic model-size allocation produces both faster convergence and lower cumulative energy use. The paper supports this with 15-round experiments on CIFAR-10 and MNIST using real solar-forecast and hardware-load traces, where the dynamic-size runs report higher final accuracy and lower total energy than the full-model-only baseline while also spreading participation more evenly across clients. If the claim holds, federated learning can be scheduled to follow renewable supply instead of competing with the grid.","feed_headline":"Federated learning trains on excess solar by resizing models per client","feed_subtitle":"CAMA adapts each device's model size to available renewable energy, reporting higher accuracy at lower energy use than a fixed-size…","key_machinery":"The load-bearing mechanism is dynamic model-size allocation driven by a batch-capacity estimate. The allocation algorithm starts at model rate $m_r = 1$ and repeatedly halves it down to a default of $\\mu = 0.0625$ until the number of batches the client can execute with its available energy and spare compute covers the required work; even a client that cannot handle the smallest rate is still given the default tiny model. Ordered dropout makes the submodels nested subsets of the full network, so a layer of size $d_g \\times k_g$ is reduced to $(d_g m) \\times (k_g m)$, and the server aggregates present elements by a weighted average over the local models. The energy accounting is Equation 3, $E_{c,i} = e_p \\times b_c \\times m_r$, which scales a client's per-batch energy linearly by the model rate.","core_discovery":"CAMA's central discovery is that a client's contribution to federated learning should be scaled to its energy budget rather than gating participation on being able to run the full model. In each round the server excludes power domains with no surplus energy and clients that have over-participated, estimates how many batches each remaining client can afford, and assigns the largest model rate from $\\{1, 0.5, 0.25, 0.125, 0.0625\\}$ that the client's batch capacity supports, with a default tiny model for clients below even the smallest threshold. The server then partitions a global layer of dimensions $d_g \\times k_g$ into a sublayer of $(d_g m) \\times (k_g m)$ for a client with model rate $m$, and aggregates the returned submodels by weighted averaging over the elements each submodel contains. On a Dirichlet-split CIFAR-10 task with ResNet-18, the paper reports a final accuracy of 69.6% at 1.85 kWh after 15 rounds, versus 67.0% at 2.16 kWh for the full-model-only baseline; on MNIST it reports 93.4% at 2.26 kWh versus 91.6% at 2.59 kWh.","pith_inferences":["Beyond the paper, the linear energy model in Equation 3 is an assumption, not a measurement; an obvious extension is to verify on real GPUs whether a quarter-size model really costs one quarter of the energy per batch, since fixed overheads would change the reported kWh savings.","Because the paper describes static batch normalization and cumulatively updated global statistics for privacy, a natural follow-up is to combine the dynamic-size aggregation with secure aggregation or differential privacy to protect those statistics.","The same ordered-dropout allocation could be applied to other resource dimensions, such as memory or network bandwidth, suggesting a general resource-aware model-pruning view of federated training."],"forward_implications":["Federated learning can include devices whose renewable-energy budget is too small for a full model, instead of dropping them from training.","Communication cost falls for small-model clients, because only the submodel is transmitted in each round.","Training schedules can be tied to solar and wind forecasts, shifting work to moments when surplus green energy exists.","Selection probabilities that penalize recent participation and large-model participation produce more even client involvement over rounds.","The reported accuracy gains suggest that model-size heterogeneity itself can help under non-IID label distributions, not only under energy constraints."],"supporting_citations":[{"why":"Supplies the ordered-dropout model partitioning and the aggregation rule for varying model sizes.","marker":"[1]"},{"why":"Defines the baseline method and the power-domain, client-registration, and solar-forecast setup that CAMA extends.","marker":"[3]"},{"why":"Supplies the statistical utility function used to score and filter clients for selection.","marker":"[6]"},{"why":"Provides the ResNet-18 architecture used in the CIFAR-10 experiments.","marker":"[7]"},{"why":"Provides the real solar and solar-forecast data used to model power domains.","marker":"[8]"},{"why":"Provides the dropout mechanism that ordered dropout adapts for nested submodel training.","marker":"[5]"}],"fun_headline_variants":["Resize models to fit solar power, boost FL efficiency","Carbon-aware FL: size models to energy budget","Dynamic model sizes cut FL energy and carbon","FL adapts model size to renewable energy supply","Energy-aware federated learning: train when sun shines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The energy savings are computed with Equation 3, which assumes a client's energy per batch falls in direct proportion to the fraction of the model it trains; if the real relationship has fixed overheads, the reported kWh reductions could be wrong.","fun_headline_variants_meta":{"raw":{"variants":["Resize models to fit solar power, boost FL efficiency","Carbon-aware FL: size models to energy budget","Dynamic model sizes cut FL energy and carbon","FL adapts model size to renewable energy supply","Energy-aware federated learning: train when sun shines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000465,"raw_usage":{"total_tokens":2332,"prompt_tokens":964,"completion_tokens":1368,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1305}},"tokens_in":580,"tokens_out":1368,"duration_ms":9460,"temperature":1.0,"reasoning_tokens":1305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:14:04.549402+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a GPU class such as the one the paper calls 'small' hardware, train the same model at model rates 1, 0.5, 0.25, and 0.125 on the same batch count, and measure wall-power draw over the run. If measured energy per batch does not scale linearly with $m_r$, the central energy-efficiency result fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ordered-dropout model partitioning and the aggregation rule for varying model sizes."},{"cited_title":"Wiesner, R","cited_arxiv_id":null,"evidence_quote":"Defines the baseline method and the power-domain, client-registration, and solar-forecast setup that CAMA extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the statistical utility function used to score and filter clients for selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the real solar and solar-forecast data used to model power domains."},{"cited_title":"Srivastava, G","cited_arxiv_id":null,"evidence_quote":"Provides the dropout mechanism that ordered dropout adapts for nested submodel training."}],"review_version":1}