{"id":"739fe17e-b864-4137-b38d-5ffb86ba056d","arxiv_id":"2505.12523","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"A survey of energy-aware deep learning methods for resource-constrained devices, covering energy-aware design, adaptive inference, on-device training, and scheduling on energy-harvesting systems.","lead":"Deep learning on small battery-powered devices faces strict energy limits, and this paper reviews a growing set of techniques for designing, running, and training networks with energy as the main goal. It organizes dozens of existing methods into categories and highlights open problems, especially accurate energy prediction without running a model.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Survey's motivating SqueezeNet-MAC claim is misattributed, and the §4 NEq equation is uninterpretable, exposing the load-bearing assumption that summaries of cited works are accurate.","rationale":"The reader's weakest_assumption is exactly the load-bearing one for a survey: the accuracy of its summaries of cited works. My own pass through the manuscript exposes the same concrete example (SqueezeNet MACs vs. parameters) and the same uninterpretable NEq equation, plus the HarvSched/HarvNet inconsistency. These are not mere typographical nitpicks; they affect the paper's central motivating example that MACs are not a good energy proxy. If SqueezeNet actually has comparable or fewer MACs than AlexNet, the statement 'exhibits greater energy consumption' is the key evidence that energy is not proportional to MACs; if that evidence is misattributed, the illustrative point still might be true (data movement is indeed often dominant, as also supported by the GoogLeNet/feature-map references), but the survey's reliability is undercut. The NEq equation, similarly, is presented as a formal definition and is internally impossible. Since the survey's value lies in its map of the field, these errors make a conditional recommendation appropriate: the high-level taxonomy and future-work directions are plausible and useful, but the paper should be accepted only after the authors verify every quantitative claim in Table 1/Table 2 and the surrounding text, and after they fix the named inconsistencies. I did not find evidence of any deeper central-claim failure (no circular derivation, no hidden fundamental flaw), so I do not recommend rejection. The agreement with the reader is full.","tokens_in":28428,"tokens_out":1901,"duration_ms":17738,"concrete_test":"Pull the SqueezeNet paper (arXiv:1602.07360) and the two cited energy references ([103] Hot Chips 2016 talk and [173] Yang et al. 2016) and verify the exact MAC and energy comparison. Specifically, recompute SqueezeNet's MAC count from its published architecture (1.7G for SqueezeNet with skip connections, ~3.7G for the 1.1 version without) against AlexNet's ~666M MACs, and check whether [103] or [173] state that SqueezeNet uses more energy than AlexNet. If the MAC comparison is unsupported or reversed, the paper must correct the claim and soften the 'proxies fail' opening.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper is a survey, so its central claim is that the reviewed literature can be accurately summarized and organized into a useful map. The reader identified the weakest assumption: the fidelity of those summaries. In §2, the paper states that 'SqueezeNet [51] contains 50x fewer MACs than AlexNet [68], yet exhibits greater energy consumption on various platforms [103, 173].' The cited SqueezeNet reference (arXiv:1602.07360) explicitly reports '50x fewer parameters and <1MB model size'—not 50x fewer MACs. This is a specific, load-bearing misattribution because it is used as the motivating example that MAC proxies do not predict energy, and it could be wrong in either direction: SqueezeNet has roughly 1/50th of AlexNet's parameters, but its MAC count is only around 1/48th or 1/70th depending on the version, and it is not a universally established result that SqueezeNet has greater energy consumption than AlexNet on 'various platforms.' The supporting citations [103, 173] do not clearly state this comparison. The NEq equation in §4 is another concrete fidelity failure: the reported equilibrium condition '|v_t_i| < epsilon, epsilon <= 0' is impossible for a non-negative magnitude, and the undefined beta/epsilon makes the description unusable. These are not stylistic slips: they are the exact evidence that the survey's summaries of its own citations cannot be trusted without checking. The paper provides no independent verification mechanism, and the omission of a clear statement of the HarvSched/HarvNet naming and reference inconsistency (§3.2.2 and §6) compounds the issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This survey reviews energy-aware deep learning techniques for resource-constrained IoT and mobile hardware. It organizes the field into energy-aware DNN design (pruning, quantization, neural architecture search), energy-adaptive inference (right-sizing, multi-exit networks, offloading), on-device training, and applications such as energy-harvesting systems and federated learning. The paper argues that energy consumption is not captured by MAC/FLOP proxies and that accurate, hardware-agnostic energy estimation remains an open problem, concluding with a set of future research directions and tables summarizing existing methods, estimation approaches, and embedded ML frameworks.","tokens_in":28728,"tokens_out":6276,"duration_ms":61075,"significance":"If the survey's summaries are accurate, it provides a valuable synthesis of a fragmented literature and a useful starting point for researchers entering this area. The organizational scheme, the coverage of intermittent computing and on-device training, and the tables of NAS methods, energy-estimation approaches, and MCU frameworks are concrete contributions. However, the survey's value as a reference depends on the fidelity of its descriptions of primary sources; the misattributed SqueezeNet example in §2 and the uninterpretable NEq equation in §4 undermine confidence in the cited summaries and must be corrected. The paper also demonstrates areas that are rarely surveyed together, such as energy-harvesting and federated learning, which gives it a distinct niche among existing reviews.","major_comments":[{"comment":"The motivating example for why MAC proxies fail to predict energy is misattributed: the paper states that 'SqueezeNet [51] contains 50x fewer MACs than AlexNet [68], yet exhibits greater energy consumption on various platforms [103, 173].' Reference [51] (Iandola et al., SqueezeNet) reports 50x fewer parameters and a model size below 1MB, not 50x fewer MACs. The cited references do not clearly establish the comparative energy-consumption claim on 'various platforms.' Because this example is the central evidence for the paper's thesis that MAC counts are not an energy proxy, the authors should replace it with a documented MAC/energy comparison or rephrase the claim to match the sources.","section":"§2, first paragraph"},{"comment":"The equilibrium condition for NEq is uninterpretable as written: the inequality '|v_t_i| < epsilon, epsilon <= 0' is impossible for the non-negative magnitude |v_t_i|. The following sentence mentions beta ('β is an arbitrary threshold'), but β is not defined in Eq. (2) or the surrounding text, and the text also refers to 'epsilon/beta' as if the two are interchangeable. Readers cannot determine NEq's actual selection rule from this description. The authors should reproduce NEq's exact condition with all symbols defined, or remove the equation and explain the method in words.","section":"§4, Eq. (2)"}],"minor_comments":[{"comment":"The orphaned footnote '0https://www.st.com/resource/en/datasheet/stm32l4r5zi.pdf' at the bottom of page 2 should be removed or converted into a proper citation in the text.","section":"§2, footnote"},{"comment":"The YOLOv5 reference contains the unfinished placeholder 'Accessed: insert date here.' and needs to be completed before publication.","section":"Reference [150]"},{"comment":"Several distinct reference numbers point to the same work (e.g., [125] and [126]; [173] and [174]; [107] and [108]; [143] and [144]; [55] and [56]; [175] and [176]; [49] and [50]; [39] and [67]). Deduplicate these entries and renumber consistently so that readers can trace claims to unique sources.","section":"Reference list"},{"comment":"The weight α in the class-dependent threshold equation is not defined in the text; please state its meaning, typical range, and whether it is a hyperparameter set by the user.","section":"§3.3, Eq. (1)"},{"comment":"The sentence introducing β appears to be a leftover from a different version of the manuscript, since Eq. (2) uses only ε. Define β and explain how it relates to ε, or remove the mention.","section":"§4, text after Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and addresses a timely topic. The organizational framework is sound, but the accuracy of several load-bearing summaries needs to be verified and corrected; I view these issues as fixable within the manuscript's scope. The self-citation [188] is used only as an example application and does not raise a conflict concern. The many duplicated references and unfinished placeholders suggest a rushed final editorial pass; the revision should include a careful check of both the reference list and the quantitative claims against their sources."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful survey of a young subfield, but it ships with a few concrete factual slips—including a misattributed SqueezeNet claim that sits right at the front door. Worth a serious referee, but only after the authors fix the errors.\n\nWhat the paper does well: it collects a broad set of recent works on energy-aware DL for constrained devices, organizes them into a reasonable taxonomy (design-time, adaptive inference, on-device training, harvesting, federated learning), and gives an honest assessment of what is missing. The emphasis on intermittent/energy-harvesting devices as a first-class topic is the most distinctive part. The gap analysis in §7—cross-platform, execution-free energy estimation—is the right problem to name. The tables are handy for orientation.\n\nThe soft spots are real but localized. The §2 motivating example says SqueezeNet has 50x fewer MACs than AlexNet yet consumes more energy; the cited SqueezeNet paper actually claims 50x fewer parameters, and the two supporting refs [103, 173] don't clearly establish the energy comparison. Since this is the example the survey uses to argue that MAC proxies are insufficient, it needs to be correct or reworded. The NEq description in §4 is another concrete failure: the equilibrium condition is printed as |v_t_i| < ε with ε ≤ 0, and then β is mentioned as the threshold. That's uninterpretable as written. There are also duplicated references (e.g., [107]/[108], [173]/[174], [49]/[50]) and the HarvSched/HarvNet naming inconsistency.\n\nThese are not deep flaws. The survey's central argument still holds on the strength of other cited work (e.g., Yang et al., Eyeriss). The errors are the kind that careful copyediting catches. But as-is, they undermine the paper's reliability as a reference.\n\nWho is this for? Someone looking to get oriented in energy-aware tinyML, especially on the harvesting side, will find the map useful. An expert might not learn much, though the gap list is a good prompt for research. I'd bring it to a reading group as a critical-reading exercise, but I'd pair it with the original papers before trusting the specifics.\n\nFor peer review: yes, send it out. The scope is timely, the coverage is broad, and the structural contribution is real. Ask the authors to fix the factual claims, check every equation against the source, deduplicate references, and clarify the HarvNet/HarvSched naming. After that, it's a solid survey reference.","headline":"Useful survey of energy-aware DL on constrained devices, but the motivating SqueezeNet claim and the NEq equation need fixing before this is reliable.","tokens_in":29255,"tokens_out":4527,"would_cite":false,"duration_ms":43695,"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":"This survey argues that energy consumption is a distinct optimization axis for deep learning on constrained hardware, one that MAC and FLOP counts do not capture, and that the field still lacks accurate hardware-agnostic energy estimation.","keywords":["energy-aware deep learning","TinyML","on-device inference","energy estimation","neural architecture search","multi-exit networks","energy harvesting","federated learning"],"falsifier":"Measure the end-to-end energy of SqueezeNet and AlexNet on the same microcontroller while controlling for batch and input resolution: the survey's motivating claim predicts SqueezeNet can consume more energy despite having far fewer multiply-accumulate operations, so a dataset of platforms where energy tracks MAC count monotonically would falsify the proxy-failure premise.","tokens_in":28194,"feed_emoji":"⚡","tokens_out":5406,"duration_ms":54775,"temperature":0.7,"pith_summary":"This survey argues that energy use is a first-class constraint for deep learning on IoT and mobile hardware, separate from memory and compute budgets. Its central case is that conventional efficiency proxies—multiply-accumulate (MAC) and FLOP counts—do not track energy consumption, because data movement, not arithmetic, dominates the energy bill and because memory hierarchies differ across devices. The paper organizes the field into energy-aware architecture design, adaptive inference (right-sizing, early exits, offloading), on-device training, and energy-harvesting or federated deployment, and it identifies accurate hardware-agnostic energy estimation as the unsolved problem on which the whole area depends.","feed_headline":"Energy, not MAC counts, decides how long tiny AI lasts","feed_subtitle":"A survey maps energy-aware deep learning for battery-free and intermittent devices, and names the missing estimator.","key_machinery":"The load-bearing mechanism is the data-movement cost model: fetching data from memory far from the compute unit can cost $10$ to $100+$ times an arithmetic operation, and feature-map movement rather than computation is what dominates DNN energy. This model is what turns the proxy-failure observation into a design principle: since energy depends on where data lives and how it flows, any energy-aware method must estimate or measure per-layer, per-platform movement costs rather than counting operations.","core_discovery":"The authors claim that energy-aware deep learning is a distinct optimization axis that today's MAC- or FLOP-based proxies systematically miss, and they support this with evidence that a network with far fewer operations can consume more energy on a given platform, and that only about 10% of a typical CNN's energy goes to computation while the rest goes to moving feature maps. They therefore classify existing work by how it makes energy itself the objective: pruning and quantization guided by layer-wise energy estimates, neural architecture search that predicts energy from measurements or regressions, inference policies that trade accuracy for energy at runtime, and training or fine-tuning methods that limit which parameters are updated. The survey concludes that the field's binding constraint is the lack of a universal, execution-free way to estimate a DNN's energy on arbitrary hardware, and proposes architecture representations such as abstract syntax trees as a route toward hardware-agnostic estimation.","pith_inferences":["If cross-platform energy estimation is ever solved, compiler-style cost models for DNNs could make energy-awareness an automated part of the build, much as latency is today.","The data-movement emphasis predicts that the best compression recipe for one memory hierarchy will not transfer to another, so per-device calibration may be unavoidable even with a hardware-agnostic core model.","Energy-aware neural architecture search with uncertainty-aware energy prediction could avoid overfitting to the few benchmarked devices and make search practical for microcontroller-scale deployment.","Intermittent energy-harvesting machine learning, if made reliable, would let remote sensors run vision or audio classifiers for years without batteries, changing where on-device AI is economically sensible."],"forward_implications":["Energy-aware pruning and quantization must be evaluated on measured energy, not on MAC or parameter counts, or they may silently increase consumption.","Adaptive inference—early exits, input-dependent quantization, and offloading—becomes the main lever for staying inside a fluctuating energy budget on battery-free devices.","On-device fine-tuning for data shifts will need parameter-selection or rehearsal methods, since full backpropagation is too energy-expensive on microcontrollers.","Federated learning over heterogeneous, intermittently powered devices needs energy-aware participation policies, or the global model becomes biased toward well-powered devices.","Progress on any of these fronts is gated by the same missing capability: a cheap, accurate, execution-free energy estimator that works across hardware."],"supporting_citations":[{"why":"Supplies the per-operation energy model showing that data movement, not arithmetic, dominates CNN energy.","marker":"[14]"},{"why":"The compressed network whose energy-versus-MAC discrepancy motivates the proxy-failure argument.","marker":"[51]"},{"why":"Reported platform measurements cited as evidence that a network with far fewer MACs can consume more energy.","marker":"[103]"},{"why":"Maps the current benchmarking landscape and its limited coverage of microcontrollers.","marker":"[148]"},{"why":"Current state-of-the-art regression-based energy predictor, the baseline for generalization claims.","marker":"[149]"},{"why":"Introduces energy-aware pruning by ranking layers on measured energy, the basis for the energy-aware pruning category.","marker":"[173]"},{"why":"Provides the method and breakdown (roughly 10% compute, rest data movement) that underpins the energy model.","marker":"[175]"},{"why":"A layer-level energy profiling method for edge CPUs, one of the few estimators built for constrained devices.","marker":"[117]"}],"fun_headline_variants":["Tiny AI's true cost: energy, not MACs","Energy-aware DL: the missing estimator","Counting watts, not MACs, for IoT deep learning","For battery-free AI, energy is the real metric","Deep learning's hidden energy cost exposed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's taxonomy and recommendations assume that its summaries of the cited works are accurate, including quantitative details such as SqueezeNet's '50x fewer MACs' comparison and NEq's equilibrium condition; the paper itself contains at least one misstatement (SqueezeNet's actual cited result is 50x fewer parameters, not MACs) and a misprinted NEq inequality.","fun_headline_variants_meta":{"raw":{"variants":["Tiny AI's true cost: energy, not MACs","Energy-aware DL: the missing estimator","Counting watts, not MACs, for IoT deep learning","For battery-free AI, energy is the real metric","Deep learning's hidden energy cost exposed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000198,"raw_usage":{"total_tokens":1312,"prompt_tokens":833,"completion_tokens":479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":406}},"tokens_in":449,"tokens_out":479,"duration_ms":4851,"temperature":1.0,"reasoning_tokens":406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:31:57.431851+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the end-to-end energy of SqueezeNet and AlexNet on the same microcontroller while controlling for batch and input resolution: the survey's motivating claim predicts SqueezeNet can consume more energy despite having far fewer multiply-accumulate operations, so a dataset of platforms where energy tracks MAC count monotonically would falsify the proxy-failure premise.","supporting_citations":[{"cited_title":"Fine-Grained Energy and Performance Profiling framework for Deep Convolutional Neural Networks","cited_arxiv_id":"1803.11151","evidence_quote":"A layer-level energy profiling method for edge CPUs, one of the few estimators built for constrained devices."}],"review_version":1}