{"id":"7b08b8f5-8f4e-40ad-9ec9-b28c4f754a1b","arxiv_id":"2505.17626","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Stochastic-depth-trained ResNets can skip many residual blocks at inference with small accuracy loss, and sensitivity-based selection of which blocks to skip enables zero-overhead adaptive inference.","lead":"This paper shows that ResNets trained with stochastic depth tolerate skipping many internal blocks at inference time with little accuracy loss. It uses that property to build a configurable adaptive inference system that achieved up to 2x power efficiency on an edge board during tests.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'near Pareto-optimal' claim rests on an unvalidated additivity assumption in the sensitivity-based configuration selection, and the same test split is reused for selection and reporting; the headline accuracy/efficiency numbers are therefore not yet established.","rationale":"I read the paper as making an empirical systems claim: stochastic-depth training makes ResNet blocks individually more skip-tolerant, and a design-time sensitivity analysis can identify a Pareto front of skip configurations that a runtime can switch among. The first part is qualitatively supported by Figure 1's shaded distributions over 500 random configurations. The second part is where the headline numbers stand or fall, and it rests on an additivity assumption that the paper does not validate. The selection rule skips the N individually least-sensitive blocks; this is only near-optimal if the importance of skipping a block is roughly independent of which other blocks are skipped. The 49.61% gap between best and worst 37-block configurations in Figure 1 shows that block identity has a large effect, but it does not by itself prove additivity fails; the point is that the paper provides no evidence for additivity. The paper even acknowledges the configurations are 'not guaranteed to deliver Pareto operating points,' yet still labels them 'near Pareto-optimal.' This is a correctness risk rather than an internal contradiction. The same test split being used for sensitivity ranking, Pareto filtering, and accuracy reporting makes the reported accuracy numbers optimistic in a way that is hard to quantify without a held-out split. A brute-force check on ResNet-20 would settle both questions. I therefore agree with the reader's weakest assumption, and I do not see a reason to move the verdict: the paper should remain conditional because the claim is plausible, testable, and the framework is modular, but the quantitative trade-off claims need stronger validation before they can be accepted as stated.","tokens_in":10894,"tokens_out":5805,"duration_ms":47289,"concrete_test":"On ResNet-20, whose B_s is 7, the paper reports only 7 sensitivity evaluations, so exhaustive enumeration is feasible: split the dataset into a selection split and a held-out evaluation split, enumerate all 2^7 = 128 skip configurations, measure accuracy and inference time on the selection split, compute the true Pareto front, and compare it with the paper's sensitivity-based front. Then report accuracy of both fronts on the untouched evaluation split. If the sensitivity front is dominated by the true front (e.g., a configuration with equal or fewer FLOPs has more than 1% higher accuracy) or its held-out accuracy is materially lower than the selection-split accuracy, the near-Pareto/additivity claim fails. For ResNet-110, where enumeration is infeasible, a random or evolutionary search over about 1,000 configurations under the same protocol would provide a practical lower-bound check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's quantitative headline depends on the Pareto-front step (Sections II-A3 and II-A4) selecting configurations whose accuracy is near the best achievable for a given number of skipped blocks. The selection rule skips the N blocks with the lowest individual sensitivity. This is only valid if block importance is approximately additive. The paper itself admits in II-A4 that the sensitivity list 'regards blocks individually' and that evaluated configurations are 'not guaranteed to deliver Pareto operating points,' yet the abstract and contributions call them 'near Pareto-optimal' without testing additivity. Figure 1 shows a 49.61% accuracy gap between the best and worst configurations when 37 of 54 blocks are skipped in a stochastic-depth-trained ResNet-110, so block identity matters enormously, and individual sensitivity rankings cannot detect whether the best N-block set is the union of the N best individual blocks. If interactions are non-additive, the N-least-sensitive greedy set can be much worse than alternative N-block configurations, making the reported 0.71% drop and 2x power-efficiency numbers optimistic. A second confound compounds this: the same CIFAR test split is used to rank sensitivities, Pareto-filter configurations, and report accuracy, so the reported numbers include selection bias and there is no held-out estimate. The footnote in II-A3 asserting that sensitivity 'performed better than' l2-norm, fisher, hessian, and random search is not accompanied by data, and all experiments appear to be single-run with no error bars. Thus the central resilience observation is plausible, but the quantitative trade-off claims are not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for adaptive inference in residual networks: models are trained once with Stochastic Depth, then a design-time sensitivity analysis ranks residual blocks by individual accuracy impact, and the resulting one-per-depth skip configurations are filtered into an approximate Pareto front of accuracy versus inference time. At runtime, an adaptive scheduler switches among these configurations, without weight reload, depending on device load and idle time. The paper reports on CIFAR-10/100 with ResNet-20/110 deployed on an Odroid XU4 board, claiming up to 2x power efficiency and up to 1.97x processed inferences relative to original ResNets, at small accuracy drops.","tokens_in":11101,"tokens_out":3981,"duration_ms":34226,"significance":"The central observation—that Stochastic Depth training confers resilience to inference-time layer skipping—is well motivated and, if it holds, practically valuable: it enables a single model to serve multiple accuracy/efficiency operating points with deterministic per-input runtime and no additional gate parameters. The paper also contributes an actual deployment on embedded hardware with energy measurements, which is a strength. However, the quantitative headline claims rest on an unvalidated additivity assumption in the Pareto-front construction and on test-set reuse in selection and evaluation; these issues must be addressed before the reported numbers can be taken as reliable.","major_comments":[{"comment":"The Pareto-front generation assumes that block importance is additive: configurations are constructed by skipping the N least-sensitive blocks according to individual sensitivity, and only these one-per-N configurations are evaluated. The paper itself states that the sensitivity list 'regards blocks individually' and that evaluated configurations are 'not guaranteed to deliver Pareto operating points,' yet the abstract and contributions call the resulting points 'near Pareto-optimal.' This is load-bearing because the headline accuracy drops (e.g., 0.71% and 3.13% in Table I) come from these configurations. Since Figure 1 shows a 49.61% accuracy gap among configurations that skip the same number of blocks in ResNet-110, interactions clearly matter; the greedy construction could be arbitrarily worse than the true Pareto front. A concrete fix is to validate the additivity assumption on ResNet-20, where the full skip-configuration space is only 2^6 = 64 configurations and exhaustive evaluation is trivial, and to report a comparison between the greedy front and the exhaustive front. For ResNet-110, a random or beam-search sample of configurations at matched skip counts would provide a meaningful sanity check.","section":"II-A4"},{"comment":"The same CIFAR test split is used for three dependent purposes: ranking block sensitivity, Pareto-filtering the skip configurations, and reporting the final accuracy numbers in Table I and Figure 6. This is selection on the test set, so the reported accuracy drops and efficiency gains are optimistically biased. The paper should hold out a validation split for sensitivity ranking and Pareto selection, and report only the held-out test accuracy, or use repeated cross-validation. Without this, the quantitative claims—especially the 0.71% drop and 2x power-efficiency figure—are not established for unseen data.","section":"II-A3/II-A4, Table I"},{"comment":"The footnote states that sensitivity ranking 'performed better than' l2-norm, fisher, hessian, and random search, but no protocol, comparison tables, or error bars are provided. Since the sensitivity ranking is the core mechanism that turns the combinatorial skip-design space into a tractable Pareto front, this comparison should be documented, at least for the small ResNet-20 model where exhaustive evaluation is feasible. Without these details, the choice of sensitivity analysis is not independently verifiable.","section":"II-A3 (footnote 1)"}],"minor_comments":[{"comment":"The phrase 'improvements of up to 2X in power efficiency at accuracy drops as low as 0.71%' should be clarified: in Table I, the 2.00x power-efficiency gain occurs at a 3.13% accuracy drop (ResNet-110/CIFAR-10), while the 0.71% drop case (ResNet-110/CIFAR-100) yields only 1.26x. If these are separate extremes, the text should say so explicitly to avoid implying simultaneous attainment.","section":"Abstract / IV-B"},{"comment":"There is a typo: 'untracktable' should be 'intractable'.","section":"II-A3"},{"comment":"The final sentence of the runtime evaluation says 'original ResNet-20 and ResNet-100 models' and repeats the phrase; the model is ResNet-110, and the duplicated clause should be removed.","section":"IV-B"},{"comment":"The figure caption notes that the Original and SkipNet curves 'do not follow the number of skipped blocks in the x-axis,' which is visually confusing because the x-axis is shared. Plotting these baselines as horizontal reference lines or with separate markers would improve readability.","section":"Figure 6"},{"comment":"The 'zero-overhead' claim should be qualified: the gated model still reads a skip-configuration array and evaluates a branch per block, and while this overhead is tiny and not input-dependent, calling it literally zero may invite unnecessary criticism.","section":"II-A2"}],"recommendation":"major_revision","confidential_remarks":"The central resilience observation is plausible and well supported by Figures 1 and 6, but the quantitative contribution currently hinges on an unvalidated greedy-additivity assumption and on test-set reuse. I would like to see, at minimum, the exhaustive validation on ResNet-20 and a clean train/validation/test split. The paper does not mention code or artifact release; for a systems-oriented ML paper, releasing the modified IREE flow would substantially increase reproducibility. The contribution is within scope for a venue interested in efficient inference, but the evaluation protocol needs strengthening."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real systems paper with a plausible central observation and one unvalidated shortcut that the headline numbers lean on. The abstract's 'up to 2x power efficiency at accuracy drops as low as 0.71%' should be read as conditional until the selection bias is addressed.\n\nWhat is actually new: the observation that stochastic-depth-trained ResNets tolerate arbitrary block skipping much better than standard ResNets is supported by Figure 1's mean and spread, and it is not just a restatement of stochastic depth. The weightless gate design and compile-time selection of skip configurations, with a runtime scheduler that switches between them, is a practical combination: no extra memory, no retraining, predictable execution time. The hardware evaluation on Odroid XU4 with real power measurements is more than many papers in this area do.\n\nSoft spots: the Pareto-front step (II-A4) constructs multi-block skip sets by skipping the N least-sensitive blocks. That is an additivity assumption, and the paper itself notes the sensitivity list treats blocks individually and does not guarantee Pareto points. The stress-test is right to push on this: Figure 1 shows block identity matters enormously when many blocks are skipped, so a greedy union of individually unimportant blocks can miss much better N-block sets. The headline 0.71% drop and 2x power numbers come from this step. Second, the same test split is used to rank sensitivities, filter configurations, and report final accuracy. That is selection bias, and without a held-out split the numbers are optimistic. Third, single runs, no error bars, and no released code or data. The footnote claiming sensitivity beats l2-norm, fisher, hessian, and random search has no data behind it.\n\nNone of this kills the central claim. The resilience observation is independent of the Pareto-front construction, and it is convincing. But the quantitative trade-off claims are not yet established. The paper would be much stronger with a validation set for selection, a check of additivity (e.g., compare greedy N-block sets against a few random or search-based alternatives of the same size), and repeated runs.\n\nBottom line: worth a serious referee, and likely publishable after revision. The audience is edge-AI and efficient-inference researchers. I'd bring it to reading group if someone cares about adaptive inference on commodity edge hardware. Cite it for the stochastic-depth resilience observation, not for the specific Pareto-front numbers.","headline":"Stochastic-depth skipping resilience is real and useful; the headline efficiency numbers need a validation split and an additivity check before I'd trust them.","tokens_in":11711,"tokens_out":2805,"would_cite":true,"duration_ms":21108,"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":"Stochastic-depth training makes ResNets resilient to layer skipping, enabling a single trained model to serve many accuracy/energy operating points at runtime.","keywords":["adaptive inference","layer skipping","stochastic depth","residual networks","edge computing","Pareto front","sensitivity analysis","runtime adaptation"],"falsifier":"Exhaustively evaluate every skip configuration on a small model: for a stochastically-trained ResNet-20 with 7 skippable blocks, compare the accuracy of the paper's sensitivity-ranked N-skip configurations against all other N-skip configurations across all 128 possibilities; if the sensitivity-ranked ones are not consistently at or near the top, the Pareto front is not doing the work the accuracy numbers claim. The same comparison on a conventionally trained ResNet-20 would directly test the claimed stochastic-depth resilience advantage.","tokens_in":10626,"feed_emoji":"⚡","tokens_out":8216,"duration_ms":62893,"temperature":0.7,"pith_summary":"Adaptive inference usually means per-input decisions: a network examines each sample and decides how many layers to run, at the cost of extra decision-gate parameters, complex training, and unpredictable execution time. This paper argues for a different, simpler approach: train a residual network once with Stochastic Depth, the training-time trick of randomly dropping residual blocks, and it becomes unusually tolerant of having blocks skipped wholesale at inference. A design-time sensitivity analysis ranks blocks by importance and builds a near-Pareto front of skip configurations trading accuracy against inference time, and a tiny runtime moves along that front by toggling a binary array, with no weight reloading and no extra computation. The payoff, measured on a real edge board, is up to 2x power efficiency and 1.97x processed inferences relative to a conventional ResNet, at accuracy drops as low as 0.71%.","feed_headline":"Stochastic-depth training makes ResNets skip-tolerant for edge AI","feed_subtitle":"One stochastically-trained model serves many accuracy/speed points; up to 2x power efficiency and 1.97x processed inferences.","key_machinery":"The load-bearing mechanism is the pairing of Stochastic Depth training with a sensitivity-ranked Pareto-front search. Stochastic Depth randomly drops residual blocks during training with a depth-dependent probability, which produces the skipping resilience measured in the paper's Figure 1. The design-time pipeline treats every residual block except the first in each segment as skippable, computes a sensitivity list by skipping one block at a time and measuring accuracy, then constructs configurations of the form \"skip the N least sensitive blocks\" for N from 0 to the number of skippable blocks. Each candidate is evaluated for accuracy and inference time on the full test set, and only the nondominated configurations form the approximate Pareto front that the runtime navigates. The gates themselves are weightless pass-throughs compiled through the paper's compiler pipeline, so a configuration is just a binary array passed alongside the input—no learned parameters, no retraining, no weight reloading.","core_discovery":"The paper's central claim is that Stochastic Depth training—originally a regularization and speedup trick for deep ResNets—implicitly makes the trained network resilient to arbitrary block-skipping at inference, and that this resilience can be turned into a zero-overhead adaptive-inference system. The authors insert weightless skip-or-not gates before residual blocks, rank each block by the accuracy lost when it alone is skipped, then generate candidate configurations by skipping the N least important blocks and keep only those that lie on the approximate Pareto front of accuracy versus inference time. At runtime, a simple controller increases skipping when the device is overloaded and decreases it after idle periods, following a standard convention of dropping requests that arrive at a busy device. On an edge platform with ResNet-20 and ResNet-110 on CIFAR-10 and CIFAR-100, the method reports up to 2.00x inference-per-watt and 1.97x processed inferences over the original ResNets, with the smallest accuracy drop being 0.71% (ResNet-110 on CIFAR-100); in all but one configuration it also beats the SkipNet baseline in power efficiency while avoiding SkipNet's per-input timing variability.","pith_inferences":["The sensitivity-ranking step is the most improvable component: if block-removal effects interact, greedy \"skip the N least important blocks\" configurations sit below the true Pareto frontier; exhaustively testing a small network's full skip space would show how much headroom remains.","The underlying principle—training-time stochastic sub-network dropout yields inference-time drop tolerance—may generalize beyond ResNets to any modular architecture, including vision transformers or large language models, where depth or expert modules could be dropped at serving time for latency scaling.","The runtime controller is reactive rather than predictive; coupling it with a short workload forecast or a learned policy for the idle threshold and minimum-accuracy parameter could smooth the accuracy/throughput trajectory and reduce dropped requests during bursts."],"forward_implications":["One stochastically-trained ResNet can be deployed as many different inference models: every Pareto configuration lives in the same executable and is activated by a binary array, so runtime switching costs no weight reloads and no extra memory.","Edge systems gain a predictable accuracy/energy knob: because skip counts are chosen at design time, each configuration has a fixed execution time, unlike per-input routing whose time can vary (the paper measures up to 13x variation for SkipNet).","The gains grow with model size: on ResNet-110, skipping 20 of 54 blocks drops CIFAR-10 accuracy by only 0.23% while skipping 37% of all blocks, and reported runtime gains reach 1.97x throughput and 2.00x inference-per-watt.","The approach makes adaptive inference compatible with real-time and quality-of-service constraints: the controller only sacrifices accuracy when requests would otherwise be dropped, and restores it after idle periods."],"supporting_citations":[{"why":"Supplies the Stochastic Depth training procedure whose skipping resilience is the paper's central observation.","marker":"[9]"},{"why":"Defines the residual block architecture that the skip configurations are built from.","marker":"[10]"},{"why":"The adaptive-inference baseline against which accuracy, speed, and efficiency are compared.","marker":"[1]"},{"why":"Provides the compilation pipeline that lets skip configurations be passed as an input array to the deployed model.","marker":"[20]"},{"why":"Defines the \"drop a request that arrives at a busy device\" semantics used by the runtime adaptation loop.","marker":"[21]"},{"why":"Supplies the image datasets used in all experiments.","marker":"[22]"}],"fun_headline_variants":["Stochastic depth training unlocks skip-tolerant ResNets","Zero-overhead adaptive inference via stochastic depth","One model, many speed points: stochastic depth for edge","ResNets trained with stochastic depth handle flexible skipping","Stochastic depth makes ResNets skip-ready for adaptive inference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hangs on the assumption that a block's importance measured by removing it alone predicts the best multi-block configurations: the Pareto front is built by skipping the N least-sensitive blocks, and if removing blocks together interacts non-additively, the reported accuracy and efficiency operating points are optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Stochastic depth training unlocks skip-tolerant ResNets","Zero-overhead adaptive inference via stochastic depth","One model, many speed points: stochastic depth for edge","ResNets trained with stochastic depth handle flexible skipping","Stochastic depth makes ResNets skip-ready for adaptive inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1871,"prompt_tokens":959,"completion_tokens":912,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":845}},"tokens_in":575,"tokens_out":912,"duration_ms":9447,"temperature":1.0,"reasoning_tokens":845,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:44:09.289928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Exhaustively evaluate every skip configuration on a small model: for a stochastically-trained ResNet-20 with 7 skippable blocks, compare the accuracy of the paper's sensitivity-ranked N-skip configurations against all other N-skip configurations across all 128 possibilities; if the sensitivity-ranked ones are not consistently at or near the top, the Pareto front is not doing the work the accuracy numbers claim. The same comparison on a conventionally trained ResNet-20 would directly test the claimed stochastic-depth resilience advantage.","supporting_citations":[{"cited_title":"Deep networks with stochastic depth,","cited_arxiv_id":null,"evidence_quote":"Supplies the Stochastic Depth training procedure whose skipping resilience is the paper's central observation."},{"cited_title":"Skipnet: Learning dynamic routing in convolutional networks,","cited_arxiv_id":null,"evidence_quote":"The adaptive-inference baseline against which accuracy, speed, and efficiency are compared."},{"cited_title":"An mlir-based compiler and runtime for ml models from multiple frameworks,","cited_arxiv_id":null,"evidence_quote":"Provides the compilation pipeline that lets skip configurations be passed as an input array to the deployed model."},{"cited_title":"Mlperf inference benchmark,","cited_arxiv_id":null,"evidence_quote":"Defines the \"drop a request that arrives at a busy device\" semantics used by the runtime adaptation loop."},{"cited_title":"Learning multiple layers of features from tiny images,","cited_arxiv_id":null,"evidence_quote":"Supplies the image datasets used in all experiments."}],"review_version":1}