{"id":"d87fc6c8-b9fd-4d40-ad4c-f2068cc62f2f","arxiv_id":"2507.00061","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Smooth-Distill applies EMA parameter averaging as a self-distillation teacher for multitask HAR and placement detection, and reports consistent but modest gains over multitask baselines.","lead":"This paper proposes Smooth-Distill, a training method that uses a smoothed historical copy of the model itself as a teacher to jointly recognize human activities and sensor placement from accelerometer data. A new sleep posture dataset with 12 postures and 3 wearing positions is released alongside experiments on MHealth and WISDM.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline performance claim rests on a possibly leaky 80:20 split; the paper never states that the split is subject-independent, and the per-class tables make subject leakage a plausible driver of the reported gains.","rationale":"The reader's weakest assumption is exactly this split question, and the full-text evidence supports it as the most load-bearing concern. The paper is otherwise internally coherent: the EMA update in Eq. 4 is correctly implemented as the Mean Teacher rule, the loss in Eq. 5 is a reasonable multitask distillation objective, and the efficiency comparison in Table 7 is plausible. The novelty concern (Mean Teacher not cited) is real but secondary; it affects framing, not correctness. The statistical-significance weakness is honestly reported in Table 8 and Section 6.2, which actually concedes that Smooth-Distill and Born-Again SD are statistically indistinguishable (p > 0.7) on the Sleep dataset, so the strongest defensible claim is efficiency at parity, not accuracy dominance. However, even that efficiency claim is conditional on the split and model-selection protocol being sound. A subject-level split is the standard required for HAR generalization claims and is conspicuously absent. The requested test is concrete and would settle the matter: re-run with held-out participants. If the results survive, the paper's conditional acceptance is warranted; if not, the central claim collapses to a within-user interpolation result. I therefore agree with the reader's conditional verdict and recommend making the subject-exclusive split a hard condition. Note the paper itself flags dataset rigor as future work ('constructing a more robust and methodologically rigorous original dataset... incorporating stricter validation protocols'), which is an internally acknowledged limitation that should be weighed in the verdict.","tokens_in":21290,"tokens_out":1731,"duration_ms":17096,"concrete_test":"Request the data-split script or rerun the protocol with an explicit subject-exclusive split: for each of the three datasets, hold out entire participants (e.g., 20% of participants) for testing, generate windows only after splitting, and re-run all five methods with the same seeds and hyperparameters. If the test accuracies in Table 5 drop or the Smooth-Distill margin over multitask shrinks to within one standard deviation, the headline generalization claim is not supported; if the rankings and margins survive, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that Smooth-Distill 'consistently outperforms' multitask baselines and matches Born-Again distillation at roughly half the training time. The load-bearing assumption is that the 80:20 train/test partition does not place windows from the same participant in both partitions. Section 4.2.1 says only 'The processed dataset was partitioned into training and testing sets in an 80:20 ratio' and 'Five-fold cross-validation with identical fold divisions ensured fair comparison,' with no statement that splitting is per participant. If the split is per window, then adjacent overlapping windows (window length 100, step size 60) from the same recording can appear in both train and test, directly inflating test accuracy. This matters most for the Sleep dataset, where 12 posture classes are recorded in controlled one-minute segments per participant and are nearly perfectly separable; the reported Task 1 test accuracy of 92.22% with tiny standard deviations (0.53) is suspiciously high for classifying 12 fine-grained sleep postures on a new participant. The same concern applies to WISDM Task 1 (68.59%), where the 4% improvement claim is made. The paper also reports p-values for paired fold t-tests (Table 8) but does not report the split protocol; if folds are drawn per window, all methods share the same leakage and pairwise comparisons are uninformative about generalization. The distinction between 'new user generalization' and 'within-user window memorization' is not established anywhere in the manuscript. Since the abstract explicitly claims consistent outperformance across evaluation scenarios and the dataset contribution is partly sold as enabling generalization studies, the split protocol is the single weakest link in the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Smooth-Distill, a self-distillation framework for multitask learning on wearable accelerometer data. The method uses an exponential moving average of the student network's parameters as a teacher (Eq. 4), eliminating the separate teacher-training phase of conventional distillation. The authors introduce a new sleep-posture dataset with 12 classes and three sensor placements, and evaluate on that dataset plus MHealth and WISDM. They report that Smooth-Distill achieves higher accuracy and F1 than single-task, multitask, SD-Dropout, and Born-Again distillation baselines, with substantially lower training time than Born-Again. The paper also provides ablation studies on the distillation weight, model architecture, and training ratio, as well as statistical significance tests.","tokens_in":21511,"tokens_out":4703,"duration_ms":51082,"significance":"If the empirical claims are validated, the contribution is practically relevant: it offers a computationally cheaper alternative to two-stage knowledge distillation for multitask wearable-sensor classification, and the new 12-posture, 3-placement sleep dataset is a useful public resource. The paper ships a code link and describes careful ablations, which are strengths. However, the headline claim of consistent superiority is weakened by the paper's own significance tests, and the central generalization claim depends on an unspecified data-partitioning protocol. The novelty of the EMA-teacher mechanism is also somewhat overstated given prior work on mean-teacher style self-distillation.","major_comments":[{"comment":"The train/test partition is described only as an '80:20 ratio' with no statement about whether the split is per participant or per window. Because windows are extracted with length 100 and step 60, overlapping windows from the same participant or recording can appear in both partitions if the split is per window. This would inflate test accuracy and make the generalization claims in Sections 5.1 and 6.5 unsupported. The authors must clarify the split granularity; if it is per window, they should add a subject-independent evaluation (e.g., leave-one-subject-out) to demonstrate generalization to new users, especially for the Sleep dataset where the reported 92.22% Task 1 test accuracy with a 0.53 standard deviation is suspiciously high for 12 fine-grained posture classes.","section":"Section 4.2.1, Data Preprocessing"},{"comment":"The claim that Smooth-Distill 'consistently outperforms alternative approaches' is contradicted by the paper's own significance tests: the p-value against Born-Again SD is 0.7173 for accuracy and 0.7431 for F1, which the text itself describes as 'statistically indistinguishable performance.' I recommend tempering the abstract and conclusion to 'matches Born-Again self-distillation at substantially lower training cost' unless the authors provide evidence of superiority on a metric that matters for deployment, such as new-user generalization. Additionally, the paper should clarify how many observations enter the paired t-test (five folds) and whether multiple-comparison corrections were applied, given that 10 pairwise comparisons are reported.","section":"Section 5.1, Table 8, and Section 6.2"},{"comment":"The training-time comparison is internally inconsistent. The text says Smooth-Distill 'increases training time to approximately 1.72 times the baseline,' but Table 7 lists Singletask training time as 5012 s and Smooth-Distill as 4368 s, which is a ratio of 0.87. If the intended baseline is the full singletask approach (two separate models, 2 x 5012 = 10024 s), then Smooth-Distill is about 0.44x of that baseline, not 1.72x. The same inconsistency appears in Section 7. Please correct the ratio and the '72% more training time' statement. The supported claim is that Smooth-Distill takes roughly 47% less time than Born-Again (4368 vs 8280 s), which should be stated precisely.","section":"Section 6.1 and Table 7"}],"minor_comments":[{"comment":"The contribution bullet states the Sleep dataset was collected from 23 participants, while Section 4.1.1 and Table 2 state 24 participants; please correct this inconsistency.","section":"Section 1, Section 4.1.1, Table 2"},{"comment":"The relation between 'five-fold cross-validation' and 'five independent random seeds' is ambiguous; please state how many runs are performed per fold and how seeds interact with fold divisions.","section":"Section 4.2.1"},{"comment":"The EMA-teacher mechanism is closely related to the Mean Teacher method (Tarvainen and Valpola, 2017), which is not cited; please discuss the difference relative to that work (e.g., distillation loss versus consistency loss, and the multitask setting) to position the novelty more precisely.","section":"Section 3.3"},{"comment":"The significance test is described as a 'paired t-test comparing mean fold accuracy,' but with only five folds the test has very low power; please report the number of paired observations and consider a signed-rank test or confidence intervals.","section":"Table 8"},{"comment":"The confusion matrices are difficult to read in the grayscale preprint; please ensure the figure is legible in the final printed version.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the unspecified split protocol. If the split is per window, the reported accuracies and the comparison across methods are not valid evidence for generalization to new users; the authors must either confirm a per-participant split or add subject-independent experiments. The statistical insignificance versus Born-Again is a further reason to soften the headline claim. The paper has useful empirical content and a new dataset, so I would not reject it outright, but the requested changes are load-bearing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent application paper, not a new method. Eq. 4 is exactly the Mean Teacher parameter update (Tarvainen & Valpola, 2017), which the paper never cites or compares against. What is actually new: the multitask application of an EMA self-teacher to joint activity and placement recognition, and a new 12-posture sleep dataset. The experiments are extensive and the efficiency comparison is useful.\n\nCredit where it is due: the Sleep dataset is a real resource with 24 participants, three sensor positions, and fine-grained posture labels; the MTL-net architecture is straightforward but appropriate. The training-time comparison against Born-Again distillation (nearly half the time) is a fair point, and the convergence analysis is qualitatively reasonable.\n\nThe central problem is the train/test split. Section 4.2.1 says only that data were partitioned 80:20 and folds were identical; it never says the split is per participant. With window length 100 and step 60, a per-window split puts overlapping windows from the same recording in both train and test. That matters most for the Sleep dataset, where each posture was recorded in a controlled one-minute segment per participant; the 92.22% test accuracy with std 0.53 would be far less informative under that protocol. The paper doesn't claim subject-independent generalization, but it also doesn't rule out the leaky reading, and the abstract's 'consistently outperforms' invites that interpretation. This must be fixed: state the split protocol explicitly, and if it is per-window, redo the experiments per participant.\n\nThe statistical analysis is another soft spot. Table 8 shows p=0.717 (accuracy) and p=0.743 (F1) between Smooth-Distill and Born-Again SD on the Sleep dataset. Section 6.2 honestly states they are statistically indistinguishable, but the abstract and introduction still say 'consistently outperforms alternative approaches.' That claim is too strong as written.\n\nThe authors also admit in Section 6.5 that 'a more robust and methodologically rigorous original dataset' is needed for future work, which is an honest limitation but undermines the dataset contribution as a generalization benchmark. Minor issues: the architecture comparison (Figures 5–7) lacks error bars or significance tests, and the 'reduced overfitting' claim rests on eyeballing training curves.\n\nWho is this for? The wearable-HAR community. A serious referee should definitely see it—not because the method is novel, but because the dataset and the empirical comparison are worth evaluating once the split protocol is clarified and the Mean Teacher lineage is acknowledged. My recommendation: send to review, and require the split description, a Mean Teacher baseline, and toned-down claims. If the split turns out to be per-window, the results are likely to change substantially.","headline":"Solid engineering paper: the new sleep dataset and multitask application are useful, but Eq. 4 is uncredited Mean Teacher and the split protocol is ambiguous, so the headline claims are not yet supportable.","tokens_in":22152,"tokens_out":3136,"would_cite":false,"duration_ms":31842,"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 wearable-sensing CNN that learns from its own smoothed weights matches two-stage self-distillation at half the training time, beating single-task and multitask baselines on both activity and placement recognition.","keywords":["self-distillation","multitask learning","human activity recognition","sensor placement detection","wearable accelerometer data","sleep posture recognition","exponential moving average teacher","knowledge distillation"],"falsifier":"Re-run the three experiments with a strict per-participant split (train on some participants, test on held-out participants, with identical folds and hyperparameters): if Smooth-Distill's edge over plain multitask training shrinks to within noise or reverses on the Sleep and WISDM datasets, window leakage rather than the distillation mechanism explains the reported gains. A second, cheaper check is to train Smooth-Distill with $\\lambda = 0$ (distillation term removed) and with the teacher frozen at its initialization; if either variant matches the full method, the EMA teacher itself is not the active ingredient.","tokens_in":21041,"feed_emoji":"⌚","tokens_out":12091,"duration_ms":108996,"temperature":0.7,"pith_summary":"The paper claims that the regularization benefits of knowledge distillation for multitask learning need no separate teacher: a CNN predicting both activity/sleep posture and sensor placement from accelerometer windows can be trained to match a smoothed copy of its own parameters, updated after every step as an exponential moving average of the student (Eq. 4 with $\\beta = 0.999$). On three datasets, a newly collected 12-posture sleep set and the public MHealth and WISDM sets, Smooth-Distill reports the highest or tied-highest test accuracy and F1 for both tasks in nearly every cell of Table 5, with its largest gains on the imbalanced MHealth activity task (test F1 53.97 versus 48.87 for born-again distillation). Because no teacher is pre-trained, training time on the Sleep dataset drops from 8280 to 4368 seconds and energy from 0.920 to 0.485 kWh compared with the two-stage baseline, which matters for wearable settings where models are retrained frequently. If true, the result turns self-distillation from a doubling of training cost into a one-pass regularizer.","feed_headline":"Self-taught model cuts distillation training time nearly in half","feed_subtitle":"A smoothed copy of its own weights teaches both sensing tasks in one pass, matching the costly two-stage teacher.","key_machinery":"The load-bearing object is the smoothed teacher, defined by the recurrence $\\theta^T_t = \\beta\\,\\theta^T_{t-1} + (1-\\beta)\\,\\theta^S_t$ with $\\beta = 0.999$: at every minibatch the student is updated by backpropagation on the combined task losses, and the teacher is then nudged a fraction $(1-\\beta)$ toward the student, so the teacher lags behind the student and carries a lower-variance average of its history. Its job is to emit stable soft targets whose KL divergence to the student's predictions, added to each task's cross-entropy with weight $\\lambda = 0.5$ in Eq. 5, steers the student toward confident and consistent predictions on both tasks while smoothing the noise of individual gradient steps, which the paper credits for the observed convergence stability and reduced overfitting relative to plain multitask training. The accompanying architecture MTL-net is a CNN that takes 100-sample accelerometer windows with a step of 60 and branches into two classification heads, one per task.","core_discovery":"Smooth-Distill's central claim is that a historical average of the training model is a sufficient teacher for multitask distillation. The teacher is never trained: after each student update its parameters move by $\\theta^T_t = \\beta\\,\\theta^T_{t-1} + (1-\\beta)\\,\\theta^S_t$ with $\\beta = 0.999$, so the teacher is an exponential moving average of the student's own weights, and the student minimizes, for each of two tasks, its cross-entropy to the labels plus $\\lambda$ times the KL divergence between its softened logits and the teacher's, with temperature $\\tau = 3.0$ and task weight $\\alpha = 0.5$ (Eq. 5). With $\\lambda = 0.5$, this yields the best or tied-best test accuracy and F1 across the Sleep, MHealth, and WISDM datasets for both posture/activity recognition and device-placement detection (Table 5), including 68.59% accuracy on WISDM Task 1, about four points above the single-task baseline. On the Sleep dataset the training time is 4368 seconds versus 8280 for Born-Again self-distillation, with 0.485 kWh versus 0.920 kWh of energy (Table 7). The paper's paired t-tests place Smooth-Distill and Born-Again in a statistically indistinguishable top tier ($p > 0.7$), both significantly ahead of single-task, multitask, and dropout-based distillation ($p < 0.001$); so within the paper's own statistics the claim is parity in accuracy with Born-Again, superiority in cost, and clear superiority over the non-distillation baselines.","pith_inferences":["I would expect the same EMA-teacher mechanism to transfer beyond accelerometry: nothing in the loss or update rule is modality-specific, so multitask gyroscope, PPG, ECG, or IMU pipelines, and task pairs beyond activity-plus-placement, should see similar one-pass regularization, although the paper only tests acceleration data.","An untested boundary condition follows from $\\beta = 0.999$: early in training the teacher barely differs from the student, so under very short training budgets or very small batches the distillation signal is weak and could even destabilize; a sweep over training length and batch size would reveal where the advantage appears.","Because the paper reports an 80:20 split without stating whether it is per participant or per window, the honest generalization estimate is still open; a leave-participants-out rerun on all three datasets would settle whether the gains survive on unseen wearers.","Since the paper's own statistics tie Smooth-Distill to born-again accuracy, the practical reading is that the method's value is chiefly operational: near-parity accuracy at roughly half the training cost, which a deployment-style comparison on retraining frequency and energy budgets would demonstrate directly."],"forward_implications":["Knowledge distillation for multitask wearable sensing becomes a one-pass procedure: no teacher is pre-trained, so a model can be retrained as often as new sensor data or labels arrive at roughly half the training time and energy of born-again distillation.","On imbalanced activity sets the benefit concentrates where it matters: Smooth-Distill raises MHealth Task 1 test F1 to 53.97 versus 48.87 for born-again and about 50 for single-task training, indicating the smoothing regularizer helps rare classes rather than only easy ones.","The benefit is not tied to one architecture: Smooth-Distill training improves LSTM, bidirectional LSTM, GRU, AnpoNet, and the CNN-based MTL-Net across the three datasets, and models retain usable performance with as little as 10% of the training data, with gains leveling off near a 70% training ratio.","For accuracy-critical deployments the practical choice is cost-driven, because the t-tests place Smooth-Distill and born-again distillation in the same statistical tier; the roughly 47% reduction in training time and energy becomes the deciding factor.","The released Sleep dataset, with 12 postures across three wearing positions, provides a public benchmark for joint sleep-posture and sensor-placement recognition."],"supporting_citations":[{"why":"Supplies the knowledge distillation formulation (softened targets, temperature, KL divergence) that Smooth-Distill adapts to a self-distillation setting.","marker":"[26]"},{"why":"Defines the Born-Again two-stage self-distillation baseline that Smooth-Distill matches in accuracy and beats in training time and energy.","marker":"[31]"},{"why":"Provides the dropout-based self-knowledge distillation method used as the SD-Dropout baseline, which underperforms on accelerometer tasks.","marker":"[36]"},{"why":"Supplies the MHealth public dataset used for evaluating both activity recognition and sensor placement tasks.","marker":"[33]"},{"why":"Supplies the WISDM public dataset, where Smooth-Distill reports its largest task-1 accuracy gain.","marker":"[35]"},{"why":"Motivates distilling knowledge in multitask learning and is extended here from separate teacher-student training to self-distillation.","marker":"[24]"},{"why":"Supports the viability of multitask learning for human activity recognition and sensor placement, the premise of the shared architecture.","marker":"[18]"},{"why":"Provides the AnpoNet architecture used in the ablation study and prior sleep-posture classification results this work builds on.","marker":"[37]"}],"fun_headline_variants":["Self-teacher from EMA weights slashes training time","Historical self-teacher halves distillation cost","No teacher needed: EMA self-distillation cuts cost","EMA weights teach student, halving training energy","Self-distillation without a teacher: cheaper and accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results stand or fall on the 80:20 train/test split being leakage-free: the paper says the windows were split 80:20 but never states whether all windows of a participant stayed on one side, so if windows from the same person appear in both partitions the reported accuracies are optimistically inflated and generalization to new users is not demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Self-teacher from EMA weights slashes training time","Historical self-teacher halves distillation cost","No teacher needed: EMA self-distillation cuts cost","EMA weights teach student, halving training energy","Self-distillation without a teacher: cheaper and accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000595,"raw_usage":{"total_tokens":2891,"prompt_tokens":1159,"completion_tokens":1732,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":775,"completion_tokens_details":{"reasoning_tokens":1660}},"tokens_in":775,"tokens_out":1732,"duration_ms":12669,"temperature":1.0,"reasoning_tokens":1660,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:14:17.727876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the three experiments with a strict per-participant split (train on some participants, test on held-out participants, with identical folds and hyperparameters): if Smooth-Distill's edge over plain multitask training shrinks to within noise or reverses on the Sleep and WISDM datasets, window leakage rather than the distillation mechanism explains the reported gains. A second, cheaper check is to train Smooth-Distill with $\\lambda = 0$ (distillation term removed) and with the teacher frozen at its initialization; if either variant matches the full method, the EMA teacher itself is not the active ingredient.","supporting_citations":[{"cited_title":"Knowledge Distillation for Multi-task Learning","cited_arxiv_id":"2007.06889","evidence_quote":"Motivates distilling knowledge in multitask learning and is extended here from separate teacher-student training to self-distillation."},{"cited_title":"Accurate Human Activity Recognition with Multi-task Learning","cited_arxiv_id":null,"evidence_quote":"Supports the viability of multitask learning for human activity recognition and sensor placement, the premise of the shared architecture."},{"cited_title":"Human Sleep Position Classification Using A Lightweight Model And Acceleration Data","cited_arxiv_id":null,"evidence_quote":"Provides the AnpoNet architecture used in the ablation study and prior sleep-posture classification results this work builds on."}],"review_version":1}