{"id":"2386ad5f-4b08-44e8-9491-a29f6abe3fb0","arxiv_id":"2502.03364","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Self-supervised human activity recognition pre-training follows power-law scaling in data hours and model size, and user diversity improves scaling more than per-user data volume.","lead":"This paper reports the first explicit scaling laws for self-supervised pre-training in wearable human activity recognition: pre-training loss falls as a power law in data hours and model parameters, and adding new users improves scaling about three times faster than adding more data from the same users. If the findings hold, HAR researchers and device makers can choose model capacity and data collection strategies from quantitative curves instead of heuristics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 500k-step convergence assumption is not demonstrated; if large-config runs are under-trained, the fitted power-law exponents in Table 2 and Figure 3 mix optimization budget effects with data-scaling effects.","rationale":"The reader's weakest_assumption correctly identifies the fixed 500k-step convergence assumption as load-bearing, and I agree. My independent review finds no other concern that is more central: the power-law form itself is supported by consistent exponents across folds (Table 2), the downstream transfer results are plausible given the pre-training gains, and the 'first known scaling laws' claim is qualified but not falsified. However, the convergence assumption is not just a missing error bar; it is a premise that, if false, changes the meaning of every fitted exponent. The paper explicitly says convergence was 'found experimentally' (Section 3.1) but provides no curves, so this is a verifiable omission rather than a demonstrated failure. The concrete test I propose would settle whether the exponents are trustworthy or need to be recomputed with converged runs. Since the paper is already CONDITIONAL and this concern is a condition rather than a refutation, the reader's verdict does not need to change; the condition is simply reinforced. I therefore set verdict_should_be to UNCHANGED, with agreement_with_reader set to agree.","tokens_in":11956,"tokens_out":3477,"duration_ms":35841,"concrete_test":"For a single Extrasensory fold, run a small subset of the grid (e.g., the smallest and largest data fractions, the 1M and 63M parameter models, and both USER and RANDOM sampling) while logging pre-training loss every 10,000 steps up to 1,500,000 steps. Then compare the loss at 500,000 steps to the loss at 1,000,000 steps: if any configuration shows a relative decrease greater than, say, 1%, the fixed 500k-step budget is not converged for that configuration. If so, re-fit the exponents in Table 2 and Figure 3 using only converged runs, or explicitly characterize the bias introduced by the fixed budget. A secondary check: report the loss curves in the paper's supplement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central scaling-law claim (Section 4.2, Table 2, Figures 2–3) presumes that the reported pre-training losses are converged losses for every data size, model size, and sampling strategy. Section 3.1 states: 'We fix the number of pre-training steps to 500,000, which was found experimentally to be sufficient for convergence given our model and data sizes.' No convergence curves, validation curves, or evidence for this assertion are provided anywhere in the paper. This is load-bearing because the fitted power laws are only interpretable as scaling laws if each point is the converged loss; otherwise they are scaling laws of a fixed-budget training run, which conflate optimization convergence with data and model scaling. The risk is concrete: with 500,000 steps at batch size 2048, smaller-data runs complete vastly more epochs (Section 3.6 states 100% data is over 100 epochs, so a 10% data fraction would be over 1000 epochs), while larger-data runs receive far fewer epochs. If the largest models or largest data fractions have not reached their converged loss, the loss-vs-data exponent will be artificially flattened (large-data runs under-reported) and the loss-vs-parameter exponent will be artificially steepened or flattened depending on which models are under-trained. This also threatens the headline USER-vs-RANDOM comparison (Table 2), since different sampling strategies produce different data distributions that may converge at different rates; a fixed step count could differentially under-train one sampling strategy. The manuscript's own statement 'found experimentally' is an internal flag that this premise was checked but the evidence is omitted, so the condition must be settled before the fitted exponents can be used for quantitative guidance (e.g., choosing model capacity for a given data budget).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to establish scaling laws for wearable human activity recognition (HAR) by pre-training Masked Autoencoder Vision Transformers on the public Extrasensory dataset. The authors perform a grid search over data size (six fractions), sampling strategy (adding users vs. random examples), and model capacity (nine Transformer architectures), with three learning rates each, totaling 1,620 pre-training runs. They report that pre-training test loss follows a power law with respect to data hours and parameter count, that the USER sampling strategy yields an exponent roughly three times larger than RANDOM sampling, and that these trends translate to improved downstream linear-evaluation F1 on UCI HAR and WISDM Phone/Watch. They conclude that model capacity should be increased with data volume and that some previously published HAR works were under-parameterized.","tokens_in":12280,"tokens_out":3873,"duration_ms":37071,"significance":"If the findings hold, the paper would provide a practical, quantitative tool for choosing model capacity and pre-training data diversity in HAR, a domain where such design rules are currently lacking. The diversity result (more users beats more data per user) is potentially important and actionable. Strengths of the work include the use of public datasets, a large and carefully enumerated grid search, attention to the inference-compute constraints typical of HAR, and explicit downstream validation on three benchmarks. The main risks are methodological: the convergence assumption behind the scaling-law extraction is asserted but not demonstrated, the power-law fits are presented without uncertainty quantification, and the downstream trends are based on best-of-many selection without addressing multiple comparisons. These issues bear directly on the central claims, so they must be addressed before the paper can be accepted.","major_comments":[{"comment":"The claim that 500,000 pre-training steps is 'sufficient for convergence given our model and data sizes' is load-bearing for all of Section 4.2, but no convergence curves, validation curves, or stopping criteria are provided. With a fixed step count and batch size of 2048, smaller data fractions receive many more epochs than larger ones (e.g., 10% data would be over 1000 epochs if 100% is over 100 epochs), so if large-data runs are under-trained, the fitted loss-vs-data exponents in Table 2 and Figure 2 will be systematically flattened. Please provide representative pre-training loss curves for the smallest and largest data fractions and model sizes, for both sampling strategies, and show that the reported losses are within a small tolerance of the converged value, or otherwise qualify the scaling laws as fixed-budget scaling laws rather than converged-loss scaling laws.","section":"Section 3.1, Section 3.6"},{"comment":"The power-law fits are reported without confidence intervals, goodness-of-fit statistics, residual analysis, or even the number of points used in each fit. With only six data sizes per fold and exponents as small as -0.015 to -0.052, the claim that the USER exponent is 'roughly 3 times greater' than RANDOM needs uncertainty estimates; the observed ratios could easily be within noise. Please report standard errors or bootstrap intervals for the exponents, show the raw per-fold points overlaid with the fitted curves, and describe the fitting procedure (e.g., least squares on log-log loss vs. log hours, and whether intercepts are free per fold).","section":"Section 4.2, Table 2, Figure 2"},{"comment":"The downstream performance claims are based on selecting the best F1 score across 27 runs (9 architectures x 3 learning rates) in Figure 4 and across 180 runs in Figure 5. Selecting the best of many runs inflates the expected F1 and can create spurious monotonic trends with data size or capacity, especially since the number of runs per point is not matched across the compared conditions. Please report the mean and standard deviation or the full distribution across the 27 or 180 runs, and show that the best-run trend is consistent with the median-run trend. Without this, the claim that downstream performance 'translates' the pre-training scaling laws is not fully supported.","section":"Section 4.3, Figures 4 and 5"}],"minor_comments":[{"comment":"The text says 'we do not fix the amount of compute' immediately after stating that pre-training steps are fixed to 500,000 with a fixed batch size; this is confusing and should be clarified (e.g., the authors mean they do not use a compute budget in the Chinchilla sense, but the actual compute per run is fixed by the protocol).","section":"Section 3.1"},{"comment":"The description of the 9 encoder architectures (3 widths x 3 depths) would benefit from a table listing width, depth, and exact parameter count for each model, since parameter count is the independent variable in Figure 3 and Figure 5.","section":"Section 3.2"},{"comment":"The abstract mentions 5000 hours of data from 60 users, while Section 3.5 states the filtered dataset is approximately 1589 hours; these numbers should be reconciled to avoid misleading readers about the pre-training data scale.","section":"Abstract and Section 3.5"},{"comment":"The WISDM Phone F1 of 31.9% from scratch and 34.3% with linear evaluation is very low; the text explains it is due to including all 18 classes, but the caption and text should state this explicitly and cite the class list so readers do not misinterpret the result as a failure of the method.","section":"Table 1 and Section 4.3"},{"comment":"The augmentation comparison uses 'best' scores across data sizes and learning rates; it is unclear whether the same model and data selections were used for the with- and without-augmentation conditions, which could introduce selection bias. Please clarify the selection protocol.","section":"Section 4.5, Figure 7"},{"comment":"There are several typos and formatting issues, including 'paralellization' in Section 3, 'Téllez Villamizar and other' in the reference list, and inconsistent citation keys in Section 2.3 (e.g., 'Logacjov et al. (2024) Logacjov and Bach (2024)').","section":"References"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision rather than reject because the central direction is plausible and the missing evidence (convergence curves, uncertainty on the exponents, and selection-bias analysis) is obtainable within the scope of the paper. The 'first known scaling laws' claim should be tempered given that adjacent sensor-foundation-model work already introduces scaling-law analyses; the differentiation from Narayanswamy et al. (2024) is clear but should be stated in the abstract. The manuscript also appears to be a preprint with a large industrial-affiliation author list; the editor may wish to confirm that all authors have approved the submission and that the public-data claims are reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is the first to fit explicit power laws for HAR pre-training loss as a function of data hours and parameter count on raw sensor data. It also reports a user-diversity effect: adding new users gives a roughly 3x steeper exponent than adding more data per user. That is a genuinely useful result for people building sensor foundation models, and it contrasts with earlier saturation claims in self-supervised HAR. The compute is substantial (1620 runs, 62k TPU-hours), they use public datasets, and the exponents are consistent across five folds, which is real evidence the relationship is not noise.\n\nThe soft spots are real but addressable. The stress-test concern is on point: the paper fixes pre-training at 500,000 steps for every data size and model size, states this was \"found experimentally\" to be sufficient for convergence, and shows no convergence curves. If the large-data or large-model runs are under-trained, the fitted exponents mix optimization budget effects with data scaling. That matters for the headline USER-vs-RANDOM comparison, since sampling strategies may converge at different rates. This is the load-bearing premise of the paper, and the evidence is omitted. Also: no confidence intervals or residual analysis for the fits, best-of-many selection over 27 or 180 runs without correction, and no code or data release. These are not fatal—they are conditions—but they keep the paper from being fully reproducible as shipped.\n\nThe central claim is plausible and I lean toward believing it holds up, but the convergence premise must be demonstrated before I would use these exponents to choose model capacity. A few convergence curves plus a paragraph on how they verified all configs converged would settle most of my worry.\n\nThis paper is for researchers working on HAR pre-training, wearable sensing, and scaling laws in non-language domains. It deserves a serious referee. I would send it to review with the expectation of heavy revision, mainly around uncertainty quantification, the convergence evidence, and artifact release. As is, I'd bring it to a reading group to dissect the methodology, and I'd cite it with caveats if I referenced HAR scaling.\n\nMy bottom line: accept for review, conditional on the authors showing convergence curves and releasing code and data.","headline":"First explicit HAR scaling laws with a user-diversity effect, but the fixed-budget convergence premise needs evidence before the exponents are taken as quantitative guidance.","tokens_in":12839,"tokens_out":1239,"would_cite":true,"duration_ms":14065,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Wearable activity recognition follows power-law scaling with data and model size, and user diversity gives roughly three times steeper gains than more same-user hours.","keywords":["scaling laws","wearable human activity recognition","self-supervised pre-training","masked autoencoder","power law","data diversity","Transformers"],"falsifier":"Train one encoder architecture on a small and a large data slice from the same Extrasensory fold with step budgets of, say, 125,000, 500,000, and 2 million steps, and check whether the fitted loss-versus-data exponent steepens as the budget grows; if the exponent moves, the reported scaling laws are contaminated by the fixed training budget.","tokens_in":11762,"feed_emoji":"📈","tokens_out":8572,"duration_ms":66932,"temperature":0.7,"pith_summary":"This paper aims to establish the first quantitative scaling laws for wearable human activity recognition (HAR). It argues that the pre-training loss of a masked-autoencoder Transformer falls as a power law with both the amount of pre-training data and the number of model parameters, so model capacity can be chosen deliberately rather than by trial and error. The key practical claim is that adding new users to the pre-training set improves the loss about three times faster per hour of data than adding more data from existing users, making data diversity the dominant lever. The paper further shows that these pre-training laws translate into downstream gains on three benchmarks, beating from-scratch training, and that prior HAR studies that fixed small encoders may have underused their datasets.","feed_headline":"Adding users beats adding hours in wearable activity AI","feed_subtitle":"New scaling laws link model size and data volume to HAR accuracy, guiding design choices.","key_machinery":"The machinery is a masked autoencoder built on a Vision Transformer adapted to raw accelerometer and gyroscope windows: 128 samples at 50Hz, patches of 4 samples, 70% random patch masking, a small 2-block decoder, and a 9-model grid over widths of 512, 1024, and 2048 hidden units and depths of 5, 10, and 20 blocks, spanning 1M to 63M parameters. Pre-training fixes 500,000 steps at batch size 2048, which the paper takes as convergence for all configurations. The scaling behaviour is extracted by fitting power laws $L = aD^{-\\alpha}$ for loss versus data hours and $L = bN^{-\\beta}$ for loss versus parameters, separately for two sampling strategies, USER, adding whole users, and RANDOM, drawing random examples from all users, and then checking that these fits predict downstream linear-evaluation F1.","core_discovery":"The paper's central claim is that HAR pre-training has scaling laws of the same form as language and vision, and that these laws are useful for model design. On the Extrasensory dataset, about 1,589 hours from 60 users, pre-training loss of a masked-autoencoder ViT falls as $L = aD^{-\\alpha}$ in data hours and as a power law in parameter count across architectures from 1M to 63M parameters. The fitted exponents are stable across five user-split folds, and the USER sampling exponent, roughly $\\alpha \\approx 0.044$ to $0.052$, is about three times the RANDOM sampling exponent, roughly $\\alpha \\approx 0.015$ to $0.020$, so increasing the number of users is the more effective way to scale data. The laws transfer to downstream linear-evaluation F1 on UCI HAR, WISDM Phone, and WISDM Watch, where pre-training outperforms from-scratch training, and the best downstream results come from the largest encoder, suggesting that several earlier HAR studies were under-parameterized.","pith_inferences":["Beyond the paper, the fitted exponents imply a practical data-budgeting rule: for a fixed pre-training hour budget, maximizing the number of distinct users should generally beat maximizing hours per user, so collection protocols should prioritize subject coverage over recording length.","The paper's fixed 500,000-step convergence assumption is testable: re-fitting the exponents with multiple step budgets would reveal whether large-data runs were under-trained and whether the reported slopes change.","Because downstream F1 keeps improving with the largest model even at minimal pre-training data, the loss-based scaling law may understate the value of capacity for classification, and a direct loss-to-F1 transfer model would be a useful follow-up.","The same user-versus-random comparison could be applied to other sensor modalities and to contrastive pre-training objectives, which the paper lists as future directions."],"forward_implications":["Larger encoders are required to convert additional pre-training data into downstream gains, and the optimal capacity grows monotonically with data size for pre-training loss.","Data diversity is a primary lever: spending a fixed data budget on new users gives roughly three times steeper pre-training loss reduction than spending it on additional examples from existing users.","Pre-trained HAR models beat from-scratch baselines on UCI HAR, WISDM Phone, and WISDM Watch, with the largest 63M encoder performing best downstream.","Previously published HAR studies that fixed small encoder capacities, around 1.5M to 10M parameters, should be revisited with capacities of at least 30M parameters to see whether their reported plateaus disappear.","Augmentations such as random rotation and scaling improve downstream performance, especially at larger model scales, and can shift the optimal capacity upward."],"supporting_citations":[{"why":"Supplies the power-law scaling-law methodology from language models that this paper adapts to HAR.","marker":"Kaplan et al. (2020)"},{"why":"Provides the compute-optimal scaling framework and the practice of co-scaling data and parameters that motivates the HAR analysis.","marker":"Hoffmann et al. (2022)"},{"why":"Defines the masked autoencoder pre-training objective and architecture used as the backbone.","marker":"He et al. (2022)"},{"why":"Provides the Vision Transformer encoder that is adapted to accelerometer and gyroscope patches.","marker":"Dosovitskiy et al. (2021)"},{"why":"Contributes the Extrasensory dataset used for all pre-training runs.","marker":"Vaizman et al. (2017)"},{"why":"Documents the prior claim of saturating returns in self-supervised HAR that the paper's scaling laws and capacity argument challenge.","marker":"Haresamudram et al. (2022)"},{"why":"Provides the large 700,000-person-day dataset study whose fixed 10M-parameter ResNet is cited as a likely under-parameterized baseline.","marker":"Yuan et al. (2024)"},{"why":"Supplies the UCI HAR downstream benchmark used for linear-evaluation F1.","marker":"Reyes-Ortiz et al. (2013)"},{"why":"Supplies the WISDM Phone and Watch downstream benchmarks used for linear-evaluation F1.","marker":"Weiss (2019)"}],"fun_headline_variants":["User diversity beats hours in wearable activity AI","For wearable AI, more users beats more data","Scaling laws for HAR: user variety wins","Wearable activity models scale with users, not hours","HAR scaling laws reveal power law in user count"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that every pre-training run, regardless of data size or model size, reaches convergence within a fixed 500,000 training steps, so that loss differences reflect data and capacity rather than unfinished training.","fun_headline_variants_meta":{"raw":{"variants":["User diversity beats hours in wearable activity AI","For wearable AI, more users beats more data","Scaling laws for HAR: user variety wins","Wearable activity models scale with users, not hours","HAR scaling laws reveal power law in user count"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00069,"raw_usage":{"total_tokens":3129,"prompt_tokens":955,"completion_tokens":2174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":2101}},"tokens_in":571,"tokens_out":2174,"duration_ms":15138,"temperature":1.0,"reasoning_tokens":2101,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T04:57:25.236911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train one encoder architecture on a small and a large data slice from the same Extrasensory fold with step budgets of, say, 125,000, 500,000, and 2 million steps, and check whether the fitted loss-versus-data exponent steepens as the budget grows; if the exponent moves, the reported scaling laws are contaminated by the fixed training budget.","supporting_citations":[],"review_version":1}