More Features Are Not More Evidence: Limits of Training-Free Human Activity Recognition with Jev

Orhan Konak

cs.AISep 28, 2026

More Features Are Not More Evidence: Limits of Training-Free Human Activity Recognition with Jev

Authors: Orhan Konak

Organizations: Hasso Plattner Institute, University of Potsdam Potsdam, Germany

Abstract

General-purpose models promise sensor-based decisions without training a task-specific classifier, which could reduce the dependence of Human Activity Recognition (HAR) on labeled data. Yet it remains unclear whether such models can directly interpret deterministic descriptions of physical sensor signals well enough to replace or complement trained HAR models. We study this question using Jev, a fixed general-purpose probabilistic decision model, on 1,800 class-balanced accelerometer windows from WISDM, UCI341, and PAMAP2. Jev receives no labeled examples, retrieval context, or HAR-specific parameter updates. We evaluate three deterministic sensor representations and compare 5,400 Jev decisions with a generative baseline and three supervised HAR models. Jev remains far below supervised recognition, with its strongest representation reaching macro-F1 of 0.038, 0.118, and 0.089 across the three datasets, compared with 0.686 to 0.907 for the supervised models. More numerical features do not improve Jev. Instead, they reduce recognition on all three datasets, while augmenting the same numerical evidence with a deterministic semantic rendering partially recovers performance, although the experiment does not isolate semantics from the accompanying serialization and redundancy changes. Jev is fast and inexpensive to query, but its probabilities are not reliably calibrated for recognition. A post-hoc fusion analysis finds a small improvement on WISDM that does not replicate on UCI341 or PAMAP2. These results show that training-free sensor decisions depend not only on the information available in the signal, but also on whether the model can use the representation through which that information is exposed. The sensor-to-model interface should therefore be treated as part of the model evaluation rather than as a neutral preprocessing step.

Figures & tables

Dataset	Subjects	Classes	Selected signal	Native rate	Eligible	Benchmark
WISDM (UCI507)	51	18	Watch acceleration, xyz	Timestamped	31,945	900
UCI341 (HAPT)	30	6	Smartphone total acceleration, xyz	50 Hz	2,579	300
PAMAP2 (UCI231)	9	12	Hand/wrist $\pm 16g$ acceleration, xyz	100 Hz	3,833	600

Table 1. Frozen study inputs. Each window has 100 samples at 20 Hz, spans 5 s, and has zero overlap. “Eligible” is the full preprocessed population; “benchmark” is the matched evaluation subset.

Dataset	Model	Representation	Accuracy [95% CI]	Macro-F1 [95% CI]
WISDM	Jev	R1	0.0733 [0.0653, 0.0822]	0.0231 [0.0153, 0.0320]
WISDM	Claude	R1	0.1278 [0.1173, 0.1386]	0.0507 [0.0428, 0.0587]
WISDM	Jev	R2	0.0556 [0.0546, 0.0562]	0.0059 [0.0058, 0.0059]
WISDM	Claude	R2	0.1422 [0.1288, 0.1569]	0.0666 [0.0536, 0.0815]
WISDM	Jev	R3	0.1122 [0.1001, 0.1237]	0.0378 [0.0341, 0.0414]
WISDM	Claude	R3	0.1467 [0.1321, 0.1616]	0.0891 [0.0757, 0.1019]

Table 2. All matched main recognition results, reported as estimate [95% subject-cluster interval]. Scores are proportions. Jev uses native choice on every window; the other systems use their frozen class predictions.

Dataset	Rep.	Valid/total	Native $\neq$ argmax	BS	NLL	ECE	AURC	argmax acc.
WISDM	R1	882/900	17	1.0282	3.3020	0.2495	0.8408	0.0737
WISDM	R2	883/900	0	1.1350	3.4987	0.4154	0.8908	0.0555
WISDM	R3	890/900	22	1.0571	3.5944	0.2532	0.8600	0.1135
UCI341	R1	297/300	4	0.9523	1.9385	0.3665	0.7022	0.1751
UCI341	R2	292/300	0	1.1045	2.2255	0.5018	0.7334	0.1610
UCI341	R3	291/300	2	0.9400	1.8631	0.3489	0.6306	0.2405

Table 3. Frozen Jev probability metrics on valid vectors only; no repair. “Native

\neq

argmax” counts disagreements between the native choice and the canonical probability argmax. BS, NLL, ECE, and AURC are point estimates; the frozen analysis does not provide intervals for these metrics.

		Jev				Claude
Dataset	Rep.	Median (s)	p95 (s)	Max (s)	USD/1k	Median (s)	USD/1k
WISDM	R1	0.296	0.668	5.327	0.0376	1.457	1.6342
WISDM	R2	0.292	0.561	2.225	0.0492	1.471	2.0587
WISDM	R3	0.300	0.561	1.930	0.0707	1.519	3.5469
UCI341	R1	0.292	0.480	1.184	0.0348	1.489	1.5446
UCI341	R2	0.292	0.507	5.335	0.0461	1.491	1.9632

Table 4. Observed sequential API latency and frozen list-price cost. Each row is a dataset/representation condition. Costs are USD per 1,000 windows, calculated from logged usage; timing excludes acquisition and preprocessing.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

$\lambda$	Accuracy	Macro-F1	$\Delta$ Accuracy	$\Delta$ Macro-F1
0.00	0.7056 [0.6562, 0.7494]	0.7050 [0.6578, 0.7469]	0.0000 [0.0000, 0.0000]	0.0000 [0.0000, 0.0000]
0.05	0.7056 [0.6562, 0.7494]	0.7049 [0.6577, 0.7469]	0.0000 [0.0000, 0.0000]	-0.0001 [-0.0004, 0.0000]
0.10	0.7067 [0.6581, 0.7497]	0.7060 [0.6596, 0.7474]	0.0011 [0.0000, 0.0034]	0.0010 [-0.0003, 0.0034]
0.15	0.7101 [0.6618, 0.7528]	0.7091 [0.6632, 0.7498]	0.0045 [0.0011, 0.0091]	0.0041 [0.0006, 0.0087]
0.20	0.7135 [0.6652, 0.7561]	0.7120 [0.6662, 0.7527]	0.0079 [0.0033, 0.0136]	0.0070 [0.0024, 0.0125]
0.25	0.7112 [0.6640, 0.7531]	0.7093 [0.6644, 0.7496]	0.0056 [-0.0011, 0.0127]	0.0043 [-0.0024, 0.0113]

Appendix

Table 5. WISDM: all exploratory fusion weights. Scores and paired differences are proportions with 95% subject-cluster intervals.

$\lambda$	Accuracy	Macro-F1	$\Delta$ Accuracy	$\Delta$ Macro-F1
0.00	0.8969 [0.8648, 0.9288]	0.8974 [0.8648, 0.9291]	0.0000 [0.0000, 0.0000]	0.0000 [0.0000, 0.0000]
0.05	0.8935 [0.8627, 0.9241]	0.8940 [0.8627, 0.9246]	-0.0034 [-0.0108, 0.0000]	-0.0034 [-0.0106, 0.0000]
0.10	0.8935 [0.8627, 0.9241]	0.8940 [0.8627, 0.9246]	-0.0034 [-0.0108, 0.0000]	-0.0034 [-0.0106, 0.0000]
0.15	0.8900 [0.8552, 0.9231]	0.8905 [0.8547, 0.9234]	-0.0069 [-0.0169, 0.0000]	-0.0069 [-0.0170, 0.0000]
0.20	0.8866 [0.8497, 0.9221]	0.8870 [0.8492, 0.9226]	-0.0103 [-0.0254, 0.0035]	-0.0103 [-0.0253, 0.0034]
0.25	0.8832 [0.8462, 0.9181]	0.8835 [0.8455, 0.9185]	-0.0137 [-0.0304, 0.0000]	-0.0139 [-0.0306, 0.0001]

Appendix

Table 6. UCI341: all exploratory fusion weights. Scores and paired differences are proportions with 95% subject-cluster intervals.

$\lambda$	Accuracy	Macro-F1	$\Delta$ Accuracy	$\Delta$ Macro-F1
0.00	0.8054 [0.7795, 0.8265]	0.8061 [0.7786, 0.8255]	0.0000 [0.0000, 0.0000]	0.0000 [0.0000, 0.0000]
0.05	0.8054 [0.7782, 0.8276]	0.8063 [0.7776, 0.8266]	0.0000 [-0.0046, 0.0050]	0.0002 [-0.0045, 0.0052]
0.10	0.8037 [0.7783, 0.8248]	0.8048 [0.7783, 0.8241]	-0.0017 [-0.0081, 0.0063]	-0.0013 [-0.0084, 0.0067]
0.15	0.8020 [0.7764, 0.8234]	0.8033 [0.7768, 0.8225]	-0.0034 [-0.0086, 0.0034]	-0.0028 [-0.0089, 0.0046]
0.20	0.8003 [0.7750, 0.8222]	0.8018 [0.7755, 0.8211]	-0.0051 [-0.0128, 0.0029]	-0.0043 [-0.0125, 0.0042]
0.25	0.7970 [0.7696, 0.8208]	0.7978 [0.7679, 0.8195]	-0.0085 [-0.0167, 0.0000]	-0.0083 [-0.0174, 0.0011]

Appendix

Table 7. PAMAP2: all exploratory fusion weights. Scores and paired differences are proportions with 95% subject-cluster intervals.

Evidence	Artifact or directory
Preprocessing/representations	docs/protocol_v1.md ; docs/dataset_adaptation_protocol_v1.md ; configs/semantic_thresholds_v1.yaml
Benchmark and deviations	docs/training_free_evaluation_protocol_v2.md ; docs/training_free_v2_r3_protocol_deviation.md
Jev request contract	docs/jev_inference_protocol_v1.md
Main scores and intervals	results/training_free_v2/scored_jev/summary.csv ; paired_differences.csv in the same directory
Probability audit	probabilistic.csv ; probabilistic_details.json in the main scoring directory
Operational evidence	operational.csv in the main scoring directory and results/training_free_v2/scored/

Appendix

Table 8. Compact artifact map. Paths are relative to the study repository; full hashes and provenance reside in the accompanying artifact index and frozen manifests.

Explore similar work

Joint-Embedding Predictive Architecture for Sensor-based Activity Recognition
Jul 17, 2026Mohd Halim Mohd Noor, Abdulrahman M. A. BarakaHuman Activity Recognition Joint-Embedding Predictive Architectures
School of Computer Sciences, Universiti Sains Malaysia, 11800 Pulau Pinang, Malaysia · Artificial Intelligence Research Center, Ajman University, Ajman, United Arab Emirates · Al-Quds Open University, Ramallah, Palestine
BenchHAR: Benchmarking Self-Supervised Learning for Generalizable Sensor-based Activity Recognition
May 8, 2026Yize Cai, Rui Feng, Anlan Yu +2Human Activity Recognition Self-Supervised Learning
The Hong Kong University of Science and Technology (Guangzhou) · Peking University · Singapore-MIT Alliance for Research and Technology
RAG-HAR+: Towards Cost-Efficient LLM-Based Human Activity Recognition for Edge Deployment
Jul 29, 2026Hansi Karunarathna, Nirhoshan Sivaroopan, Chamara Madarasingha +2Human Activity Recognition Multi-Modal Sensing
University of Sydney, Australia · Curtin University, Australia · Colorado State University, USA

Jul 17, 2026, eess.SPJ/K move · Enter open · S save

Joint-Embedding Predictive Architecture for Sensor-based Activity Recognition

Mohd Halim Mohd Noor, Abdulrahman M. A. Baraka

School of Computer Sciences, Universiti Sains Malaysia, 11800 Pulau Pinang, Malaysia · Artificial Intelligence Research Center, Ajman University, Ajman, United Arab Emirates · Al-Quds Open University, Ramallah, Palestine

Human Activity Recognition Joint-Embedding Predictive Architectures Covariance

Open paper AskPDF arXiv

Sensor-based human activity recognition (HAR) has achieved significant progressed in fully supervised learning settings. However, these supervised learning models rely on large amount of labeled data, which require labor-intensive collection and meticulous annotation. To address these challenges, this paper proposes a Joint Embedding Predictive Architecture framework tailored for sensor-based HAR, designed to learn robust and generalizable representations from unlabeled datasets. The proposed framework features an encoder designed to explicitly model both the fine-grained local temporal representations within individual window and the long-term temporal sequence of adjacent windows. Furthermore, we introduce an improved Variance-Invariance-Covariance Regularization (VICReg) objective function that incorporates computationally lightweight norm term to stabilize the JEPA pre-training phase. This term balances variance, invariance and covariance constraints to prevent representation collapse. The proposed HAR-JEPA framework is evaluated using two benchmark continuously performed activity datasets. The results show that high-quality representations are successfully learned by the proposed framework. Furthermore, the representations learned by HAR-JEPA demonstrates superior generalization on minority, high variance transitional activities such as sit-to-stand and sit-to-lie where supervised learning tend to overfit due to limited support.

Abstract

Explore similar work

Joint-Embedding Predictive Architecture for Sensor-based Activity Recognition

BenchHAR: Benchmarking Self-Supervised Learning for Generalizable Sensor-based Activity Recognition

RAG-HAR+: Towards Cost-Efficient LLM-Based Human Activity Recognition for Edge Deployment