PREFIXKD-EDGE: REAL-TIME LOCOMOTION-INTENT PREDICTION WITH PREFIX SELF-DISTILLATION ON LIGHTWEIGHT TEMPORAL NETWORKS
Keywords:
Early time-series classification, Edge intelligence, Human activity recognition, Knowledge distillation, Lightweight neural networks, Wearable sensingAbstract
Edge intent inference must make useful decisions before a sensing window is complete, while respecting tight memory and latency budgets. This paper studies early locomotion-mode prediction from wearable inertial sequences and introduces PrefixKD-Edge, a deployment-oriented framework that combines a compact temporal convolutional backbone, training-only full-window-to-prefix self-distillation, and confidence-gated decision timing. The task is evaluated on the public UCI Human Activity Recognition Using Smartphones dataset using its official subject-disjoint split. We retain three dynamic classes—level walking, stair ascent, and stair descent—and use the six supplied body-acceleration and gyroscope sequence channels rather than the 561 engineered features. A causal four-sample block mean reduces 50-Hz windows to 12.5 Hz, after which predictions are tested at 25%, 50%, 75%, and 100% observation. Across three random seeds, the 7,907-parameter PrefixKD-CNN attains macro-F1 scores of 94.30±1.76%, 97.64±1.62%, 98.00±1.43%, and 98.13±1.33%, respectively. Its measured FP32 TorchScript batch-1 median latency is 0.031–0.034 ms on one server CPU thread. With a confidence threshold of 0.90, the adaptive policy reaches 98.20±1.18% accuracy while observing 42.72±1.97% of the window on average (1.09 s). The same-backbone distillation gain is positive but small, so no statistical-significance claim is made. The results show that hardware-friendly backbones, prefix-aware training, and explicit decision policies can provide a practical accuracy–earliness operating point without adding inference parameters.References
[1] Lara O D, Labrador M A. A survey on human activity recognition using wearable sensors. IEEE Communications Surveys & Tutorials, 2013, 15(3): 1192-1209. DOI: 10.1109/SURV.2012.110112.00192.
[2] Bulling A, Blanke U, Schiele B. A tutorial on human activity recognition using body-worn inertial sensors. ACM Computing Surveys, 2014, 46(3): Art. 33. DOI: 10.1145/2499621.
[3] Yang J, Nguyen M N, San P P, et al. Deep convolutional neural networks on multichannel time series for human activity recognition. In: Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2015: 3995-4001.
[4] Hammerla N Y, Halloran S, Plötz T. Deep, convolutional, and recurrent models for human activity recognition using wearables. In: Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2016: 1533-1540.
[5] Ordóñez F J, Roggen D. Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity recognition. Sensors, 2016, 16(1): Art. 115. DOI: 10.3390/s16010115.
[6] Ronao Y S, Cho S B. Human activity recognition with smartphone sensors using deep learning neural networks. Expert Systems with Applications, 2016, 59: 235-244. DOI: 10.1016/j.eswa.2016.04.032.
[7] Murad A, Pyun J-Y. Deep recurrent neural networks for human activity recognition. Sensors, 2017, 17(11): Art. 2556. DOI: 10.3390/s17112556.
[8] Ignatov A. Real-time human activity recognition from accelerometer data using convolutional neural networks. Applied Soft Computing, 2018, 62: 915-922. DOI: 10.1016/j.asoc.2017.09.027.
[9] Yao S, Hu S, Zhao Y, et al. DeepSense: A unified deep learning framework for time-series mobile sensing data processing. In: Proceedings of the International Conference on World Wide Web (WWW), 2017: 351-360. DOI: 10.1145/3038912.3052577.
[10] Xing Z, Pei J, Yu P S. Early classification on time series. Knowledge and Information Systems, 2012, 31(1): 105-127. DOI: 10.1007/s10115-011-0400-x.
[11] Mori U, Mendiburu A, Dasgupta S, et al. Early classification of time series by simultaneously optimizing the accuracy and earliness. IEEE Transactions on Neural Networks and Learning Systems, 2018, 29(10): 4569-4578. DOI: 10.1109/TNNLS.2017.2764939.
[12] Schäfer P, Leser U. TEASER: Early and accurate time series classification. Data Mining and Knowledge Discovery, 2020, 34: 1336-1362. DOI: 10.1007/s10618-020-00690-z.
[13] Hochreiter S, Schmidhuber J. Long short-term memory. Neural Computation, 1997, 9(8): 1735-1780. DOI: 10.1162/neco.1997.9.8.1735.
[14] Cho K. Learning phrase representations using RNN encoder–decoder for statistical machine translation. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014: 1724-1734. DOI: 10.3115/v1/D14-1179.
[15] Lea C, Flynn M D, Vidal R, et al. Temporal convolutional networks for action segmentation and detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017: 156-165.
[16] Bai S, Kolter J Z, Koltun V. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv:1803.01271, 2018.
[17] van den Oord A. WaveNet: A generative model for raw audio. arXiv:1609.03499, 2016.
[18] Howard A G. MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv:1704.04861, 2017.
[19] Sandler M, Howard A, Zhu M, et al. MobileNetV2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018: 4510-4520.
[20] Jacob B. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018: 2704-2713.
[21] Lin J, Chen W-M, Lin Y, et al. MCUNet: Tiny deep learning on IoT devices. In: Advances in Neural Information Processing Systems, 2020, 33: 11711-11722.
[22] David R. TensorFlow Lite Micro: Embedded machine learning for TinyML systems. In: Proceedings of Machine Learning Systems, 2021, 3: 800-811.
[23] Banbury C R. MicroNets: Neural network architectures for deploying TinyML applications on commodity microcontrollers. In: Proceedings of Machine Learning Systems, 2021, 3.
[24] Banbury C. Benchmarking TinyML systems: Challenges and direction. arXiv:2003.04821, 2020.
[25] Teerapittayanon S, McDanel B, Kung H T. BranchyNet: Fast inference via early exiting from deep neural networks. In: Proceedings of the International Conference on Pattern Recognition (ICPR), 2016: 2464-2469.
[26] Geifman Y, El-Yaniv R. Selective classification for deep neural networks. arXiv:1705.08500, 2017.
[27] Guo C, Pleiss G, Sun Y, et al. On calibration of modern neural networks. In: Proceedings of the International Conference on Machine Learning (ICML), PMLR, 2017, 70: 1321-1330.
[28] Hinton G, Vinyals O, Dean J. Distilling the knowledge in a neural network. arXiv:1503.02531, 2015.
[29] Zhang L, Song J, Gao A, et al. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019: 3713-3722.
[30] Ioffe S, Szegedy C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Proceedings of the International Conference on Machine Learning (ICML), PMLR, 2015, 37: 448-456.
[31] Anguita D, Ghio A, Oneto L, et al. A public domain dataset for human activity recognition using smartphones. In: Proceedings of the European Symposium on Artificial Neural Networks (ESANN), 2013: 437-442.
[32] Anguita D, Ghio A, Oneto L, et al. Human Activity Recognition Using Smartphones. UCI Machine Learning Repository, 2013. DOI: 10.24432/C54S4K.
[33] Loshchilov I, Hutter F. Decoupled weight decay regularization. In: Proceedings of the International Conference on Learning Representations (ICLR), 2019.
[34] Loshchilov I, Hutter F. SGDR: Stochastic gradient descent with warm restarts. In: Proceedings of the International Conference on Learning Representations (ICLR), 2017.
[35] Paszke A. PyTorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems, 2019, 32: 8024-8035.