Sensor-based Human Activity Recognition (HAR) using data from wearable and smartphone sensors such as accelerometers and gyroscopes is a mature yet rapidly evolving field with wide-ranging applications in healthcare monitoring, sports analytics, fitness tracking, mobile computing, and ubiquitous human-computer interaction. Deep learning models using convolutional, recurrent, and transformer architectural components have become the de facto standard for modeling these multivariate time series, achieving strong performance on benchmark datasets but often struggling with deployment on edge devices under realistic constraints such as limited compute, energy budgets, and inter-subject variability.

Liquid Neural Networks (LNNs), including Liquid Time-constant Networks (LTCs) [1] and Neural Circuit Policies (NCPs) [2], are a recent class of continuous-time recurrent neural architectures in which each unit’s time constant is dynamically modulated by its state and input, yielding adaptive memory and strong expressivity for time-series modeling. LNNs have demonstrated superior robustness, interpretability, and generalization in control and navigation tasks [3], achieving competitive or better performance than much larger recurrent networks while using only a handful of neurons and parameters [4]. Their continuous-time formulation and input-dependent dynamics make them particularly attractive for irregular, noisy sensor streams and edge deployment scenarios.

Despite these promising properties, the potential of Liquid Neural Networks for sensor-based HAR and highly constrained real-world deployment remains largely unexplored. This thesis proposes to systematically investigate LNN architectures for sensor-based HAR, focusing on their ability to (i) achieve competitive or superior recognition performance, (ii) maintain inherent robustness regarding inter-subject variability, and (iii) operate efficiently on edge hardware in terms of parameters, latency, and energy usage. The work will compare LNN-based HAR models against performant and efficient baselines such as TinyHAR.

References

1. Hasani, R., Lechner, M., Amini, A., Rus, D., & Grosu, R. (2021). Liquid Time-Constant Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 35(9), 7657–7666.
2. Lechner, M., Hasani, R., Amini, A., Henzinger, T. A., Rus, D., & Grosu, R. (2020). Neural Circuit Policies Enabling Auditable Autonomy. Nature Machine Intelligence, 2, 642–652.
3. Chahine, M., et al. (2023). Robust Flight Navigation Out of Distribution with Liquid Neural Networks. Science Robotics, 8(74), eadc8892.
4. Zong, S., Bierly, A., Boker, A., & Eldardiry, H. (2025). Accuracy, Memory Efficiency and Generalization: A Comparative Study on Liquid Neural Networks and Recurrent Neural Networks.

• Review the theory and practice of Liquid Neural Networks (e.g., LTC, NCP, CfC) and their applications to multi-channel sensor signal data.
• Design, implement, train and evaluate LNN-based architectures for sensor-based HAR using established datasets in terms of classification performance .
• Investigate training regimes that maximize the liquid dynamics to achieve out-of-the-box robustness against inter-subject variability.
• Evaluate the proposed models in terms of computational efficiency representative edge platforms.
• Benchmark LNN-based models against strong generalized baselines

• Solid understanding of deep learning concepts, particularly recurrent architectures and optionally neural ODEs.
• Proficient programming skills in Python and PyTorch.
• Experience working with time series or sensor data is a plus but can be acquired during the thesis.
• Interest in model efficiency and edge deployment is a plus.

• A paragraph explaining your motivation.
• Your study program (Bachelor/Master), current semester, and field of study.
• A transcript of records (courses and grades).
• Your programming experience.
• Any areas of interest relevant to the topic.
• Your CV (if available)