Experimental Analysis of Deep Echo State Networks for Ambient Assisted Living

Claudio Gallicchio, Alessio MicheliAI*AAL@AI*IA 2017 · OpenReview

Abstract

Ambient Assisted Living (AAL) systems rely on sequential sensor data to monitor and predict human behavior. However, the inherent noise and multi-scale temporal dynamics of such data pose significant challenges for traditional modeling techniques. This paper evaluates the effectiveness of Deep Echo State Networks (DeepESNs) for AAL applications. By stacking recurrent units into a hierarchical architecture, DeepESNs can capture complex temporal patterns more effectively than shallow networks. Our experimental results on indoor movement forecasting and human activity recognition demonstrate that DeepESNs achieve superior accuracy without increasing the computational footprint, making them ideal for deployment on resource-constrained edge devices in Wireless Sensor Networks (WSNs).

Introduction

The primary goal of Ambient Assisted Living (AAL) is to enhance the independence and quality of life for the elderly and disabled by proactively monitoring daily activities. This is typically achieved through Wireless Sensor Networks (WSNs) that collect non-intrusive data, such as Received Signal Strength (RSS). Despite their utility, RSS signals are notoriously noisy and heavily dependent on the temporal context of human movement.

While Recurrent Neural Networks (RNNs) are well-suited for sequential data, their training complexity and high computational demands often preclude their use on the low-power microcontrollers found in WSN nodes. Reservoir Computing (RC), particularly the Echo State Network (ESN), addresses this by utilizing a fixed, randomly initialized recurrent "reservoir" where only the linear output layer is trained. However, standard shallow ESNs often struggle to represent the multiple time scales inherent in human behavior—ranging from rapid signal fluctuations to slow transitions between activities. This study investigates whether a deep reservoir architecture can overcome these limitations.

Methodology

The core innovation explored in this work is the Deep Echo State Network (DeepESN). Unlike a standard ESN that uses a single large reservoir, a DeepESN organizes recurrent units into a stacked, hierarchical structure.

Architecture of a Deep Echo State Network (DeepESN) showing the hierarchical stacking of reservoir layers.

In this architecture, the first layer receives the external input $\mathbf{u}(t)$, and each subsequent layer receives the state of the previous layer as its input. This hierarchical processing allows the network to develop progressively more abstract representations of the temporal dynamics. To isolate the impact of depth, the authors implemented a strictly fair comparison: a DeepESN with 10 layers of 10 units each was compared against a shallow ESN with a single 100-unit layer. Both models shared the same total number of recurrent units and trainable parameters, ensuring that any performance gains were attributable to the architectural depth rather than model size.

Experimental Evaluation

The models were evaluated on two real-world benchmark datasets from the UCI Machine Learning Repository, both utilizing RSS data from WSNs:

  1. Indoor Movement Forecasting: A sequence-to-element binary classification task aimed at predicting whether a user will change rooms or remain in the current one. The input consists of 4-dimensional RSS streams from fixed anchors and a mobile wearable node.
  2. Human Activity Recognition (HAR): A sequence-to-sequence multi-class classification task involving 7 daily activities (e.g., sitting, standing, walking, cycling). This task uses 6-dimensional RSS data from three wearable sensors.
Examples of noisy RSS input signals from the Indoor Movement Forecasting dataset.

As shown in Figure 2, the input signals are characterized by significant noise and rapid fluctuations, highlighting the difficulty of extracting meaningful behavioral patterns from raw RSS data.

Results and Discussion

The experimental results consistently favored the deep architecture across both tasks:

  • Indoor Movement Forecasting: The DeepESN achieved a test accuracy of 90%, significantly outperforming the shallow ESN's 84%.
  • Human Activity Recognition: The DeepESN reached a 7-class accuracy of 77%, compared to 74% for the shallow counterpart.

These findings suggest that "deep is better than wide" for temporal AAL data. The hierarchical structure of the DeepESN naturally captures multi-time-scale dynamics, allowing it to filter high-frequency noise in early layers while modeling low-frequency behavioral shifts in deeper layers. Furthermore, the DeepESN demonstrated superior generalization to unseen environments in the movement forecasting task.

Crucially, these performance gains were achieved with a very small total network size (100 units). This efficiency is vital for AAL, as it allows sophisticated behavioral modeling to be embedded directly onto the microcontrollers of WSN nodes, enabling real-time, local processing without relying on cloud infrastructure.

Conclusion

This research demonstrates that Deep Echo State Networks are a powerful and efficient tool for Ambient Assisted Living. By leveraging hierarchical reservoirs, DeepESNs provide a significant accuracy boost over shallow models while maintaining a low computational profile suitable for edge devices. Future work will focus on deploying these models in live smart home environments and investigating the specific features learned by different layers of the deep reservoir.

Summary updated

BibTeX
@inproceedings{gallicchio2017experimental,
  author = {Claudio Gallicchio and Alessio Micheli},
  title = {{Experimental Analysis of Deep Echo State Networks for Ambient Assisted Living}},
  booktitle = {AI*AAL@AI*IA},
  year = {2017},
  url = {https://openreview.net/forum?id=4twkuUF5PY},
}
Figure