Speakers
Description
Anomaly detection is a critical task in modern AI systems, from healthcare monitoring to IoT sensor networks. However, traditional approaches, whether statistical methods or deep learning models, often require significant computational resources, labeled training data, and lack interpretability. This makes them unsuitable for deployment on resource-constrained edge devices, where real-time, energy-efficient, and adaptive detection is essential.
We present HaNDS (Habituation-Driven Neuromorphic Anomaly Detection System), a biologically-inspired framework that mimics the habituation behavior of biological neurons to detect novel or surprising events in time-series data. Unlike traditional machine learning models, HaNDS does not require training data or complex optimization. Instead, it leverages adaptive thresholds and neuronal dynamics, decay, recovery, and novelty spiking, to identify anomalies in a lightweight, interpretable manner.
HaNDS simulates habituating neurons where repeated exposure to the same stimulus weakens the response (habituation), while novel input triggers a spike (novelty detection). Each neuron is characterized by three parameters: decay rate ($\beta$), recovery rate ($\alpha$), and novelty threshold ($\theta$). By deploying multiple neurons with different sensitivities, HaNDS provides diverse novelty profiles, enabling adaptive detection across varying signal contexts.
We implemented HaNDS in C++ for maximum efficiency and benchmarked it on the ECG5000 dataset (5000 ECG sequences, 140 timesteps each). Our results demonstrate that HaNDS achieves competitive anomaly detection performance compared to classical methods (Z-score, MAD, Isolation Forest, One-Class SVM) while consuming significantly less energy. We also introduce entropy-based analysis to quantify the "surprise level" of novelty triggers and energy estimation to measure computational efficiency.
HaNDS is designed for the full AI lifecycle, from prototyping on Ubuntu Desktop to deployment on Ubuntu Server for batch processing, and finally to Ubuntu Core for real-time edge inference on IoT devices like Raspberry Pi. Its lightweight, CPU-only design makes it ideal for resource-constrained environments, aligning perfectly with the growing demand for edge intelligence in smart homes, industrial IoT, and healthcare monitoring.
We envision extending HaNDS to multi-modal sensor data (audio, vibration, network traffic), exploring spiking neural network hardware compatibility, and integrating it into federated learning frameworks for privacy-preserving edge AI.
This work bridges neuroscience, edge computing, and open-source AI, offering a novel, efficient, and interpretable approach to anomaly detection for the Ubuntu community and beyond.