Een Neuro-Symbolisch Raamwerk voor Veilige 'Edge AI': Faseruimtereconstructie Gecombineerd met Fysieke 'Reservoir Computing'

The deployment of Artificial Intelligence on resource-constrained edge devices is hindered by the high energy cost of iterative training, the need for large datasets, and the risk of transmitting sensitive data to cloud servers. This thesis proposes Phase Space Reconstruction-based Physical Reservoir Computing (PSR-PRC), a neuro-symbolic framework that addresses these challenges by combining temporal signal preprocessing with the intrinsic physics of electronic devices to enable one-shot learning entirely on-chip. At the algorithmic level, Phase Space Reconstruction with time-delay embedding unfolds one-dimensional time series into higher-dimensional trajectories, which are quantized using Equal Density Bins. This preprocessing significantly reduces the state richness requirements of the reservoir, allowing the system to operate with a small number of physically independent devices without complex interconnections. At the hardware level, the framework is implemented using three distinct mechanisms: Negative Bias Temperature Instability in pFETs, the Floating Body effect in bulk nFETs, and Ovonic Threshold Switching. All implementations exploit the fading memory, nonlinearity, and history-dependent dynamics inherent in these devices. Foundry-fabricated CMOS devices and SiGeAsSe OTS devices are characterized through both simulation and hardware measurements, demonstrating that standard commercial components or novel devices can serve as effective reservoir elements. The system is validated on multiple real-world temporal signal processing tasks. In gait authentication using the OU-ISIR dataset, the algorithm achieves an Equal Error Rate of 0.08% for 72 subjects with one-shot learning. Hardware implementations with 9 devices achieve EERs ranging from 2.9% to 5.1% depending on the device mechanism. For EEG classification, the system achieves an average Area Under the Curve of 0.94 for six-class hand movement recognition without artifact removal preprocessing, competitive with deep learning approaches that require significantly more computational resources. Walking activity classification is demonstrated using IMU sensor data from 14 subjects across multiple sessions. Finally, the thesis addresses security through the concept of User-Device Entanglement. The inherent manufacturing variability of CMOS devices provides each hardware instance with a unique physical signature, functioning as a Physical Unclonable Function. The system physically entangles the user's biometric data with this hardware-specific signature at the signal level, creating an inseparable composite feature vector. This dual-factor design ensures that successful authentication requires both the correct biometric input and the original physical chip, providing intrinsic protection against data theft and replay attacks without requiring cloud communication.