Due to the addition of new team members and parallel projects, we request a Storage Takeover from our current allocation (NAISS 2024/22-1182). This is an update on our current projects: We developed an interpretable self-supervised network optimized for resource-constrained wearable devices. Unlike traditional deep learning methods, our method requires no seizure data during initial training and incrementally improves detection performance at runtime using newly acquired data—eliminating energy-intensive retraining. This enables real-time, adaptive seizure alerts while maintaining low power consumption. Biologically Inspired Lightweight Inference: To address the energy inefficiency of artificial/deep neural networks (ANN/DNN), we implement a Forward-Forward algorithm-based inference scheme. This biologically plausible approach reduces computational overheads compared to back-propagation-driven models, aligning closer to the human brain’s efficiency. We introduce a verification method to enhance the reliability of safety-critical applications. This method mitigates adversarial vulnerabilities and reduces over-approximation errors in formal verification, improving scalability and precision for robust model certification.