Industry Expert Talk: Signal Processing Inspired AI for Sensing - an Industry Perspective
This talk explores "AI for Sensing," a paradigm shifting physical sensing systems away from handcrafted, application-specific data pipelines toward lightweight, sensor-centric artificial intelligence. Rather than relying on rigid downstream processing, the primary goal is to integrate advanced machine learning directly into every phase of the sensing workflow—from initial transducer acquisition and adaptive calibration to noise reduction and multi-modal feature representation. By embedding these soft-sensing pipelines directly onto sensing hardware, the approach establishes an application-agnostic, sensor-specific soft architecture capable of running efficiently on-board at low power, low latency, and low memory footprint.
Key evidence and implementations demonstrate the versatility of this framework across diverse physical sensor modalities. By using attention mechanisms and auto-encoders, the system learns rich feature embeddings that act as a master feature representation or "vocabulary" for a given sensor type, allowing downstream applications to perform supervised learning without retraining the core sensing layer. Practical applications highlighted include closed-loop control via reinforcement learning for adaptive sampling and auto-gain control, neural signal enhancement for ECG physiological monitoring, microwave radar for concealed and in-body imaging, acousto-optic sensing, and quantum electromagnetic field sensing. Additionally, generative AI models can automate the structural design of complex physical hardware, such as plasmonic sensors used for micro-plastic detection in water.
Ultimately, unifying signal processing morphology with machine learning creates Sensor Specific Intelligent Chips (SSICs) that combine physical transducers with intelligent, self-adapting acquisition pipelines. This approach elevates sensor sensitivity, dynamic range, and system reliability while making AI-generated feature representations syntactically interpretable and scalable. By transforming raw sensor outputs into standardized, software-defined feature representations, this architecture establishes a foundational building block for future intelligent cyber-physical and IoT systems.
