Integrated Sensing and ML Processing Semiconductor Device

The system segments the processing architecture by separating sensing functions performed at edge devices from complex machine learning model training performed at remote devices. Edge devices perform local preprocessing and inference using compressed models, while remote devices handle only model updates and retraining. This segmentation enables ML processing capability at the edge with reduced energy consumption compared to transmitting all sensing data to remote devices.

The patent introduces a new dimension of processing by implementing federated learning across multiple edge devices. Instead of a single centralized processing architecture, the system operates across distributed devices with coordinated model updates. This dimensional change enables local processing capability while sharing learning benefits across the network, reducing the need for energy-intensive data transmission to centralized remote devices.

The system extracts and processes only the essential features and compressed model representations at edge devices rather than transmitting complete raw sensing data to remote devices. This extraction approach maintains the necessary information for ML inference while significantly reducing privacy risks associated with transmitting sensitive raw data across networks.

The patent implements local quality by enabling each edge device to perform ML inference locally using compressed models downloaded from remote devices. This local processing ensures that sensitive sensing data remains on local devices rather than being transmitted to remote devices, thereby maintaining data privacy while still providing ML processing capability. Each device processes its own data locally with appropriate computational resources.

The system performs preliminary actions by pre-training complete machine learning models at remote devices with abundant computational resources. These pre-trained models are then compressed and downloaded to edge devices, which only need to execute the already-trained models for inference. This preliminary model training and compression at remote devices enables real-time local inference at edge devices without requiring them to have full ML training capabilities.

The patent applies parameter changes by transforming full-precision machine learning models into compressed versions suitable for edge devices. This involves changing model parameters through techniques such as quantization, pruning, and knowledge distillation. The compressed models have reduced computational requirements and can be executed in real-time on resource-constrained edge devices while maintaining acceptable accuracy, thus enabling real-time processing without requiring edge devices to have the same computational capabilities as remote devices.