Selective and compressive sensing for energy-efficient implantable neural decoding

The spike classification is a critical step in implantable neural decoding. The energy efficiency issue in the sensor node is a big challenge in the entire system. Compressive sensing (CS) provides a potential way to tackle this problem. However, the overhead of signal reconstruction constrains the compression in sensor node and analysis in remote server. In this paper, we design a new selective CS architecture for wireless implantable neural decoding. We implement all the signal analysis on the compressed domain. To achieve better energy efficiency, we propose a two-stage classification procedure, including a coarse-grained screening module with softmax regression and a fine-grained analysis module based on deep learning. The screening module completes the low-effort classification task in the front-end and transmits the compressed data of high-effort task to remote server for fine-grained analysis. Experimental results indicate that our selective CS architecture can gain more than 50% energy savings, yet keeping the high accuracy as state-of-the-art CS architectures.