A 73.8K Inference/mJ SVM Learning Accelerator for Brain Pattern Recognition

Sensing brain dynamics for neural disease control and brain-computer interface has become feasible through brain pattern recognition. The applications include seizure detection/prediction for epilepsy control, neural signal sensing for neuroprostheses, and brainwave processing for motor-assistive devices [1]. Fig. 1 shows the workflow for brain activity recognition, in which the acquired neural signals are sent to a machine learning (ML) kernel for pattern recognition. The outputs could be used to drive a stimulator or an actuator for control/interfacing. Of the ML kernels, support vector machine (SVM) has been widely used for neural signal classification [2]–[3] due to the effectiveness given the limited training samples. Online SVM adaptation (through training) also becomes essential to combat time-varying neural signals [2]. As signal dimension increases with the electrode array size for more sophisticated tasks [4], conventional SVM becomes inefficient due to the inherent sequential update characteristic. In this work, cluster-partitioning (CP)-SVM [5] is adopted to reduce the latency by processing data in parallel. Energy dissipation is another issue for energy-constrained devices. This work presents an energy-efficient SVM learning accelerator that supports a high throughput for neural sensing devices with a massive electrode array.