PhD Thesis Defense

Neural interfaces are entering an era where what once was science fiction is becoming a reality. As neural interfaces move out of the lab and into people's lives, the stability of neural decoding algorithms becomes ever more pressing. It is an unfortunate reality that neural implants degrade from long-term exposure to the neurological environment, however prior work has shown enhanced decoding stability in the application of 1D convolutional neural networks to neural feature extraction. However, these algorithms have high memory and processing requirements, prohibiting them from meeting the low area and power restrictions of implantable brain machine interface decoding pipelines.

This dissertation addresses the difficulties of implementing these algorithms on streamed neural data with high parallelism and low area and power costs. We address the unique dataflow characteristics of the feature extraction workload by designing a tailored processing element that reduces the memory access requirements by 2x. We further reduce system memory requirements through efficient process scheduling and memory partitioning. We then address the model complexity through retraining and analysis of the effect of various system parameters on the accuracy of kinematic decoding and hardware performance.

Results show that these design choices were able to successfully implement these intensive but performant algorithms within the power and area budgets of implantable devices. The architecture supports 192 channels that achieve state-of-the-art decoding stability at 1.8 𝜇𝑊 and 12801  𝜇𝑚2  per channel in 65 nm CMOS technology. The device is a fully configurable, scalable, area and power efficient solution that supports models with 2-8 feature layers and a total kernel length of up to 256. This architecture reduces caching requirements by $5\times$ over conventional computation schemes. We show our hardware optimized models maintain superior stability over time using recorded data from tetraplegic human participants with spinal cord injury. The models and hardware were validated in real time with a human subject in online closed-loop center-out cursor control experiments with micro-electrode arrays that were implanted for 6 years. Decoders using features generated with this work substantially improve the viability of long-term neural implants compared to other feature extraction methods currently present in low-power BMI hardware.