Neuromorphic Decoders: Paving the way towards adaptive, low-power and low-latency Brain Computer Interfaces.Rasetto, Marco (2024) Neuromorphic Decoders: Paving the way towards adaptive, low-power and low-latency Brain Computer Interfaces. Doctoral Dissertation, University of Pittsburgh. (Unpublished)
AbstractIntracortical Brain-Computer Interfaces (iBCIs) intercept neuronal signals, allowing paralyzed individuals to perform movements and regain daily function. However, these advancements are mostly confined to laboratory settings due to high power consumption and bandwidth requirements for communication and computation in decoders, limiting their portability. Decoders are often trained offline, requiring significant memory to store neural recordings, and are typically implemented on standard hardware architectures, which increases latency and power consumption. This research aims to develop mobile BCIs with low-power, low-latency decoders that can be used outside labs or hospitals. Neuromorphic decoders present a promising solution by addressing power and latency constraints. These architectures process spiking data from iBCIs directly, eliminating the need for binning or spike counting and reducing latency and power consumption. Hierarchy of Time-Surfaces (HOTS) is particularly promising for BCI applications, using clustering to analyze data patterns, potentially making HOTS more interpretable than other machine-learning techniques. Online clustering may offer solutions to continual and incremental learning challenges, allowing BCIs to adapt to shifts in neural activity and new tasks without recalibrating or retraining decoders. However, HOTS presents some challenges: It requires exponential decay kernels that are difficult to implement efficiently on digital hardware, and it shows lower accuracy compared to backpropagation-based spiking neural networks. Additionally, HOTS's current learning rule does not support continual and incremental learning. This thesis addresses these issues. For hardware, it explores using electrochemical (ECRAM) memristor dynamics to implement exponential decay in HOTS decoders, potentially reducing circuit complexity and energy consumption. For software, it proposes Sup3r, a learning algorithm that improves accuracy and skips uninformative events, enhancing efficiency and stability in online learning. Sup3r demonstrates continual and incremental learning, making it a fundamental advancement for HOTS models, applicable beyond BCI to various fields. The hope is that these combined solutions will pave the way for low-power, adaptive neuromorphic decoders, enabling patients to regain autonomy outside the laboratory setting. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}