On-chip machine learner for spike sorting in implantable brain machine interfaces (BMI)

Advances in neuroscience have enabled the rapid development of electronics that abet the functioning of prosthetic limbs. Multi-electrode arrays (MEAs) have been successfully implanted in the brain, and the resultant neural signals or ‘Action Potentials’ have been amplified and recorded to be processed, thus forming a ‘brain-to-electronic and electronic-to-prosthetic’ interface. Neural amplifiers have greatly improved over the years to address the issues of high (and thus hazardous) power dissipation to the brain tissue and high background recording noise. This has allowed for MEAs to reach the order of one-thousand or more. Modern-day neural recordings taken by an implanted MEA consist of a single probe recording multiple neuron activity, since studying cells in isolation does not illustrate the big picture [1]. Each spike signal is characteristic to a certain neuron, and it is imperative to accurately identify spike signals with their corresponding neurons. This issue known as ‘Spike Sorting’ consists of a two-step process: Feature Extraction and Clustering. The former formidably reduces the dimensionality of the Spike data, while retaining its salient features that can help differentiate between the various neuron groups. The latter implements algorithms on the extracted features to classify the data into the corresponding neuron groups. The motivation behind this research is to propose novel bio-inspired feature extraction for the purpose of spike sorting. These features are based on the output of an Integrate-and-Fire neuron model, representing a formal spiky neuron model. First, a thorough literature review was conducted, focusing on the best features available. Research papers claiming that derivative-based features performed better than the well-established PCA in terms of noise and error were studied and the results were successfully matched with ours. Next, a complete spike sorting circuit was algorithmically modelled using MATLAB. Features based on the CCO outputs tested successfully using k-means classification. The Gain of the Integrate-and-Fire was 6.465 x 1012 Hz/A. Lastly, a circuit-level implementation of the same was repeated on CADENCE. The Gain was a comparable 4.4 x 1012 Hz/A, thus successfully proving that results from the functional model matched those of the real-time simulated model, for the same set of parameters and settings.Bachelor of Engineerin