Dictionary Construction for Accurate and Low-Cost Subspace Learning in Unsupervised Spike Sorting
Majid Zamani, Salinna Abdullah, Andreas Demosthenous · International Journal of Simulation Systems Science & Technology · 2020
This paper discusses and outlines the construction of highly reliable and power efficient dictionaries as the main block in unsupervised feature learning from evolving sub-spaces. Three types of dictionaries are considered in this paper for unsupervised subspace learning including Hadamard φ_(H_h (k) ), equiangular tight frame φ_ETF(k) and random Bernoulli φ_Bern(k) . The constructed dictionaries are then utilized in unsupervised feature learning algorithm and the classification results are investigated using a library-based neural simulator consists of various noise levels and 300 different average spike shapes. The proposed dictionaries obtain high performance with classification error of around 7% over 100 windows of generated data using the developed neural signals for 3 to 6 clusters and noise levels σ_N between 0.05 and 0.3. In summary, the combination of constructed dictionaries and subspace learning present a new class of implantable feature extractors robust to extreme signal variations and well-suited for hardware implementation.