Problem ● Problem: lack of generic toolbox for designing neural signal processor for bio-medical applications ● Various techniques and algorithms for filtering: ● Wavelet Transform, Short-Timr Fourier Transform, Kalman, Independent Component Analysis, Principle Component Analysis, etc. ● Support multiple sensory data in biomedical applications: EMG, ECG, EOG, EEG, etc. ● Different classifiers are available: ● Neural networks, Supper Vector Machine, KNN, etc. difficult to evaluate both from the implementation and suitability perspectives! ● What kind of embedded computing system to choose?
Selebriti Toolbox for customizing the classifier with multiple constraints ● Generating the classifier via multi-objective evalutionary methods ● Ensures the accuracy level ● Satisfies the performance and energy constraints of the interface ● The target classifer will be spiking neural network – Mimicing mammallian neural system – Ultra low power -> combining with ReRAM tech -> extereme low power ● But can generate MLP and SVM for evaluation and more anaylses ● Self-learning Method (to learn new activities and actions) ● Able to equip the classifer with unsupervised techniques like SOM and STDP ● Light weight reinforcement learning techniques ● Suggest possible embedded hardware platforms via the design space exploration (DSE) technique