We aim to reduce the energy consumption of computationally complex AI algorithms for biomedical applications while delivering real-time performance. By bridging between algorithms and hardware (including circuits, architectures, and systems), we develop efficient VLSI architecture that can be applied to a broad range of biomedical applications including wearable EEG analysis systems, brain-machine-interface, visual prosthesis, etc. We utilized neural network search to find the best network for seizure prediction applications [TBME], and developed hardware friendly neural network models such as CNN [AICAS], BNN [ISCAS], Bio-inspired SNN [ISCAS] and CNN-LSTM hybrid network [AICAS] to classify signals with different characteristics. Not like conventional research that only run AI algorithms on computer, our algorithm can run on a power and resource constrained devices. To make models more general and personalize for patents, we developed dedicated Generative Adversarial Networks to argument limited personal biomedical data for training and avoiding overfitting . Combine with latest event-driven sensors, our spike processing algorithm is now able to mimic the missing functionality of AMD and RP patients.
Figure 5. CNN-LSTM hybrid neural network for seizure signal classification
Figure 6. Seizure signal argumentation with GAN Networks