| For the past few years, flexible, active, multiplexed recording devices for high resolution recording over large, clinically relevant areas in the brain have been manufactured. While this technology has enabled a much higher-resolution view of the electrical activity of the brain, the analytical methods to process, categorize and respond to the huge volumes of data produced by these devices have not yet been developed.;In the first part of this thesis, we propose an unsupervised learning framework for spike analysis, which reveals typical spike patterns of the microECoG data. And we further explore using these patterns for seizure prediction and spike wavefront prediction. These methods have been applied to in-vivo feline seizure recordings and yielded promising results.;Being able to predict the neural signal in the near future from the current and previous observations has the potential to enable real-time responsive brain stimulation to suppress seizures. We explore two approaches for neural activity prediction using recurrent nerual networks. Firstly, to exploit the multiple activtiy pattern clusters present in the signal, a multiple choice learning model is proposed. An ensemble-awareness loss is used to jointly solve the model assignment problem and error minimization problem. Secondly, to ensure an accurate prediction for a long time horizon, two multi-resolution networks are proposed. To overcome the blurring effect associated with video prediction in the pixel domain using standard mean square error (MSE) loss, energy based adversarial training is used to improve the long-term prediction.;We also looked at the problem of object tracking at the pixel level. We present a novel pixel-wise visual object tracking framework that can track any anonymous object in a noisy background. The framework consists of two submodels, a global attention model and a local segmentation model. The global model generates a region of interests (ROI) that the object may lie in the new frame based on the past object segmentation maps; while the local model segments ROI to detect pixels belonging to the object. Once the models are trained, there is no need to refine them to track specific objects, making this method ecient compared to online learning approaches. |
