Adaptive statistical algorithms for learning and control of neural dynamics Project Summary / Abstract The lack of real-time and closed-loop machine learning tools limits experimentalists from investigating learned behavior and neural dynamics. Real-time and closed-loop methods can bring about new experimental investi- gations aimed at understanding neural information processing at the system and organism levels. They allow experimentalists to monitor the animal's internal states and probe their internal dynamics to infer the animal's information processing architecture. To this end, we develop real-time adaptive algorithms for modeling nonlinear dynamical systems, feedback control strategies, and adaptive behavior training algorithms. Compared to offline data analysis, real-time and closed-loop experiments are more challenging from a statistical machine learning perspective. Typically these algorithms have many tunable knobs that cannot be changed during the experimen- tal session, and repeating the same experiment with different settings is orders of magnitude more expensive than re-analyzing data offline. To overcome this difficulty and address potential convergence speed issues, we pro- pose to exploit the commonalities across animals, recording sessions, and datasets from different tasks through hierarchical modeling and meta-learning to find the best parameters for each experiment. Although the exact set of parameters may not be similar, we can learn hyper-priors that can smooth over datasets and meta-learn to learn faster from streaming data (Aim 1). We also propose to further refine the implementations of our algorithms and branch out to more challenging complex behaviors. Specifically, we will use deep-learning in loop adaptive experimental design with the aim to (1) find the optimal stimulus to train deep neural networks to predict neural response, (2) synthesize best images to train an interpretable deep neural network model, and (3) train models using natural and perturbed behavior (Aim 2). Finally, we propose to improve the robustness and stability of the real-time data analysis pipeline to enhance the quality of the closed-loop experiments (Aim 3).