Learning from dynamic and large volumes of data is expected to bring significant science and engineering advances along with consequent improvements in quality of human life, and robustness to tactical networks. The overarching goal of this tutorial is a unifying framework encompassing resilient and scalable interactive learning and learning-aided decision making through contemporary machine learning and optimization techniques.
Exposition will start with robust active learning on graphs that is featured with performance comparable to state-of-the-art deep neural networks yet at reduced complexity. Robustness to adversaries will be monitored and accounted for via moment-based blind ensemble learning. A host of scalable function approximation schemes will be presented for online (and possibly interactive) machine learning tasks. Those will leverage random features to cope with the `curse of dimensionality’ associated with online (multi-) kernel learning, and thus obtain the sought nonlinear function ‘on the fly’ with quantifiable performance, even in adversarial environments with unknown dynamics. Building on this robust and scalable function approximation framework, a privacy-preserving feature- and graph-aware learning approach will be also developed even for dynamically changing topologies. To further boost interactive learning, the last part will highlight risk-cognizant bandit and reinforcement learning approaches empowered with novel scalable function approximation modules. Impact of the unified framework will be demonstrated through extensive tests.