Deep Kernel Machines
Johan Suykens - KU Leuven, ESAT-Stadius and Leuven.AI Institute
With neural networks and deep learning flexible and powerful architectures have been proposed, while with support vector machines and kernel machines solid foundations in learning theory and optimization have been achieved. In recent work on restricted kernel machines (RKM), new connections have been established between restricted Boltzmann machines (RBM), kernel principal component analysis (KPCA) and least squares support vector machines (LS-SVM). An important role for revealing the unexpected connections is played by duality principles. It enables to conceive Deep Kernel Machines for supervised and unsupervised learning, such as deep forms of KPCA and Deep RKMs. Within the framework one can either work with explicit (e.g. multi-layered, convolutional) feature maps or implicit feature maps in connection to kernel functions. New developments will be shown for generative kernel machines, multi-view and tensor based models, latent space exploration, robustness and explainability. Future perspectives and challenges will be outlined.
We consider environments where a set of human workers needs to handle a large set of tasks while interacting with human users. The arriving tasks vary: they may differ in their urgency, their difficulty and the required knowledge and time duration in which to perform them. Our goal is to decrease the number of workers, which we refer to as operators that are handling the tasks while increasing the users’ satisfaction. We present automated intelligent agents that will work together with the human operators in order to improve the overall performance of such systems and increase both operators' and users’ satisfaction. Examples include: home hospitalization environments where remote specialists will instruct and supervise treatments that are carried out at the patients' homes; operators that tele-operate autonomous vehicles when human intervention is needed and bankers that provide online service to customers. The automated agents could support the operators: the machine learning-based agent follows the operator’s work and makes recommendations, helping him interact proficiently with the users. The agents can also learn from the operators and eventually replace the operators in many of their tasks.
Deep Reinforcement Learning has been very successful in solving a variety of hard problems. But many RL architectures treat the action as coming from an unordered set or from a bounded interval. It is often the case that the actions and policies have a non-trivial structure that can be exploited for more efficient learning. This ranges from game playing settings where the same action is repeated multiple times, to supply-chain problems where the action space has a combinatorial structure, to problems that require a hierarchical decomposition to solve effectively. In this talk, I will present several scenarios in which taking advantage of the structure leads to more efficient learning. In particular, our talk about some of our recent work on action repetition, actions that are related via a graph structure, ensemble policies, and policies learnt through a combination of hierarchical planning and learning.
Overparametrized models are widely used in modern machine learning applications. However, they exist behaviors that are not explained by the traditional statistical analysis. In this talk, I will discuss some interesting questions and new theoretical analysis to understand these overparameterized models.