This dissertation delves into understanding, characterizing, and addressing dataset shift in deep learning, a pervasive issue for deployed machine learning systems. Integral aspects of the problem are examined: We start with the use of counterfactual explanations in order to characterize the behavior of deep reinforcement learning agents in visual input...
Although deep reinforcement learning agents have produced impressive results in many domains, their decision making is difficult to explain to humans. To address this problem, past work has mainly focused on explaining why an action was chosen in a given state. A different type of explanation that is useful is...
As bipedal robots move ever closer to being integrated into all manner of real world envi-ronments there is a necessity to push their dynamic capabilities to meet or exceed those of humans and animals. Advancements must be made to address ordinary challenges that arise everyday in the same environments that...
Robotic Bipedal locomotion holds the potential for efficient, robust traversal of difficult terrain. The difficulty lies in the dynamics of locomotion which complicate control and motion planning. Bipedal locomotion dynamics are dimensionally large problems, extremely nonlinear, and operate on the limits of actuator capabilities, which limit the performance of generic...
Pardoxes in voting has been an interest of voting theorists since the 1800's when Condorcet demonstrated the key example of a voting paradox: voters with individually transitive rankings produce an election outcome which is not transitive. With Arrow's Impossibility Theorem, the hope of finding a fair voting method which accurately...
Bayesian Optimization (BO) methods are often used to optimize an unknown function f(•) that is costly to evaluate. They typically work in an iterative manner. In each iteration, given a set of observation points, BO algorithms select k ≥ 1 points to be evaluated. The results of those points are...
A large number of sequential decision-making problems in uncertain environments
can be modeled as Markov Decision Processes (MDPs). In such settings, an agent
can observe at each time step the state of the environment and then executes an
action, causing a stochastic transition to a new state of the environment...
Machine learning systems are generally trained offline using ground truth data that has been labeled by experts. However, these batch training methods are not a good fit for many applications, especially in the cases where complete ground truth data is not available for offline training. In addition, batch methods do...
Data variations are prevalent in real-world applications. For example, software vendors have to handle numerous variations in the business requirements, conventions, and environmental settings of a software product. In database-backed software, the database of each version may have a different schema and content. As another example, data scientists often need...
The ability to extract uncertainties from predictions is crucial for the adoption of deep learning systems to safety-critical applications. Uncertainty estimates can be used as a failure signal, which is necessary for automating complex tasks where safety is a concern. Furthermore, current deep learning systems do not provide uncertainty estimates,...