My primary research interests are in the areas of machine learning and artificial intelligence, with a focus on the following topics:
- Lifelong learning of multiple consecutive tasks over long time scales,
- Knowledge transfer between learning tasks, including the automatic cross-domain mapping of knowledge,
- Interactive AI methods that combine system-driven active learning with extensive user-driven control over learning and reasoning processes, and
- Big data analytics through lifelong and transfer learning, with a focus on learning multiple tasks online from streaming data and adjusting for drifting task distributions.
My research applies these methods to problems in robotics, computational sustainability, and medicine.
Lifelong Machine Learning (Related Publications)
Lifelong learning is essential for an intelligent agent
that will persist in the real world with any amount of versatility.
Animals learn to solve increasingly complex tasks by continually building
upon and refining their knowledge. Virtually every aspect of
higher-level learning and memory involves this process of continual
learning and transfer. In contrast, most current machine
learning methods take a "single-shot" approach in which knowledge is
not retained between learning problems.
This project is funded under the Office of Naval Research (ONR) Active Transfer Learning program and is joint work with Terran Lane (University of New Mexico / Google) and Paul Ruvolo (Olin College). For further details on this ongoing project, please contact me directly.
Selective Knowledge Transfer (Related Publications)
My dissertation research focused on the selective transfer of knowledge between learning tasks [Eaton, 2009]. In this context, a task is a single learning problem, such as learning to recognize a particular visual object. An agent, having faced multiple tasks, would have built up a repository of learned knowledge that could be used to improve future learning. I developed methods that automatically select the particular source knowledge to transfer in learning a new target task. Until my dissertation, the problem of source knowledge selection had received little attention, despite its importance to the development of robust transfer learning algorithms. My dissertation showed that proper source selection can produce dramatic improvements in transfer performance by identifying the knowledge that would best improve learning of the new task. This aspect can be measured by the transferability between tasks -- a measure, introduced in my dissertation, of the change in performance between learning with and without transfer.
I developed selective transfer methods based on this notion of transferability for two general scenarios: the transfer of individual training instances and the transfer of model components between tasks. For instance-based transfer, I developed the TransferBoost algorithm, which uses a novel form of set-based boosting to select individual source instances to augment a target task's training data [Eaton & desJardins, 2011; Eaton & desJardins, 2009]. For model-based transfer, my approach organizes the source tasks onto a manifold that captures the transferability between tasks [Eaton, desJardins & Lane in ECML 2008]. Using the basis functions of this manifold, it learns a transfer function that represents how components of previously learned models are best transferred between tasks, and uses this function to select the model components to transfer to a new target task.
Interactive Artificial Intelligence (Related Publications)
My research on interactive AI methods seeks to give users extensive control over reasoning and learning processes. In many critical applications, especially in military and medical domains, users will reject traditional AI automation without the ability for each result to be checked and altered by a human operator. Interactive AI methods incorporate such levels of user control to facilitate the transition of AI into these types of applications. This interactive AI paradigm combines user-driven control with the complementary system-driven approach of active learning.
In recent work, we explored interactive learning of a regression function in collaboration with the user [Eaton, Holness, & McFarlane in AAAI 2010]. The user views a scatterplot of data scored by the function and can graphically correct the score of individual instances. The system then generalizes each correction to nearby instances, as determined by a manifold underlying the data, and updates the learned function and associated scatterplot. Interactive learning based on this manifold generalizes each user adjustment to other related data in an intuitive manner that monotonically improves performance with any correction. This property of continuous improvement ensures that the system will be accepted by users, in contrast to several other learning approaches which initially overfit each correction, potentially causing users to decrease their trust in the AI components. This work was applied to an information management system used to monitor threats -- an application in which users require the ability to rapidly adjust the scoring function in response to changing requirements. This technique could also be applied to other systems used by network security analysts, stock traders, and crisis monitoring centers.
Constrained Clustering (Related Publications)
Constrained clustering uses background knowledge in the
form of must-link constraints, which specify that
instances belong in the same cluster, and cannot-link
which specify that two instances belong in different
improve the resulting clustering.
My Master's thesis work [Eaton,
focused on a method for propagating a given set
of constraints to other data instances
based on the cluster geometry, decreasing the number of
needed to achieve high performance. This method for
propagation was later used as the foundation for the first
constrained clustering method that supports an incomplete
between views [Eaton,
desJardins, & Jacob in KAIS 2012; Eaton,
desJardins, & Jacob in CIKM 2010]. In this
clustering progress in one
the data (e.g., images) is propagated via a set of pairwise
to improve learning
performance in another view (e.g., associated text
The key contribution of this work is that it supports an
mapping between views, enabling the method to be
to a larger range of applications and legacy data sets that
multiple views available for only a limited portion of the
Learning User Preferences over Sets of Objects (Related Publications)
In collaboration with Marie desJardins (UMBC) and Kiri
(NASA Jet Propulsion Laboratory), I developed the DDPref
learning and reasoning
about a user's preferences for selecting sets of objects
where items in
the set interact [desJardins,
& Eaton, 2010]. The DDPref representation captures
between items within a set, modeling the user's desired
quality and diversity of items in the set. Our approach
allows a user
to either manually specify a preference representation, or
example sets that represent their desired information, from
can learn a representation of their preferences. We applied
method to identify sets of images taken by a remote Mars
transmission back to the user. Due to the limited
bandwidth, it is important to send back a set of images
captures the user's desired information. This research is
applicable to search result set creation, automatic content
presentation, and targeted information acquisition.