My dissertation research (Eaton, 2009) focused on the problem of source selection in transfer learning: given a set of previously learned source tasks, how can we select the knowledge to transfer in order to best improve learning on a new target task? In this context, a task is a single learning problem, such as learning to recognize a particular visual object. Until my dissertation, the problem of source knowledge selection for transfer learning had received little attention, despite its importance to the development of robust transfer learning algorithms. Previous methods assumed that all source tasks were relevant, including those that provided irrelevant knowledge, which can interfere with learning through the phenomenon of negative transfer.
My results showed that proper source selection can produce large improvements in transfer performance and decrease the risk of negative transfer 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 (Eaton & desJardins, 2011) and the transfer of model components between tasks (Eaton et al., 2008). In particular, my research on model-based transfer showed that modeling the transferability relationships between tasks using a manifold provides an effective framework for source knowledge selection, providing a geometric framework for understanding how knowledge is best transferred between learning tasks.
The success of transfer learning on a target task
is highly dependent on the selected source data.
Instance transfer methods reuse data from the
source tasks to augment the training data for the
target task. If poorly chosen, this source data
may inhibit learning, resulting in negative
transfer. The current most widely used algorithm
for instance transfer, TrAdaBoost, performs poorly
when given irrelevant source data.
We present a novel task-based boosting technique
for instance transfer that selectively chooses the
source knowledge to transfer to the target task.
Our approach performs boosting at both the instance
level and the task level, assigning higher weight to
those source tasks that show positive transferability
to the target task, and adjusting the weights of
individual instances within each source task via
AdaBoost. We show that this combination of task- and
instance-level boosting significantly improves
transfer performance over existing instance transfer
algorithms when given a mix of relevant and irrelevant
source data, especially for small amounts of data on
the target task.
@inproceedings{Eaton2011Selective,author={Eaton, Eric and desJardins, Marie},title={Selective Transfer Between Learning Tasks Using Task-Based Boosting},booktitle={Proceedings of the 25th AAAI Conference on Artificial Intelligence (AAAI-11)},month=aug,location={San Francisco, CA},publisher={AAAI Press},pages={337--342},year={2011},supplement={papers/Eaton2011Selective-Supplement.pdf}}
2009
Selective Knowledge Transfer for Machine Learning
Eric
Eaton
University of Maryland, Baltimore County, Aug 2009
Knowledge transfer from previously learned tasks to a new task is a fundamental component
of human learning. Recent work has shown that knowledge transfer can also improve
machine learning, enabling more rapid learning or higher levels of performance.
Transfer allows learning algorithms to reuse knowledge from a set of previously learned
source tasks to improve learning on new target tasks. Proper selection of the source
knowledge to transfer to a given target task is critical to the success of knowledge transfer.
Poorly chosen source knowledge may reduce the effectiveness of transfer, or hinder
learning through a phenomenon known as negative transfer.
This dissertation proposes several methods for source knowledge selection that are
based on the transferability between learning tasks. Transferability is introduced as the
change in performance on a target task between learning with and without transfer. These
methods show that transferability can be used to select source knowledge for two major
types of transfer: instance-based transfer, which reuses individual data instances from the
source tasks, and model-based transfer, which transfers components of previously learned
source models.
For selective instance-based transfer, the proposed TransferBoost algorithm uses a
novel form of set-based boosting to determine the individual source instances to transfer
in learning the target task. TransferBoost reweights instances from each source task based
on their collective transferability to the target task, and then performs regular boosting to
adjust individual instance weights.
For model-based transfer, the learning tasks are organized into a directed network
based on their transfer relationships to each other. Tasks that are close in this network
have high transferability, and tasks that are far apart have low transferability. Model-based
transfer is equivalent to learning a labeling function on this network. This dissertation
proposes the novel Spectral Graph Labeling algorithm that constrains the smoothness of
the learned function using the graph’s Laplacian eigenvalues. This method is then applied
to the task transferability network to learn a transfer function that automatically determines
the model parameter values to transfer to a target task. Experiments validate the success
of these methods for selective knowledge transfer, demonstrating significantly improved
performance over existing methods.
@phdthesis{Eaton2009Selective,author={Eaton, Eric},title={Selective Knowledge Transfer for Machine Learning},school={University of Maryland, Baltimore County},year={2009},}
In this paper, we propose a novel graph-based method for
knowledge transfer. We model the transfer relationships between source
tasks by embedding the set of learned source models in a graph using
transferability as the metric. Transfer to a new problem proceeds by
mapping the problem into the graph, then learning a function on this
graph that automatically determines the parameters to transfer to the
new learning task. This method is analogous to inductive transfer along a
manifold that captures the transfer relationships between the tasks. We
demonstrate improved transfer performance using this method against
existing approaches in several real-world domains.
@inproceedings{Eaton2008Modeling,author={Eaton, Eric and desJardins, Marie and Lane, Terran},title={Modeling Transfer Relationships Between Learning Tasks for Improved Inductive Transfer},booktitle={Proceedings of the 19th European Conference on Machine Learning (ECML)},year={2008},pages={317--332},publisher={Springer-Verlag},location={Antwerp, Belgium},address={Berlin, Heidelberg},}
