Collaborative Machine Learning via Model Fusion
In-Person Registration: https://forms.gle/8eQThJNsrdKkvQPv7
Virtual Registration: http://go.hawaii.edu/VCh
The classic machine learning (ML) paradigm involves algorithms that learn from centralized data, possibly pooled together from multiple data sources. The computations involved may be done on a single machine or farmed out to a cluster of machines but are inherently restricted within the availability of in-house resources. As a result, such algorithms often fail to utilize the computation and communication capability of their entire ecosystem to support their ever-growing scales.
Federated Learning (FL) addresses this pitfall by devising algorithms that allow multiple parties to collaboratively train a model with their private data sources, but without having to centralize them. These algorithms nonetheless require participants to synchronize their training processes to communicate and aggregate local parameter gradients. In practice, however, there are cases when one only has access to model artifacts that were pre-trained separately, which makes FL inapplicable. Such situations often arise in a model marketplace (e.g., Amazon’s AWS marketplace) where it is not practical nor feasible to request customers to synchronize their model training.
To overcome this challenge and enable broad collaboration among participants (who face similar analytic tasks in their businesses) in the model marketplace, I propose to explore and develop an alternative post-training fusion approach that can synthesize new models from existing model artifacts in an asynchronous manner that does not require participants to coordinate their model training. This is achieved via learning a generative mechanism that captures the latent knowledge patterns within these model artifacts expressed in terms of their trained weights and/or inferential behaviors on unlabeled data. In this talk, I will present and discuss two promising solution techniques that can fuse pre-trained neural networks, and more broadly, black-box models. To this end, I will also highlight current limitations of model fusion techniques as well as potential directions for future development.
Nghia received the Ph.D. in Computer Science from National University of Singapore (NUS) in 2015. From 2015 to 2017, he was a Research Fellow at NUS. After NUS, Nghia did another postdoc at the Laboratory for Information and Decision Systems (LIDS), MIT (2017-2018). From 2018-2020, he was a Research Staff Member and Principal Investigator at the MIT-IBM Watson AI Lab in Cambridge, Massachusetts. In Nov 2020, Nghia joined the AWS AI Labs of Amazon in Santa Clara, California as a senior research scientist. His research interests span the broad areas of deep generative modeling with applications to (personalized) federated learning, meta learning, black-box model reprogramming/reconfiguration.
Dr. Nghia Hoang