Seminars

View all Seminars  |  Download ICal for this event

Lifelong Learning of Representations with Provable Guarantees

Series: Bangalore Theory Seminars

Speaker: Santosh Vempala Georgia Tech

Date/Time: Dec 09 14:30:00

Location: CSA Seminar Hall (Room No. 254, First Floor)

Abstract:
In lifelong learning, tasks (or classes) to be learned arrive sequentially over time in arbitrary order. During training, knowledge from previous tasks can be captured and transferred to subsequent ones to improve sample efficiency. We consider the setting where all target tasks can be represented in the span of a small number of unknown linear (or nonlinear) features of the input data and propose a lifelong learning algorithm that maintains and refines the internal feature representation. We prove that for any desired accuracy on all tasks, the dimension of the representation remains close to that of the underlying representation. The resulting algorithm is provably efficient and the sample complexity for input dimension d, m tasks with k total features up to error ϵ is O~((dk^1.5 + km)/ϵ). We also prove a matching lower bound for any lifelong learning algorithm that uses a single task learner as a black box. An empirical study, with a lifelong learning heuristic for deep neural networks, performs favorably on challenging image datasets compared to state-of-the-art continual learning methods.


Speaker Website Link


Microsoft teams link:

Link


Hosts: Aditya Abhay Lonkar, Rahul Madhavan, Rameesh Paul & Aditya Subramanian