University of Cambridge > > Machine Learning @ CUED > Efficiency and Transferability of Neural Networks

Efficiency and Transferability of Neural Networks

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Robert Peharz.

Neural Networks are expensive tools for doing very specific tasks. The computational costs of neural network use hinder deployment on embedded devices for time-sensitive tasks, and the costs of learning are substantial in all settings. At the same time, neural network learning is overly dependent on large labelled training dataset that directly matches the test time task, and by default transfers poorly to slight perturbations of task.

In this talk I will explore a number of investigations in the efficiency and transferability of neural networks; we demonstrate that pruning is typically ineffective, but structured replacement can work well. We show that full pipeline optimization to hardware can provide substantial benefit. At the same time we explore how meta learning approaches can be adapted for transductive learning, by learning to learn unsupervised loss functions.

This talk is part of the Machine Learning @ CUED series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2024, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity