Explore the concept of deep ensembles from a loss landscape perspective in this 35-minute Launchpad talk. Delve into the hypothesis, background research, and methodologies used to measure function similarity within and across trajectories. Examine subspace sampling techniques and analyze loss and function similarity in prediction space. Evaluate the effects of ensembling and draw conclusions from the presented findings. Engage in a Q&A session to further discuss ideas and insights related to the arxiv.org paper 1912.02757.
Overview
Syllabus
Intro
Overview
Hypothesis
Background Research
Measuring function similarity within and
Subspace sampling within and across trajectories
Loss and function similarity in prediction space
Evaluating the effects of ensembling and
Conclusion
Questions, comments, ideas
Taught by
Launchpad
