Pitfalls with ablation in neural network architectures

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

• Christina Lioma (University of Copenhagen)
• Friday 17 June 2022, 12:00-13:00
• Computer Lab, SS03.

If you have a question about this talk, please contact Michael Schlichtkrull.

Abstract:

Ablation tests are frequently used for analysing machine learning performance. Generally, ablation refers to the removal of a component in order to understand its contribution to the overall decision-making process. For instance, it is expected that the ablation of a feature during classification will affect the output somehow analogously to the importance of that feature for the classification task at hand. Ablation is therefore routinely used to attribute feature importance, as well as to explain machine learning output in a partial, approximate, and model-agnostic way.

This talk will point out a core problem when using ablation with neural network architectures. The problem stems from the tendency of neural network architectures to ignore complex predictive features in the presence of few simple predictive features, even when the complex features have significantly greater predictive power than the simple features. This talk will provide evidence demonstrating the existence of this tendency, even in small neural network architectures, and show how this may entirely invalidate the standard interpretation of ablation tests. A discussion about why this is important and ways of moving forward will be provided.

This is joint work with Qiuchi Li (University of Copenhagen).

Bio:

Christina Lioma is a professor in machine learning at the Department of Computer Science, University of Copenhagen. Her research focuses on applied machine learning, information retrieval and web search technologies, web data mining and analytics, recommendation systems and natural language processing. She has a track-record of research collaboration with Danish and international industry, and an alumni of >20 PhD students and postdocs. Since 2012, Christina Lioma has attracted more than 50 million USD in external funding.

This talk is part of the NLIP Seminar Series series.

This talk is included in these lists:

• All Talks (aka the CURE list)
• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge Forum of Science and Humanities
• Cambridge Language Sciences
• Cambridge talks
• Chris Davis' list
• Computer Lab, SS03
• Department of Computer Science and Technology talks and seminars
• Graduate-Seminars
• Guy Emerson's list
• Interested Talks
• Language Sciences for Graduate Students
• NLIP Seminar Series
• PMRFPS's
• Simon Baker's List
• Trust & Technology Initiative - interesting events
• bld31
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.