Place: Chamberlin 5280 (Zoom link for those attending online: )
Speaker: Nicholas Roberts, University of Wisconsin-Madison
Abstract: The underlying motivation of automated machine learning, or AutoML, is to automate away tasks which require machine learning (ML) expertise--such that experts in domains other than ML can reap its potential benefits for their problems. These automation efforts are mostly siloed within the machine learning community by their reliance on the tasks or domains which are most familiar to machine learning experts--classification tasks in computer vision or NLP. Unfortunately, this neglects the heavy-tail of tasks and domains that practitioners might care about and directly contradicts a core premise of AutoML: usefulness to non-ML-experts. In this talk, we present two of our recent directions which make progress toward alleviating this issue by expanding into under-explored domains and problem types. In the first half of the talk, we will present a class of search spaces over deep neural network operations which can specialize a given CNN architecture to any domain of interest by generalizing the convolution theorem from signal processing. In the second half of the talk, we will discuss the limitations of weak supervision for semi-automated dataset curation and show how to generalize weak supervision so that it can be applied to any label space equipped with a distance metric, as opposed to categorical labels alone.