Shortcut Learning in Machine Learning

Schedule

• 4:00 PM - 6:15 PM June 22, 2021 KST

• 3:00 AM - 5:15 AM June 22, 2021 EDT

• 0:00 AM - 3:10 PM June 22, 2021 PDT

Overview

Recent advances in deep learning open a new era of practical AI applications. Despite their tremendous successes, numerous case studies have reported that even large-scale ML models with billions of parameters oftentimes fail to generalize in testing environments with heterogeneous data generating processes. For example, if the model was trained to classify the "boat" and "car" from images of [boats in the water] and [cars on the road], the model may wrongly classify new images of [boat on the road] as a "car" by taking the [road] for its decision rules. This phenomenon is called the "shortcut learning bias," which happens when ML models establish their decision rules based on undesired/unintended features [Geirhos et al., 2020], which are correlated but causally irrelevant to outcomes. The shortcut learning bias is a prevalent threat to fairness, accountability, and transparency (FAccT) of ML models because their decision rules may be based on non-causally-related features such as gender or race.

In this tutorial, we will outline notions of shortcut learning and discuss methods for alleviating the issue. Specifically,

1. We will exemplify the shortcut learning bias by demonstrating the challenges that modern deep learning models encounter.

2. We will formalize the notion of shortcut learning bias through the lens of causality and invariance.

3. We will review the state-of-the-art techniques for addressing bias.

We hope our tutorial will alert the shortcut learning bias risk and encourage developing fair, accountable, and transparent ML methods for experts in the FAccT community.