Our Motivation

Motivation: About

Traditional Vision System

Traditional vision systems use well-established (hand-crafted) feature descriptors (SIFT, SURF, BRIEF, etc.) for down-stream tasks such as image classification, object detection and 3D vision.

Motivation: Text

Data-Driven Vision Approach

Data-driven vision (Deep Learning) approaches entirely rely on the images and labels to find the underlying patterns in an end-to-end learning approach.

Performance: Deep Learning >> Traditional Vision Approach

Motivation: Text

However, the performance of the DL approach entirely relies on the distribution of the data!
Therefore, if the data has group shifts, the model would cause serious discrimination over the minor group! In other words, the model could be overfitting towards the major groups!

Motivation: Text

Below is an example of a group-shifted dataset, the Celebrity Faces dataset.
The label denotes the class categories. In Celebrity Faces, it is male and female.
Attributes denote the features within each sample which are not characterized in label space. In Celebrity Faces, it could correspond to the hair color: black or blonde.
The groups, or we call Uncertainty Set, are the combinations of Labels and Attributes. In Celebrity Faces, it corresponds to four groups of images.
Clearly, males with Blonde hair would be the minor group in this case.

Motivation: Text

In fact, group shifts, (or we call Bias), are quite common in ML
Here are the classification results under the Celebrity Faces dataset. It is obvious that the Minor group is suffering from the dominant impact of major groups, thus receiving low accuracy!

Motivation: Text

Therefore, it is our goal to improve the worst-case group classification accuracy while preserving the major groups' accuracy. On the next page, we'll introduce some state-of-art methods and their limitations.

Motivation: Text

Our Motivation

Traditional Vision System

Data-Driven Vision Approach

However, the performance of the DL approach entirely relies on the distribution of the data!Therefore, if the data has group shifts, the model would cause serious discrimination over the minor group! In other words, the model could be overfitting towards the major groups!

In fact, group shifts, (or we call Bias), are quite common in ML Here are the classification results under the Celebrity Faces dataset. It is obvious that the Minor group is suffering from the dominant impact of major groups, thus receiving low accuracy!

Therefore, it is our goal to improve the worst-case group classification accuracy while preserving the major groups' accuracy. On the next page, we'll introduce some state-of-art methods and their limitations.

However, the performance of the DL approach entirely relies on the distribution of the data!
Therefore, if the data has group shifts, the model would cause serious discrimination over the minor group! In other words, the model could be overfitting towards the major groups!

In fact, group shifts, (or we call Bias), are quite common in ML
Here are the classification results under the Celebrity Faces dataset. It is obvious that the Minor group is suffering from the dominant impact of major groups, thus receiving low accuracy!