In the past decade, deep learning models have made it possible to analyze and understand image datasets made up of thousands to millions of images. This has resulted in many new interesting research problems, such as the prediction of neighbourhood beauty, safety, and liveability. However, these datasets may be the subject of biases. For instance, neighbourhoods may be perceived as safer or more beautiful when a human reviewer only receives sunny images. If we train a model on such a dataset without accounting for these biases, we end up with a model that unfairly judges neighbourhoods by the weather conditions present in its images. Meanwhile, deep learning models are increasingly consulted by decision makers to guide policy-making. As such, it is imperative that models are not influenced by subtle biases, as they might disadvantage communities. As we increasingly make use of such datasets, we are interested in discovering biases, understanding their contributions to a research problem, and to develop methods which can remove their influence.
It is well-known that large-scale image datasets are plagued by issues, including labelling errors, lack of diversity amongst labellers, and biases, such as the level of saturation and weather effects that are visible in an image. While labelling errors are bothersome as they affect the accuracy of a model, biases and non-representative labellers are particularly troubling as they skew the model towards a biased world-view. Any decision based on such flawed models will inherit these biases, making the de-biasing of datasets and/or models an important research direction. In this research we are therefore interested in understanding and accounting for the issue of biases in image datasets. In particular, we are interested in de-biasing models trained on image datasets depicting urban areas.
The student will explore various out-of-the-box debiasing methods to test how well they can account for biases in a variety of datasets. The student may also create and test their own methods if they are interested in doing so. The student can make use of a pre-trained model for weather conditions in images, as well as various models which measure urban qualities.
- Inventorize the availability of de-biasing methods that are applicable to deep learning models
- Apply de-biasing methods and observe how they affect the prediction task
- Time-permitting, perform a study with volunteers to determine whether the de-biased models are considered more fair
- Deep Learning the City : Quantifying Urban Perception At A Global Scale (Dubey et al, 2016) - this paper will help you get a feeling for urban image datasets
- Transient Attributes for high-level understanding and editing of outdoor scenes (Laffont et al, 2014) - A dataset containing weather-related images, as inspiration for biases
- Debiasing Convolutional Neural Networks via Meta Orthogonalization (David, Liu, and Fong, 2021) - Example of a possible debiasing method applicable to neural networks
- Completion of FTE-35306 Machine Learning, GRS-34806 Deep Learning, or an equivalent background in machine/deep learning is a requirement.
Theme(s): Sensing & measuring; Modelling & visualisation