Jia, Xiaowei and Karimi, Hassan
(2023)
Preserving Fairness of Deep Learning Under Environmental Changes.
In: Pitt Momentum Fund 2023.
Abstract
Recent advances in deep learning have shown promise in discovering complex patterns from spatial datasets and greatly enhanced our ability to address some of the most critical societal problems, such as agricultural monitoring and natural disaster detection. However, direct applications of deep learning often result in spatially biased predictions due to common data variability, data quality, data volume, and a lack of spatial fairness considerations in existing deep learning techniques. This can be further exacerbated by different types of environmental changes (e.g., weather conditions, soil properties) over time. Such biases, if left unattended, may cause unfair distribution of social resources, spatial disparity, weaknesses in resilience or sustainability, etc. Existing methods for enforcing fairness do not consider long-term data distribution shifts due to environmental changes. For example, crop yield is very sensitive to soil moisture, which is highly variable over the landscape due to changes in precipitation and water table depth. As a result, a fairness-enforced model learned from historical years may fail to preserve fairness in future years. This proposal aims to develop a new generation of spatially-explicit deep learning frameworks by leveraging accumulated physical knowledge and new optimization algorithms to preserve spatial fairness under long-term environmental changes.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
Metrics
Monthly Views for the past 3 years
Plum Analytics
Altmetric.com
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}