| About this Abstract |
| Meeting |
MS&T23: Materials Science & Technology
|
| Symposium
|
Materials Informatics for Images and Multi-dimensional Datasets
|
| Presentation Title |
The Conundrum of Ambiguous Feature Sets in Materials Informatics for Images |
| Author(s) |
Kevin Field, Gabriella Bruno, Matthew J. Lynch, Ryan Jacobs, Dane D. Morgan |
| On-Site Speaker (Planned) |
Kevin Field |
| Abstract Scope |
Machine learning (ML) tools are now being widely adopted for microscopy image data analysis. These tools are evaluated against human labelled datasets using metrics such as F1. Our research shows these datasets have large bias when the feature sets are ambiguous. This suggests that F1 scores are also biased and may not provide accurate evaluation on a model’s ability to quantify a materials’ response. We will present the assessment of biases using a crowd-sourcing human labelling workflow through synthetically generated images to systematically evaluate factors such as background, contrast, and feature-to-image size variances and so on. The synthetic images are generated using a physics-based simulation technique enabling controlled bias evaluations. This work will present on the most recent findings based on this workflow and its impact on how ML models perform within object detection tasks when trained with controlled bias based on the quantitative results from the crowd-sourced based experiments. |