Deep Learning offers new possibilities for detecting subtle microscale defects in semiconductors. Yet even such a powerful tool, in its current form, has shown serious limitations when used in real world environments. Some of the major challenges that limit deployment of deep learning for visual inspection in the semiconductor industry include: 1) the rarity of defect samples and 2) the high variability of manufactured devices. Techniques that require much less (or no) supervision are the only long-term viable solution for high-throughput / high-mix precision industries.

In this context, we aim at developing computational methods and workflows that process the readily available visual streams in a largely unattended way. First, we will start by addressing the issue of interference in ML model development. Interference, also known as forgetfulness, increases the burden of model development and deployment. Second, we will also research robust methods for learning from fewer supervised samples.