Heterogeneous Data, Domain Adaptation, and Continual Learning – Orcun Goksel

In deep learning, data is everything, and in particular, well-curated and uniform data sources. In the wild, however, data comes in all types, and often in an incomplete, fractionated, and heterogeneous format. Curating unified data subsets then often has to omit large amounts of otherwise available data. I will start by presenting this problem and one approach based on attention that we developed for addressing this in semantic image segmentation from multiple input channels: https://doi.org/10.1038/s42256-021-00379-y


In the second part, I will motivate continual knowledge accumulation for both domain and class incremental settings, and argue its value in the future of deep learning. I will exemplify these with image classification and segmentation applications using ideas and results from multiple works, including: https://doi.org/10.1016/j.imavis.2024.105187 https://doi.org/10.1109/TMI.2024.3368365 https://doi.org/10.1016/j.media.2023.102924

About Orcun Goksel