ISTI-Talk: Do You Really Need More Data? What Could You Accomplish With the Data You Already Have?
- Day - Time: 24 September 2025, h.11:00
- Place: Area della Ricerca CNR di Pisa - Room: C-29
Speakers
Referent
Abstract
Data scarcity constrains progress across scientific domains, from medical imaging with limited patient samples to environmental monitoring with sparse sensor data, and from rare event detection to expensive experimental datasets. Three universal challenges persist: severe class imbalance where minority categories are underrepresented, low data resolution that obscures critical features, and limited dataset availability due to collection costs or rare phenomena. These constraints often lead researchers to pursue costly data acquisition, but what if intelligent computational methods could amplify the value of existing datasets?
This seminar will present deep learning techniques developed to address these fundamental data challenges such as feature-space oversampling to balance class distributions, curriculum learning to optimize training stability, super-resolution networks to enhance data quality, and domain adaptation to maximize generalization. Originally developed for SAR ship classification, these methods are transferable to any data-limited domain. I will share results, discuss lessons learned, and invite feedback on future directions, particularly the role of foundation and vision-language models in enabling richer semantic reasoning under data scarcity.