Learning Over Time
Sommario
On September 25th, 2025 at 2.00 pm Lorenzo Iovine, PHD Student in Information Technology, will hold a seminar on "Learning Over Time" at DEIB Seminar Room "Alessandra Alario" (Building 21).
In this seminar, I will present and discuss two key contributions showcased during the LOT Spring School 2025. The first, by Tinne Tuytelaars, explores Continual Learning as a crucial paradigm to overcome the limitations of static training in dynamic environments, emphasizing strategies for adaptation over time.
The second, by Alexei Efros, introduces Test-Time Training (TTT) as a powerful method to adapt models at inference time using unlabeled test data, enabling robust performance in out-of-distribution (OOD) scenarios.
Building upon these concepts, I will briefly introduce my recent work on satellite image classification under temporal distribution shifts. In this work, we propose a pipeline that combines embedding methods with Streaming Machine Learning (SML) to effectively mitigate OOD issues and adapt to changing data distributions over time. Our approach aligns with the motivations behind both Continual Learning and Test-Time Training.
In this seminar, I will present and discuss two key contributions showcased during the LOT Spring School 2025. The first, by Tinne Tuytelaars, explores Continual Learning as a crucial paradigm to overcome the limitations of static training in dynamic environments, emphasizing strategies for adaptation over time.
The second, by Alexei Efros, introduces Test-Time Training (TTT) as a powerful method to adapt models at inference time using unlabeled test data, enabling robust performance in out-of-distribution (OOD) scenarios.
Building upon these concepts, I will briefly introduce my recent work on satellite image classification under temporal distribution shifts. In this work, we propose a pipeline that combines embedding methods with Streaming Machine Learning (SML) to effectively mitigate OOD issues and adapt to changing data distributions over time. Our approach aligns with the motivations behind both Continual Learning and Test-Time Training.
