Periodo di svolgimento delle lezioni: da aprile a maggio 2026
Anno di corso: a scelta tra primo o secondo anno
Durata: 12 ore
Crediti dottorali:2,4 CD
Verifica finale: sì
Deep Learning models and optimization algorithms are not designed to learn continuously, and they struggle to generalize and adapt to new circumstances and environments. To achieve continual learning capabilities, an intelligent agent must contextualize past knowledge while acquiring new skills and improving existing behaviors in response to novel scenarios. However, when faced with new concepts to learn or drifts in data distribution, a model trained with classic gradient-based techniques forgets all previously accumulated knowledge, leading to “catastrophic forgetting”. The concept of learning continually from experiences has long been explored in AI and robotics. Research in this direction has been referred to by different names, such as Lifelong Learning and Continual Learning, with the latter being associated with most modern approaches. This course introduces Continual Learning, covering its key challenges, main approaches, reference benchmarks, and evaluation metrics. A hands-on part will guide students through Avalanche, the main research framework in this area.