Paul Janson
Paul Janson
Home
Publications
Contact
Light
Dark
Automatic
continual learning
Beyond Cosine Decay: On the effectiveness of Infinite Learning Rate Schedule for Continual Pre-training
We demonstrate that infinite learning rate schedules consistently outperform widely-used repeated cosine decay for continual pre-training under distribution shifts across both vision and language models, providing a more effective alternative for large-scale self-supervised learning without catastrophic forgetting.
Paul Janson
,
Vaibhav Singh
,
Paria Mehrbod
,
Adam Ibrahim
,
Irina Rish
,
Eugene Belilovsky
,
Benjamin Therien
PDF
Cite
Code
Post
Continual zero-shot learning through semantically guided generative random walk
Learning novel concepts, remembering previous knowledge, and adapting it to future tasks occur simultaneously throughout a human’s …
Wenxuan Zhang
,
Paul Janson
,
Divyansh Jha
,
Kai Yi
,
Ivan Skorodov
,
Mohammed Elhoseiny
PDF
Cite
Code
Overcoming Generic Knowledge Loss with Selective Parameter Update
Adding knowledge to the model without destroying its generalization by finetuning small set of parameters
Wenxuan Zhang
,
Paul Janson
,
Rahaf Aljundi
,
Mohammed Elhoseiny
PDF
Cite
Post
Domain Aware Zero shot learning
Continual zero-shot learning involves learning seen classes incrementally while improving the ability to recognize unseen or …
Kai Yi
,
Paul Janson
,
Wenxuan Zhang
,
Mohammed Elhoseiny
PDF
Cite
A Simple baseline that questions the use of pre-trained model in continual learning
A baseline that performs better without training in continual learning benchmarks
Paul Janson
,
Wenxuan Zhang
,
Rahaf Aljundi
,
Mohammed Elhoseiny
PDF
Cite
Code
Cite
×