MTL-ViT: A new multi-task learning framework using Vision Transformers [IEEE ICIP 2024]

June 19, 2024 ยท View on GitHub

(*Note: This is an ongoing project, hence the full code and strategy is not yet open-sourced by the author.)

We presnet a new multi-task learning strategy using Vision transformers (ViTs). Our approach is based on exploiting the class-token and self-attention mechanism of Vision Transformers in order to train multiple tasks through a single ViT, more efficiently and with limited computational budget.

alt text

Total Loss of the Multi-task system: