Writer: Yongshin Kim

Definition

• Single task
  • $input\;X \rightarrow output \;Y$
• Multi task
  • $input\;X\rightarrow output\;Y_1,\;Y_2,\;Y_3 \;...$
  • ex) Object detection: detection(classification) + location(regression)
  • ex) Korean $\rightarrow$ Chinese, English, Japanese
    • Korean $\rightarrow$ Chinese, Korean $\rightarrow$ English, Korean $\rightarrow$ Japanese

Advantage

• Knowledge transfer
  • Useful information from learning $Y_1$ is delivered to other $Y_2$ and $Y_3$.
• Prevent overfitting
  • Learning a more generalized representation to fit multiple tasks simultaneously.
• Computational efficiency
• Real-world application

Disadvantage

• Negative transfer
  • $Y$ may exist that adversely affects other tasks.
• Task unbalancing
  • If the learning difficulty varies greatly from task to task, the robustness of the model decreases.