Question about the method for the cache branch

[dilint]

For the three formulas in Section 3:

I have three questions:

1. Does $Y_{train}^I \prime\prime=Y_{train}^I$ ? or what's the formula of $Y_{train}^I \prime\prime$?
2. Are the F train features normalized? For $f_{train} \in F_{train}$ and $f_{train}F_{train}^T$, the number of $f_{train}f_{train}$ will much larger than $f_{train}f_{train} \prime$, $f_{train} \prime \in F_{train}-f_{train}$.
3. In the part Few-shot Knowledge Retrieval of Figure2, there is $\phi(\cdot)$. Is it conflict with $f_{train}F_{train}Y_{train}^I\prime \prime$?

And I am confused about what network and training protocol is used to realise Fully Supervised method?

I appreciate your time and the contributions your research makes to the field. I look forward to your response and am eager to learn from your insights.