Algorithm 1.

CHEER $\left({𝓗}_p,𝓣\left(y\mid {\mathbf{x}}^r\right),{𝓗}_o\right)$

1:	Input: rich model $𝓣\left(y\mid {\mathbf{x}}^r\right)$, poor data ${𝓗}_p$ and paired data ${𝓗}_o$
2:	Infuse rich model’s behavior via ${𝓗}_o$
3:	i ← 1
4:	while i ≤ d do
5:	${\mathit{\omega }}_p^{(i)}\leftarrow \text{argmin}{𝓛}_i\left({\mathit{\omega }}_p^{(i)}\right)$ via (8);
6:	$a_p^{(i)}\left({\mathbf{x}}^p\right)\leftarrow a_p^{(i)}\left({\mathbf{x}}^p;{\omega }_p^{(i)}\right)$
7:	$i\leftarrow i+1$
8:	end while
9:	${\mathbf{A}}_p\leftarrow \left[a_p^{(1)}\left({\mathbf{x}}^p;{\mathit{\omega }}_p^{(1)}\right)\dots a_p^{(d)}\left({\mathbf{x}}^p;{\mathit{\omega }}_p^{(d)}\right)\right]$
10:	Infuse rich model’s target via $\left({𝓗}_o,{𝓗}_p\right)$ and Ap
11:	$\left({\mathbf{Q}}_p,{\mathbf{O}}_p\right)\leftarrow \text{argmin}{𝓛}_p$ via (9)
12:	Output: poor model $𝓢\left(y\mid {\mathbf{x}}^p\right)\leftarrow \left({\mathbf{A}}_p,{\mathbf{Q}}_p,{\mathbf{O}}_p\right)$