Figure 8:

Label $\left\{\mathcal{D},{\mathcal{D}}_{\text{R}},{\mathcal{D}}_{\text{NR}}\right\}$ indicates which dataset was used to train the linear readout function. Improved performance of IFM on standard data $\mathcal{D}$ can be attributed to improved representation of robust features ${\mathcal{D}}_{\text{R}}$. See Sec. 4.3 for construction of robust $\left({\mathcal{D}}_{\text{R}}\right)$ and non-robust $\left({\mathcal{D}}_{\text{NR}}\right)$ feature datasets.