adopted, while the vertical axis represents the distance when the CCRM module is not adopted. Each scatter corresponds to one sample, and the solid regression line is calculated by fitting all samples. It is clear that the slopes of the solid lines are much less than one, indicating that the proposed OLR-SACNN can extract deep features sensitive to the sample changes, i.e., the deep features extracted are closely related to the samples. 10 20 30 40 k = 3.8427 -10 Regression Line Sample -10 010203040 Norm of Deep Feature (sEMG) (a) E4 U4 0.1 U8 U12 U16 The Distance Between the sEMG Feature and the sEMG Samples 0.4 0.3 k = 0.0061 0.2 0.1 0.1 0.2 Adopting CCRM (c) Figure 9. A function analysis of different modules of the OLR-SACNN model. (a) A 2D distribution of norms of the deep features from the sEMG and the ultrasound. (b) The correlation matrix of deep features from the sEMG and the ultrasound (the horizontal axis is the simplified deep feature of the sEMG, and the vertical axis is the one of the ultrasound). (c) The distance between the original samples and their corresponding deep features. 22 * IEEE ROBOTICS & AUTOMATION MAGAZINE * DECEMBER 2022 Sample Regression Line 0.3 0.4 0.1 0.15 0.2 0.25 0.3 0.1 0.05 U8 U12 (b) The Distance Between the Ultrasound Feature and the Ultrasound Samples k = -0.1793 U16 0.05 E8 E12 E16 0.15 U4 0.1 E4 E8 E12 E16 0.15 Sample 0.15 0.2 Regression Line 0.25 Adopting CCRM 0.3 Without CCRM Norm of Deep Feature (Ultrasound) Without CCRM