Fig. 1:
CHEER: (a) a rich model was first built using rich multimodal or multi-channel data; (b) the behaviors of rich model are then infused into the poor model using paired data (i.e., behavior infusion); and (c) the poor model is trained to fit both the rich model’s predictions on paired data and its poor dataset (i.e., target infusion).
