- Дата: 15-07-2022, 21:43
The lack of slot range diminishment paired with a lower in slot accuracy when removing the slot range time period in SCN suggests it supplies a slight regularization profit, however little else. However, a potentially more elegant resolution is to attempt to drive slot diversity by means of architectural inductive biases as an alternative of loss targets. 2020), with a temporal contrastive loss as an alternative of a reconstruction loss. In this paper, we consider picture reconstruction and set prediction as downstream tasks to showcase the versatility of our module each in a difficult unsupervised object discovery setup and in a supervised task involving set-structured object property prediction. We display that the Slot Attention module can be utilized for supervised object property prediction, where the eye mechanism learns to highlight individual objects with out receiving direct supervision on object segmentation. We examine to a number of baselines: a randomly initialized mannequin with no coaching, CSWM (Kipf et al., 2019) and a completely supervised mannequin, the place each slot within the mannequin is educated to regress the true position of one of the objects in the scene. 2019), restricting ourselves to labels that correspond to the x or y coordinates of objects. 2019); Kulkarni et al.