Spatiotemporal information during unsupervised learning enhances viewpoint invariant object recognition

摘要

Recognizing objects is difficult because it requires both linking views of an object that can be different and distinguishing objects with similar appearance. Interestingly, people can learn to recognize objects across views in an unsupervised way, without feedback, just from the natural viewing statistics. However, there is intense debate regarding what information during unsupervised learning is used to link among object views. Specifically, researchers argue whether temporal proximity, motion, or spatiotemporal continuity among object views during unsupervised learning is beneficial. Here, we untangled the role of each of these factors in unsupervised learning of novel three-dimensional (3-D) objects. We found that after unsupervised training with 24 object views spanning a 180° view space, participants showed significant improvement in their ability to recognize 3-D objects across rotation. Surprisingly, there was no advantage to unsupervised learning with spatiotemporal continuity or motion information than training with temporal proximity. However, we discovered that when participants were trained with just a third of the views spanning the same view space, unsupervised learning via spatiotemporal continuity yielded significantly better recognition performance on novel views than learning via temporal proximity. These results suggest that while it is possible to obtain view-invariant recognition just from observing many views of an object presented in temporal proximity, spatiotemporal information enhances performance by producing representations with broader view tuning than learning via temporal association. Our findings have important implications for theories of object recognition and for the development of computational algorithms that learn from examples.