Learning features and transformations with a predictive horizontal product model

According to the theory of two parallel visual pathways, the ‘dorsal pathway’ encodes spatial information, invariant of stimulus-specific properties, while the ‘ventral pathway’ encodes object feature identity, invariant of positions and sizes. Commonly, they are referred to as the ‘where’ and ‘what’ pathways. Such a distinction between spatial transformation- and identity encoding cells is already evident in the diverse response properties of V1 complex cells [2]: some complex cells are direction- and speed selective, independent of spatial frequency, hence resembling neurons in Medial Temporal Lobe(MT) of the dorsal pathway (they predict ‘where’). Other complex cells are selective to spatial frequencies independent of speed, hence coding feature identity (‘what’). Interestingly, to compensate for upstream and downstream neural transmission delays, the ‘where’ pathway should maintain a future position of an object. This can be accounted for by the representation of movement direction and velocity in the dorsal pathway. It is important, for instance, when a person detects a temporal pattern change in visual target stimulus. . . . . . . We propose a new architecture which learns object position/motion and feature identity in an unsupervised fashion based on a predictive model.

@InProceedings{ZWW12b, 
 	 author =  {Zhong, Junpei and Weber, Cornelius and Wermter, Stefan},  
 	 title = {Learning features and transformations with a predictive horizontal product model}, 
 	 booktitle = {Proceedings of the Sixteenth International Conference on Cognitive and Neural Systems (ICCNS 2012)},
 	 editors = {},
 	 number = {},
 	 volume = {},
 	 pages = {},
 	 year = {2012},
 	 month = {May},
 	 publisher = {},
 	 doi = {}, 
 }