Unsupervised learning of spatial transformations in the absence of temporal continuity

Learning features invariant to arbitrary transformations in the data is a requirement for any recognition system, biological or artificial. Such transformations may be learned using label information or from temporal data in an unsupervised manner by exploiting continuity. This paper presents a dynamical system for learning invariances from real-world spatial patterns in an unsupervised manner and in the absence of temporal continuity. The model consists of a simple and a complex layers. Given an input, the simple layer imagines all of its variations, each with a degree of consistency, and eventually settles for the most consistent reconstruction. During this imagination, the complex layer learns the consistent variations of the same pattern as a transformation in each spatial region. Experimental results are comparable to those from supervised learning. The conditions for stability of the system are analyzed.