Motion Memory: Invariant representations of sequences in cortical L2/3 by Hierarchical Temporal Memory

We aim to form stable representations of temporal sequences with key focus on semantic learning and streaming data. The state of the art in the Hierarchical Temporal Memory is represented by Numenta’s recently published “ColumnPooler” which emulates functionality of cortical L2/3 layer, forms stable allocentric representations of temporal sequences and/or objects, and has been applied to sensory-motor learning. Our designed experiments evaluate the ColumnPooler for such task and uncover its current limitations. Presented “Motion Memory” design defines needed modifications in order to be effectively used for sequence representation, namely: Semantic distance between the representations; Online learning on streams; ability to represent time; and representation of motion from static sensor. One of the main problems with the current design is the lack of semantic meaning in (continuously updated) representations of the object. The proposed improvement enables MotionMemory to do unsupervised learning on streaming data and resulting representation have semantic meaning, this has many practical applications in sensory processing (ie. vision), or hierarchical learning.