An Infinitely Scalable Learning and Recognition Network

Learning and recognition are fundamental process performed by animals, humans, robots and intelligent systems. Humans, for example, continually learn and recognize where they are in the world (place recognition), who is there with them (facial recognition) and what things are around them (object recognition). Recognition also plays a significant role in technology like smartphones, whether it be recognizing what you are saying (voice recognition) or what the consumer item in front of you is when using Google Goggles (object recognition). Google and other information aggregators perform recognition at a vast scale, recognizing and classifying billions of images in the cloud and house numbers in millions of kilometres of Google Streetview imagery. In security and surveillance, task-specific signatures (such as a specific persons voice, a bomb-carrying back pack or a persons face) must be automatically recognized amongst vast amounts of data. Common to all these artificial recognition processes are computational and storage requirements that grow with the magnitude of the task. Typically these storage and computational requirements grow linearly or worse with the size of the dataset, a critical problem in a world where data storage demand is outstripping capability, and this gap is forecast to continue growing (1). There is currently no feasible solution to this problem current techniques such as those used for video and image compression have plateaued in performance over the last decade, while the limits of hash-based approaches are known and unlikely to provide an ultimate solution. This project combines modelling of and inspiration from the spatial memory encoding system in the mammalian brain with machine learning techniques to enable sublinear storage growth; that is, as the number of items in the database (places, images, voice signatures etc.) that need to be encoded grows, the amount of storage space required per item continually decreases.