The world we live in is a spatiotemporal world, where all of the perceived input is taken as spatial signals through time. Simple tasks such as recognizing and following your pet cat have proven to be very difficult for traditional computing systems. This is arguably, in part, due to the traditional algorithms focusing on the spatial and temporal data separately, instead of as a unit. This work aims to take that work further by combining standard statistical, computing, practices with concepts inspired by the mammalian brain.
The current focus on this work is directed at Hierarchical Temporal Memory (HTM), which is a brain-inspired machine learning algorithm that operates on both spatial and temporal data. A custom implementation of HTM is being developed and it will be applied to various applications. A comprehensive study on the algorithm will also be performed. It is desired to use HTM as a starting point for developing future systems. Those systems will take abstractions from biology, combining them with statistical models to be used for more complex applications, such as anomaly detection in video streams.
The figure shows an example of a Spatial Pooler (SP) in HTM. As the name implies, the SP is responsible for working with spatial data. In that figure, the green lines represent active synapses and blue lines represent inactive synapses. In this case, a threshold of two active bits is utilized, causing the two center columns to become active and the outer columns to remain inactive. All cells in the columns are either active or inactive, as the cells share the same proximal segment. Based off the activation of the columns a new representation of the spatial data is created.