Vision Toolkit for Image Recognition

The Vision Toolkit is an easy-to-use graphical application that lets anyone train their own vision system, without knowledge of the underlying algorithms. The user drags in images (e.g. their own dataset or images from the web), organizes them into categories, and then the Toolkit trains an HTM (Hierarchical Temporal Memory) to recognize those categories. The user can experiment with their trained network by bringing in novel images, and they can upload it to our webserver with one click. Once on the webserver, they can use our web interface to recognize images from anywhere. They can even send images for recognition from any device using a simple API (e.g. an iPhone). Hierarchical Temporal Memory is Numenta's technology and is based on a theory of the neocortex. It comprises a deep hierarchy of nodes, each of which learns variable-order temporal sequences of spatial patterns. Inference is performed using Bayesian belief propagation. It also supports top-down attention, similarity search, prediction, and other useful features that we can demonstrate. The Vision Toolkit uses NuPIC, our platform for HTM development. The full NuPIC release allows complete customization of the HTM parameters, scales from single CPUs to clusters, and supports many sensory modalities.