Training

Training involves asking a network questions and then adjusting the network when it answers incorrectly. The first step in training is asking a question. Asking a question is referred to as inference. The second step is applying a correction, this is called Backpropagation.

Training, and specifically Backpropagation, is a computationally expensive endeavor. To train networks efficiently, specialized hardware called GPGPU or General Purpose Computing on graphics processing units are used. GPGPU's or just GPU for short, can accelerate training by as much as 100 times over the x86 processor found in most personal computers.

Training also requires a lot of data. Data is of two two types, labeled and unlabeled. Labeled data is data that is of a known type and training a network with this data is called Supervised Learning. For example pictures could be labelled by color or black and white. Pictures of zip codes (the mnist dataset) could be labelled by the digits within. Unlabeled data is of an unknown type. A dump of ten thousand random Twitter comments is an example of unlabeled data. Labeled data is always preferred to unlabeled. Using unlabeled data to train an network is called Unsupervised Learning. Unlabeled data can also be used indirectly during Supervised learning through Transfer Learning.

VR Experience Reconstruction and Refinement.

Somax views the world through a special set of eyes. It see's color like you and me. also like humans, it has depth perception and can understand how close / far away objects are.

Reconstruction. Unlike most humans Somax can see, remember, and reproduce objects in 3 dimensions at a level of detail beyond comprehension. On command Somax can for example capture a 3 dimensional model of a human with sub millimeter accuracy! On command it can also capture a spherical 3 dimensional picture of its visually unobstructed environment with a radius up to 10 meters. This picture will contain can contain full color image data as well the distance to each pixel in the image. Somax will use it's AI engine to transform this data into OpenGL instructions describing the location and estimated geometry of it's visible surroundings. A separate viewing application specific to a VR platform will use this information to load the OpenGL instructions and generate the world as experienced by Somax at that moment. And even better it will allow you to move to a different point of view and see what Somax imagines objects might look like.

Refinement. During refinement a Somax will ask a user questions to clarify things it is having difficulty understanding. Sometimes it may be useful to reproduce the environment of question, and it's counter environments, and then measure the human response to each. The more faithful the reproduction is the more accurate the answer. This is where VR and Somax will work together. Somax will record an accurate 3D audio-visual environment and a VR will accurately reproduce it!