The third major release of the Imaging Whiteboard has all the features of previous releases and adds new Artificial Intelligence (AI) features, including:
A tool which will allow the user to design a Convolutional Neural Network (CNN), to train and test this network and to save the network at any stage.
An Image Classifier Control that will use trained networks to classify images.
Enhancements to existing controls to support the development of training images, and the integration of the Image Classifier into Whiteboards.
The Convolutional Neural Network Configurator.
The CNN configurator will allow the user to build and configure neural networks and
convolutional neural networks by adding convolutional, pooling, and fully connected layers.
These layers can be configured, then the network trained and tested. The network can be saved
to disk at any point in this process so that training can be resumed later; or the final network can
be used by the classification control.
Fully configurable layer structure.
Sequencer for fully hands free training.
Hyperparameters: Batch size, Gradient Velocity, Gradient Clipping, and learning rate configurable dynamically.
Graphical and textual feedback during training.
Full training and testing history stored with the model, along with weights, filters, and configuration.
Xavier initialization for weights and filters.
Activation Functions: Sigmoid, ReLU, Leaky ReLU, or Softmax (final layer only)
The Image Classifier
This is a new control which will allow the user to load a trained neural network or convolutional neural network, and use it to classify images.
The Blob Counter Output feature
Selected blobs can be played (output to other controls) using the step button or the play button.
Using the Image Classifier to read handwritten numbers
Here we use a Convolutional Neural Net (CNN) that has been trained using the MNIST
database of handwritten digits.