Interactive Training of Artificial Neural Networks
Our goal is to outperform the vanilla backprop with interactive interventions into the learning process of a neural network. This requires to propose a human-machine interface, i.e. interaction variables have to be included into the Multilayer Perceptron network (MLP) model and incorporated into the backpropagation learning algorithm. Also a visualization method has to be selected to comprehend the learning process of MLPs. To maintain simplicity, only visualizable 2D tasks are investigated, i.e. remote sensing, image compression, image filtration and proposed benchmark classification tasks (circle, spiral, square, etc.).
Current research
We have proposed and implemented a framework for interactive training of MLPs. Experiments on the following remote sensing testbeds are in progress.