Knowing the ideal network architecture to solve a particular problem is a significant challenge in working with current neural networks. Many hours of training, evaluating and retraining are required and the outcome can still often be unsatisfactory. Additionally, the environment the architecture was tested for may change over time, significantly reducing efficacy.
Presenting BrightNet, an open network architecture, spontaneously responsive to each unique environment.
BrightNet simultaneously optimises the network to use only the connections necessary and explores new pathways as it continues to learn and respond. Connections, nodes and layers resolve when no longer required, opening computation resources. New connections, nodes and layers are added throughout the training process to allow the network to grow to the optimal size. This allows BrightNet be responsive to an ever changing environment.
We have tested BrightNet on various industry benchmarks including MNIST dataset (see below) and find it demonstrates results that prove both its efficacy and high performance.
Table 1: Performance on MNIST.
Showing comparison between Multilayer Perceptron (MPL) and BrightNet.
*Note: BrightNet adapts to the dataset and reduces/increases node count accordingly displaying significant improvements in efficacy.