Explainable Nodes

Most nodes in inner layers of a deep neural network are difficult to understand and explain. Using our Learning Coach and Incremental Network Growth methods, D5AI changes the question. Rather than explaining an selected node in fixed, pretrained network, D5AI adds one or more new nodes that are each designed and trained to to be duplicate some of the functionality of the original selected base node and to be easily explainable.

​

Do-it-yourself guide to creating and training an explainable node:

0. This simplified example assumes the activation function of the base node is monotonic increasing and the there is a natural threshold value T between acceptance and rejection. It assumes that the base network has been trained to minimize an error loss function.

1. Select a sample of training data

2. For each data example, compute the activation of the base node and the derivation of the network objective

     a. The activation A is compared to the threshold T. If A > T, it is regarded as detection. If A<T, it is regarded as a rejection.

     b. For minimizing an error loss function, a gradient descent update is in the direction of the negative of the derivative.

3. For each data item, determine whether the selected node has made an error on the implicit local objective of the node:

     a. If A>T, but the derivative is positive, there is a false positive detection error.

     b. If A>T, and the derivative is negative, there is a correct detection.

     c. If A<T, and the derivative is positive, there is a correct rejection.

     d. If A<T, but the derivative is negative, there is a false rejection error.

​