How does Arya-xAI work?

Arya XAI currently has two modes of operation known as default and contrastive mode which are demonstrated through the following use-cases respectively:

Network Analysis: In this mode, there is a single,positive significance value that is propagated in proportion to the contribution of each unit. This can be used to prune the underutilized parts of the network and modify the saturated ones. This is also helpful in reducing or enhancing the input features.
Decision Analysis: In this mode, there are dual significance values which can be either negative or positive depending on the relative contribution of any unit to the decision. This is helpful in providing explanations regarding the corresponding input features that are positively/ negatively affecting the decision .

The procedure for significance/relevance calculation is as follows:

From the model weights and architecture construct a graph with output nodes as roots and input nodes as the leaves.
The propagation starts at root and proceeds in breadth-first manner to avoid re-processing of any node.
The propagation completes when all the leaves(input nodes) have been assigned relevance.
For the default mode, any loss of relevance during propagation is due to network bias. However, network bias is ignored for relevance computation in the contrastive mode.
The relevance of a single sample represents local importance. For global importance, the relevance of each feature can be aggregated after normalization on the sample level.

IMPACT ON PRUNING AND MODEL RESULT ASSESSMENT

For model pruning, current explainable approaches offer limited evidence on node/layer level importance in final prediction. The limits range from type of data to network architecture. Arya-XAI uses a unique approach to figure out the importance of each node/layer on the final prediction. This allows a much deeper explainability of the complex black box models for data scientists/researchers, allowing them to gauge which features have the most impact on the model decision. Explanations can be generated at a local level as well as a global level.
Global level explanations can be used to study the overall model trend, data usage and identify possible improvements. Local level explanations are important for granular control over data usage and unit level performance of the network. Local level explanations not only help in analyzing model outputs for individual samples, they also help in establishing internal trends in any feature with respect to the decision. This facilitates in building feature-level decision boundaries and flag any adverse predictions.