Design and implementation of a new neural network structure, including analysis of its potential, with particular emphasis on interpretability possibilities.
Demonstration of known theorems related to the new structure.
Design and implementation of various explainability algorithms
Possible extension to Recurrent, Convolutional and Graph Neural Network cases.
The proposed research can be seen as a substantial generalization of the traditional neural network, whose main basic operation has remained almost unchanged since their introduction. In case of a positive outcome of the research, this could open up to numerous other works that would extend to more complex network cases, such as Recurrent, Convolutional and Graph Neural Networks. More generally, the structure created could be adapted for use in any field in which a neural network is currently employed.