"Feed forward NAL-based Neural Network".
Neural networks are one way to represent any computable function. They can be trained with a training algorithm. Non-bayesian NN's can't handle uncertainty. This paper proposes a way on how to leaverage some of the advantages of NN's to represent and learn a function which also computes uncertainty. This is archived by using functions which use NAL truth values. Advances in deep learning training methods may also get leaveraged in the future, but this is not explored in this work.
- All values are represented as NAL-Truth Values
- ring of addition is replaced with revision, ring of multiplication is replaced with deduction (TODO - verify if this makes sense)
- nonlinearities may get represented as nonlinear functions which get (f,c) as input and return (f,c) as output
- uncertainty is fed into the inputs - outputs have a uncertainty too
- input-output mapping of hidden units is learned with a learning/optimization algorithm, for example back propagation or evolutionary algorithms (EA)
TODO< write core ideas up>
TODO< write rust code to demonstrate that it really works for small problems.
TODO< reference to bayesian NN networks >
Non-axiomatic reasoning system: Exploring the essence of intelligence
Deep Learning