Abductive networks: generalization, pattern recognition, and prediction of chemical behavior

Neural networks have linear sums of weights incoming to a neuron.

Abductive network module may have a third degree polynomial function with up to 3 input variables.

Measures and scores of predicted error versus complexity required. Can adjust complexity allowance based on needs.

Most accurate networks are the ones with similar scores in both training and testing. These networks are best able to make predictions on unseen inputs.

Relies on an "abductory inductive mechanism" (AIM) which was an algorithm hidden in proprietary software.

AIM attempts to find patterns between the input examples and output examples and takes a dense mesh of polynomial functions (a deep neural network, up to 9 layers) and winnows out parameters which do not help.