Sensitivity analysis of prior knowledge in knowledge-based neurocomputing

Knowledge-based neurocomputing addresses, among other things, the encoding and refinement of symbolic knowledge in a neurocomputing paradigm. Prior symbolic knowledge derived outside of neural networks can be encoded in neural network form, and then further trained. Previous research suggested certain values for the weights that represent prior knowledge, based on an analysis of the derivative of the error function. This inductive bias is investigated empirically, and furthermore, we show how to use sensitivity analysis methods to investigate this bias. This work shows that the bias of the encoding method for the prior knowledge corresponds well with a range of good parameter values that retain the encoded knowledge and allows refinement by further training.