Abstract
In neuro-fuzzy learning by on-line first-order backpropagation,
fuzzy rules' interpretability may be lost during the training phase.
This is a so-called interpretability-precision dilemma.
We show that ``hidden-node teaching'' is a simple and effective remedy
for this dilemma. To verify hidden-node teaching effects,
we first analyzed learning results
in two small-scale regression problems. Since
the posed scheme is well suited to a practical large-scale setting,
we then applied hidden-node teaching to a CANFIS neuro-fuzzy modular
network with two local-expert multilayer perceptrons
for attacking the letter recognition problem,
a large-scale UCI machine learning benchmark.
Although our current approach is still ``ad hoc'' (and
thus more extensive investigation is required),
our preliminary experiment shows encouraging results.
