The new training will override the old.
Therefore you have to either
a. Add the old training set to the new one
or
b. Add representative examples of the new dataset to the old one.
Then retrain to obtain new coefficients.
Decades ago, the Army asked me to design a set of classifiers that would identify the different products resulting from a missile deployment (e.g., missile, decoys, & deployment hardware). New test flight data was delivered periodically.
I found the best approach was to use a radial basis function classifier. When new data was run through the classifier and misclassified, new clusters were added to the data base. Then new and old coefficients were updated to minimize error rates.
Hope this helps.
Thank you for formally accepting my answer
Greg
