As the user is supplying corrections for misrecognised characters, it would be possible to also use this information to progressively train the character recogniser as the user is using the system. The system would progressively learn to recognise the user's writing over a period of time. The only danger is that the user might not supply the correct interpretation for a symbol, and end up confusing the recogniser further. However, the system may be able to determine if this is happening when there are significant differences between the symbol supplied and all the other training examples the character recogniser has.
It takes the current character recogniser 4.0 seconds to train on 451 examples of 180 symbols. As each symbol is trained independently of the others, it takes about 0.02 seconds to train it for a single symbol. Due to this high training speed, online training is possible.
Even if the character recogniser took longer to train, the training could be executed in the background and the character recogniser's data could be updated every few seconds.