Suppose I have a continuous stream of events - say the price of a stock.
If I have a trained recurrent network, then at any point in time, we have the predicted value and the actual value.
It seems intuitive that I should be able to use the actual value to train the model on an on-going basis?
However, all the theory I have read strictly splits the training from the predictive phase? During the predictive phase the weights are held constant. Is there any reason why that is?
Yes, there's some work on this topic.
You can search for terms like
for time-series data or for whatever you need.
For an intro on continual learning with neural networks, I recommend the paper Continual Lifelong Learning with Neural Networks: A Review (2019), but if you're looking for something more specific, you can search for a more specific paper or tool, like Continual Learning Long Short Term Memory (which I have not read yet) or the river and avalanche Python libraries, which might also be useful.
