I was thinking. Why ”compress” information into a small number of weight parameters and also perform value prediction with them. This many jobs for the weights I think just makes it more difficult.
So I thought. Could you split the process of training. I mean. A time series could perhaps be split into several data series. Some sort of extra parameters.
What if the weights are not constants but stored in a time series themselves. Now that we have thought of a large storage space for parameters the training of them should also be made easier. What if they could contain errors. A good enough solution is better than no solution. So when when training the large parameter list we should only train those parameters that make sense. Otherwise it would take too long.