Let us say that you are having a Recurrent Neural Network designed $$s a LSTM$.$ Let us say that

you are training it $($at different times$)$ for two different types of sequences. In Sequence type A$,$ values from far i n the past continue to have significant impact on the output at present, while values from the Current have relatively less impact; in Sequence type B it is the exact reverse. Thenintheprocessoftraining,theweights$($Wi$,$b$)$oftheForgetGateoutputfi $($controlparameter of the output gate with range $[0,1]),$ and the weights $($Wi$,$ bi$)$ of the Input Gate output i$,$ will so adjust themselves that:

a$)$

b$)$

c$)$

d$)$

f willbe small$($near zero$)$ for the Sequence Aand large for Sequence B$,$while i$,$ will be large $($near one$)$ for Sequence Aand small for Sequence B

f$,$andi willbothbelargeforbothtypesofsequences

f$,$ willbe large for the Sequence Aand small for Sequence B$,$ while i$,$ will be small for Sequence Aand large for Sequence B

the values of f and ie willnot be amenable to simplistic classification as either large or smallfor either type of sequence.