ctc forward backward algorithm - why is it being used?
0
$begingroup$
as per ftp://ftp.idsia.ch/pub/juergen/icml2006.pdf fwd bw algo helps us speed up discovery of the happy path. For e.g. if the GT label is DOG and we have 10 time steps , the possible label sequences could be _DD_OO_GG_ DD_OOO_GGG DO__GGGGGG etc ..so one path could be "_DO" (index 0, 1 and 2 from 3 example possibilities shown above). Assuming most tasks will have many more time steps (and hence a much higher number of paths), how is the fwd bw algo speeding things up ? Lets say the fwd variable is calculating the prob of obtaining an "O" at time step T, does it mean that it picks the best possible path of getting to "O" (recursively) ? for e.g. if the 3 paths to "O" were "_DD" , "DDD" and "blank D blank" with probabilities...