next up previous contents
Next: Experiments Up: Task 5.2: Technical Description Previous: Forward-Backward

Forward-Backward training

From the second equation, a forward-backward algorithm for hybrid HMM/ANN can be derived as an application of the generalized EM algorithm where the E-step is the re-estimation of the ANN targets using forward-backward recurrences and the M-step is the ANN training (see [32] for details about the $\alpha$ and $\beta$ recurrences).


Christophe Ris