We consider problems of sequence processing and propose a solution bas
ed on a discrete-state model in order to represent past context, We in
troduce a recurrent connectionist architecture having a modular struct
ure that associates a subnetwork to each state, The model has a statis
tical interpretation we call input-output hidden Markov model (IOHMM).
It can be trained by the estimation-maximization (EM) or generalized
EM (GEM) algorithms, considering state trajectories as missing data, w
hich decouples temporal credit assignment and actual parameter estimat
ion, The model presents similarities to hidden Markov models (HMM's),
but allows us to map input sequences to output sequences, using the sa
me processing style as recurrent neural networks. IOHMM's are trained
using a more discriminant learning paradigm than HMM's, while potentia
lly taking advantage of the EM algorithm. We demonstrate that IOHMM's
are well suited for solving grammatical inference problems on a benchm
ark problem, Experimental results are presented for the seven Tomita g
rammars, showing that these adaptive models can attain excellent gener
alization.