MFM: Multiple Forward Model Architecture for Sequence Processing

A multiple forward model (MFM) architecture is proposed for sequence identi cation, learning and production. MFM is inspired by Wolpert and Kawato's [ 1998 ] multiple paired forward and inverse models architecture for motor control. In particular, learning of sequencespeci c modules and switching among multiple sequences are demonstrated. Appropriate sequence modules are chosen and maintained on the basis of feed-forward prediction errors and softmax responsibility estimation. Simulations are conducted in order to reproduce the ndings from Tanji and Shima's [ 1994 ] sequence experiments on monkeys. Simulations demonstrate that the network is capable of learning four simple three-step sequences in a supervised fashion initially using external cues and eventually reproducing them without external cues as in the monkey experiments. Preliminary hypotheses are made about how various components of the network mimic the overall function of the cortex and the basal ganglia in mammals. Finally, general implications of this architecture for hierarchical sequence processing are discussed.