Fig. 1

Schematic overview of the regression feedforward neural network topology central to the MRL framework. A time slice \({\varvec{e}}\) of the signal envelopes of \(I\)-channel sEMG at time \(t\) is fed through an encoder network, constituted by \(N\) fully connected blocks, which transforms \({\varvec{e}}\) into an alternative representation, i.e. code, \({\varvec{h}}\). A set of \(J\) decoder networks, each constituted by a single hidden block, decode this representation to independently estimate the activations of \(J\) DoFs, interpreted collectively as the proportional and simultaneous output command \(\widehat{{\varvec{y}}}\) corresponding to the movement intent of the user at time \(t\)