Integration and differentiation in dynamic recurrent networks
Edwin E. Munro, Larry Shupe, Eberhard E. Fetz · 1991
Summary form only given, as follows. Dynamic neural networks with recurrent connections were trained by backpropagation to generate the differential or the leaky integral of a nonrepeating frequency-modulated sinusoidal signal. The trained networks performed these operations on arbitrary test inputs. Reducing the network size by deleting and combining hidden units and then retraining produced smaller networks that computed the same function and revealed the underlying computational algorithm. Networks could also be trained to compute simultaneously the differential and integral of the input on two outputs; the operations were performed in distributed overlapping fashion, although the activation of the hidden units resembled the integral.>