We ran the van der Pol oscillator (A–D) or the linear decaying oscillator (F,H) learning protocol for 10,000 s for different parameter values and measured the mean squared error, over the last 400 s before the end of learning, mean over number of dimensions and time. (A) We evolved only a fraction of the feedforward and recurrent connections, randomly chosen as per a specific connectivity, according to FOLLOW learning, while keeping the rest zero. The round dots show mean squared errors for different connectivities after a 10,000 s learning protocol (default connectivity = 1 is starred); while the square dots show the same after a 20,000 s protocol. (B) Mean squared error after 10,000 s of learning versus the standard deviation of noise added to each component of the error, or equivalently to each component of the reference, is plotted. (C) We multiplied the original decoding weights (that form an auto-encoder with the error encoders) by a random factor (1 + uniform) drawn for each weight. The mean squared error at the end of a 10,000 s learning protocol for increasing values of is plotted (default is starred). (D) We multiplied the original decoding weights by a random factor (1 + uniform), fixing , drawn independently for each weight. The mean squared error at the end of a 10,000 s learning protocol, for a few values of on either side of zero, is plotted. (E,G) Architectures for learning the forward model when the reference is available after a sensory feedback delay for computing the error feedback. The forward model may be trained without a compensatory delay in the motor command path (E) or with it (G). (F,H) Mean squared error after 10,000 s of learning the linear decaying oscillator is plotted (default values are starred) versus the sensory feedback delay in the reference, for the architectures without and with compensatory delay, in F and H respectively.