Figure 9.

Adaptation of the firing threshold θ. (A) Detail of the LIF threshold θ^disruption (visualized as the red dashed line) adaptation for the left leg's early swing stop. Initially, the threshold is set to zero, thus LIF fires (in the green) at the first non-zero error e^disruption (in the black), where the error is rectified difference between the early stop event x (in the blue) and RBF anticipation a (in the magenta), h(x − a). During the LIF firing, the threshold rapidly grows; therefore, the next LIF non-zero activity at step 400 is below the threshold, and LIF does not fire. The threshold slowly decays (not observable in plots). (B) The LIF detector (in the yellow) for the left leg's contact absence behaves similarly. The last thousand steps of the LIF neuron activations are aggregated in histograms, where it is shown that the respective thresholds are upper-bound of the regular activations. (C) The swing stop perception is precise during the regular motion; thus, the LIF activity (in the green) is similar for all legs, and so are the thresholds (showed as the red dashed line). (D) However, the ground contact perception differs for each leg (probably due to different loads on the legs during the stance) and is less precise (the leg sometimes did not detect the ground contact). It resulted in the increased variance of the ground contact absence thresholds across the legs. Note that the contra-lateral legs (e.g., cL1 and cR1) have similar thresholds.