Fig. 6: Varying device asymmetry.
From: Fast and robust analog in-memory deep neural network training
Different device materials show different degrees of asymmetric conductance responses. A different device responds with varying degrees of asymmetry (changing \({w}_{\max }\) and fixing the step size). Colors of the example pulse responses to 200 up and 200 down pulses indicate the asymmetry device setting. B Weight errors (computed as in Fig. 4) achieved by the various algorithms depend on the degree of device symmetry. Note that only AGAD retains a very low error independent of the asymmetry setting (green line). Asymmetry, typically very detrimental for direct SGD implementation (red line), is necessary for TTv2 (blue line) as well as c-TTv2 (orange line). This is because the latter algorithms hinge on the assumption that the conductance quickly returns to the symmetry point (SP) and the time constant to reach the symmetry point (SP) for random updates depends on the asymmetry (see Eq. (7)). Error bars indicate standard errors over 3 construction seeds.
