Fig. 5
The impact of weight precision on functional performance. The simulation results for (a) energy distribution of RBM on Fig. 2a and (b) energies after 80th epoch of graph-partitioning problem of Fig. 3c, obtained with different assumptions for the conductance tuning. For RBM simulations, the data were collected over 500 epochs, and were averaged over 500 different trials. In neurooptimization experiment, 200 sets of weights were generated for each case of tuning error. A single data point on a graph represents an energy achieved after 80th epoch, averaged over 16 × 10 trials (10 runs for each of the 16 initial states) for a specific set of weights. For clarity, data point inside 25–75% are not shown. The tuning error was simulated by choosing randomly weights from the range of target value × [1−tuning error, 1 + tuning error]. To make the comparison meaningful, the energy is calculated assuming target (error-free) weights in both panels
