Comparison of estimated multiplicative shear bias m₁ (left) and its error Δm₁ (right) as a function of S/N for different ML hyper-parameters. Top panels: performance for different numbers of objects used in the training, from 10 000 to 400 000. Middle panels: performances for the chosen architecture for different number of epochs, from 1000 to 9000. Bottom panels: performance for different architectures. The chosen one, shown in blue, corresponds to four hidden layers of 30 nodes each. The shallow architecture, shown in green, has only two hidden layers of 30 nodes. The narrow architecture, shown in orange, corresponds to four hidden layers of 30, 20, 10, and 10 nodes.