It’s hard to grasp just how good backprop is. Normally in science you estimate the effect of 1-3 variables on 1-3 outcomes. With backprop you can estimate the effect of a trillion variables on an outcome. You don’t even need more samples! Around 100 is typical for both (n vs batch_size)
It’s hard to grasp just how good backprop is. Normally in science you estimate the effect of 1-3 variables on 1-3 outcomes. With backprop you can estimate the effect of a trillion variables on an outcome. You don’t even need more samples! Around 100 is typical for both (n vs batch_size)