I am lecturing on important "math" tricks for machine learning. One that I like a lot is softmax overflow, where we subtract the max logits from the exponents to ensure computational stability.
Most probably know of it, but I still find it nice.
#machinelearning
