How a 140-Year-Old Integral Transform Made Our Neural Network 18x Faster
I found that picking the right activation function for a Bayesian Neural Network lets you compute uncertainty analytically in a single forward pass instead of running 50 Monte Carlo samples. 18x faster inference. The math works because Gaussian-family distributions have closed-form moments under exponential-quadratic activations, a property that goes back to Weierstrass’s 1885 work on