AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
Is There a Best Explanation for Why It Works?
It's keeping information about both. What actually blows my mind about this is the way that it combines them is an addition and then a norm, which is basically just constraining like the values that can be an effector to be smaller than a certain amount. It turns out that this is enough to propagate information deeply into the network to just do that addition with the original input. This is actually one of the problems as well. It's difficult and not necessarily rewarding to experiment with little architecture tricks.