AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
I'm a Big Fan of the Spline Theory of Neural Networks
i've been speaking with rando balestriero and he's got this paper the spline theory of neural networks which is about these input sensitive polyhedra in the ambient space. i always took that to mean why they're quite interpretive and it's just an affine transformation for a single input but yeah what what he showed though is that actually even an mlp with relu's it's extremely extrapolative because you can remove a whole bunch of data and depending on how you've designed the network architecture it will still inform that region that that you've taken away.