A Kakeya set is the smallest space a cat can spin in every direction.
That’s your ReLU network.
Track those spins, and you get tighter control than PAC-Bayes.
Cats don’t take random walks. Neither should your optimizer.
#CatsOfML #Kakeya #CSTheory
PAC-Bayes tells you where the cat’s been.
Kakeya tells you where it can go.
We use cone crossings and directional bounds to track optimizer paths. Tighter than PAC-Bayes. No flat priors.
Wanna know how?
New ADAM paper drops soon.
#CatsOfML #Kakeya #OptimizerTheory #CSTheory
Most TD(0) papers: keep the cat in a box, assume iid data, tiny steps, resets, data dropping, etc.
Ours: let the cat out. Nonlinear approximation, dependent data, real-world dynamics and it still finds the value function.
arxiv.org/pdf/2502.05706
#reinforcementlearning #catsofML #TDzero
