Sounds like you’re walking the line between “less is more” and “too much is a tragedy.” Try treating expressivity as a budgeted function—something like the Frobenius norm of the weight matrix minus a sparsity penalty, maybe weighted by the Hessian trace to capture curvature. Then formulate a Lagrangian where you minimize the loss plus a λ times that sparsity term, but also add a constraint that the eigenvalue spread stays above a threshold. In practice, you could monitor the drop in mutual information between layers as you prune; if it stays flat, you’re still expressive enough. Just make sure you don’t end up with a one‑liner that looks like a punchline.

Sounds like a solid plan—just make sure the script doesn’t turn into a marathon. If the eigenvalues start dancing too closely, give them a gentle reminder to spread out. Keep an eye on that mutual information curve; it’s the real heartbeat of expressivity. Coffee’s on me, the rest is pure patience and precision.

Just remember, if the script turns into a marathon, it’s not the coffee you’re paying for—it’s the sheer weight of those parameters. Keep the eigenvalues in check and the MI curve steady, and you’ll have a pruning routine that’s both elegant and efficient. Good luck.