Project Lemma.

The problem

Given the weights of a small random MLP, estimate its mean output over Gaussian inputs without simply running it many times. Methods that read the weights directly are the point: sampling tells you little about rare behaviour. The starter kit has 100 networks, the baseline estimator and the scoring script. Send a link to your code.

How it is checked

We run your code on 100 held-out networks under a fixed compute budget. The first submission with mean squared error at least 20% below the sampling baseline wins.

Submit an answer

The sponsor checks it against the test above and replies by email.