Skip to the content ForoHackers
Log inJoin

gradient_descent_hiker

Joined 15 May 2025

Posts in the forum

  1. Replied in Monte Carlo and the 1/sqrt(N) wall

    Quasi-random sequences (Sobol, Halton) improve the rate for smooth integrands towards roughly `1/N`, at the cost of losing the clean statistical error bar. Different tool, different promises.

  2. Replied in Counting k-mers: the smallest useful bioinformatics program

    Small tip from the ML side: if you feed k-mer counts into a classifier, normalise by total count per sample first. Otherwise the model just learns "which sample had more reads", and you get a beautif…

  3. Replied in A benchmark score is a measurement. Where are the error bars?

    Agree. I have been guilty of reporting the best of five seeds. If the five runs differ by more than the gap between two methods, the gap says nothing. I now report mean and spread over seeds, and the…

  4. Replied in A learning-rate experiment you can run in 20 lines

    Correct on both. Next thing I want to test is what happens with two directions of very different curvature, say `f(x, y) = 10 x^2 + 0.1 y^2`. I expect the large one to dictate the learning rate and t…

  5. Started A learning-rate experiment you can run in 20 lines

    New here. I like small experiments that show why a knob matters. Learning rate, on the simplest possible function: `f(x) = 10 x^2`, whose gradient is `20 x`. ```python def run(lr, steps=50): x =…