Martin Roberts

With much excitement and after a 3 year hiatus, I share my latest blog post "A New algorithm on multi-armed bandits". extremelearning.com.au/multi-armed-ba… This topic that falls under the broad umbrella of #ReinforcementLearning and played a key role in the AlphaGo algorithm.

I've proven the strangest result of my career.. The classic idea that gradient descent's rate is best with constant stepsizes 1/L is wrong. The idea that we need stepsizes in (0,2/L) for convergence is wrong. Periodic long steps are better, provably. arxiv.org/abs/2307.06324

@roydanroy @Atrix256 You might like to read about isolation forests. I wouldn’t say it had a profoundly deep theoretical basis, but it was the first time in many years that I saw an approach for outlier detection that had a basis akin to information theory.

Who's an expert in anomaly detection out there? What's are the key texts in this area? Does it have a rigorous theoretical foundation? That handles the adversarial aspects? And what about deep learning and in-roads there?

It wasn't until I'd *taught* algorithms a few times that I finally understood why sorting is in the CS curriculum. Unfortunately, most curricula don't explain this! It is NOT because sorting is an important algorithm to learn to implement...

@EricLengyel That is cool.

Here's something about the quadratic equation that isn't often taught. The sign chosen in the solution x = (−b ± sqrt(b² − 4ac)) / 2a is also the sign of the derivative at the root x. Choosing plus always gives the root where the curve goes up through the x axis. 1/3

I love the Samsung Frame TV but got really tired of paying a subscription for the art—so I made a little script that lets you use a folder of your own images to rotate randomly on your TV. 👀github.com/ow/samsung-fra…

@ow This is very cool. My Samsung Frame tv is one of my favourite things in my house and all my family, friends and visitors love it so much. But I had the same irk as you. For the record, I would happily pay you to continue this, then to pay a faceless commercial company! 😎💥✨

@jrdnfrd @Atrix256 Kinja Basu was one of the first to write about this problem in 2014. Then in 2019 I did a post on it, extremelearning.com.au/evenly-distrib… Then in 2020, Alan's friend, Eric Heitz, published a method much better than either of these.... eheitzresearch.wordpress.com/749-2/

@Atrix256 How would you select low-discrepancy points on a triangle mesh? I want to choose low-discrepancy points on a terrain mesh to use as visibility probes.

Yeah and I thought it was really pushy and self-centred when I first saw it. But then I realised that everyone’s feed is so full and stuff disappears down within hours, that reposting useful things across the weekend at different times can actually be very helpful to readers.

@Atrix256 I think although Twitter has passed its zenith, it will still be around for a long long time. Nobody doubts they IBM or Yahoo have passed their prime, and virtually no-one ever mentions their name but they are still in 2023 both absolutely massive.

Our Global Gender Gap Report is up for a Webby Award! Still some hours left to vote. If you ❤️ it, plz vote for us! vote.webbyawards.com/PublicVoting#/…

Beautiful animations of the Fibonacci lattice when mapped to a disk and sphere.

@meetamit This is very cool! Are you happy for me to put a link in my post to your ObservableHQ notebook?

@TechSparx Hi, I recently made a notebook based on your Fibonacci Lattice posts, including animations of lattice creation and projection, plus interactive inputs to adjust ε, Φ and other params. If interested: @meetamit/fibonacci-lattices" target="_blank" rel="nofollow noopener">observablehq.com/@meetamit/fibo…

Here's a small (but long overdue!) blog update to a very popular post that I wrote 2 years ago. "How to evenly distribute points on a sphere more evenly than the canonical Fibonacci Lattice (Golden Spiral). bit.ly/30hyeWH

@AlexisPaques @Atrix256 Many thanks. :)

@TechSparx @Atrix256 github.com/x-hw/amazing-qr It has been done many times :)

@michael_nielsen The reason I learned at (Australian) primary school was that “every word has a vowel in it.” Almost every word has at least one occurrence of aeiou in it, but to make this rule complete, you need to include “y”. This covers words like: by, dry, fly, my, try,

A teacher in my primary school told me y was "sort of" a vowel. I have no idea why, but the network agrees!

Enjoyed this tidbit in @karpathy's explanation of how to train a neural net to generate names: the first layer learns a 2-d embedding of the alphabet; it automatically discovers vowels :-) youtube.com/watch?v=TCH_1B…

@pmddomingos This is the secret of the MacDonalds business model. It’s not the Big Mac, the real estate or cheap junk food. It’s their unrivalled ability to train teenagers and 20-somethings, to run a commercial fast-paced business at international standards.

If someone wrote a book on how to build a competent organization out of incompetent people, it would be a historic bestseller.

@Atrix256 And the absolutely classic article on how all these different special constants are the basis of specific infinite tilings… blog.wolfram.com/2019/03/07/sha…

@Atrix256 The golden ratio and the supergolden ratio are both Pisot numbers, which have the cool property that their powers are extremely close to integers.

@Atrix256 And of course as you know, the plastic constant 1.3247…. (which is also a pisot nnnber) is the basis for my quasi random sequence R_2.