mc - Network

https://mcognetta.github.io/posts/leetcode-random-seed/

https://mcognetta.github.io/posts/leetcode-random-seed/

This problem came up again, so I updated my old solution!https://sigmoid.social/@mc/111662581917469235

This problem came up again, so I updated my old solution!

https://sigmoid.social/@mc/111662581917469235

Tokenization is an often-overlooked aspect of modern #NLP, but it’s experiencing a resurgence — thanks in large part to @karpathy and...

Tokenization is an often-overlooked aspect of modern #NLP, but it’s experiencing a resurgence — thanks in large part to @karpathy and his classic tweet:

x.com/karpathy/sta...

Come hang out with us and let's fix these problems!

Today we are launching a server dedicated to Tokenization research! Come join us!discord.gg/CDJhnSvU#nlproc #machinelearning #tokenization

Today we are launching a server dedicated to Tokenization research! Come join us!

discord.gg/CDJhnSvU

#nlproc #machinelearning #tokenization

Geniushttps://youtu.be/TUtafoC4-7k#chess #programming

Genius

https://youtu.be/TUtafoC4-7k

#chess #programming

Gboard never stops innovating.https://youtu.be/EHqPrHTN1dU#keyboard #jp

Gboard never stops innovating.

https://youtu.be/EHqPrHTN1dU

#keyboard #jp

It's frustrating that you can't use the walrus operator in list comprehensions where you can in the unrolled loop.Not that I would...

It's frustrating that you can't use the walrus operator in list comprehensions where you can in the unrolled loop.

Not that I would do it often in real code, it's just annoying for when I want to golf.

#python

(Another) another day, another Japanese karaoke Korean keyboard variant. Just how many of these are there?This is like the third variant...

(Another) another day, another Japanese karaoke Korean keyboard variant. Just how many of these are there?

This is like the third variant I've seen at the same karaoke chain. Otherwise the UI has been mostly the same across different locations.

#korean #keyboard

I can't be stopped

I can't be stopped

I spent a bit of time last night and this morning over-optimizing a naive #Python #LeetCode solution to get the fastest solution on the...

I spent a bit of time last night and this morning over-optimizing a naive #Python #LeetCode solution to get the fastest solution on the site.

Enjoy: https://theoreticallygoodwithcomputers.com/posts/leetcode-gold-optimization/

Starting in 5 minutes!

Starting in 5 minutes!

Today in the Formal Languages and Neural Networks (FLaNN) seminar, we have Nur Lan presenting `Minimum Description Length Recurrent Neural...

Today in the Formal Languages and Neural Networks (FLaNN) seminar, we have Nur Lan presenting `Minimum Description Length Recurrent Neural Networks`

https://aclanthology.org/2022.tacl-1.45/

Starts at 3PM GMT (Monday)!

https://flann.super.site/

#ML #Interpretability #NeuralNetworks

The Elements of Differentiable Programminghttps://arxiv.org/abs/2403.14606Looks like an incredible resource.#ML

The Elements of Differentiable Programming

https://arxiv.org/abs/2403.14606

Looks like an incredible resource.

#ML

As discussed in the original paper, Rényi Efficiency does a pretty good job of predicting relative downstream performance, but there were...

As discussed in the original paper, Rényi Efficiency does a pretty good job of predicting relative downstream performance, but there were effects that it could not capture. Our paper gives explicit examples of some of those effects.

However, it doesn't mean that Rényi Efficiency is useless. One of our counterexamples is very unnatural and the other is also probably not useful IRL. That Rényi Efficiency fails on these isn't necessarily indicative of how it fares on other realistic tokenizers.

The original paper was written by our second author (also found here https://twitter.com/zouharvi)....

The original paper was written by our second author (also found here https://twitter.com/zouharvi).

https://aclanthology.org/2023.acl-long.284/

The goal was to find which metrics best correlated with downstream performance (e.g. BLEU).

Our counterexamples allow us to arbitrarily increase Rényi Efficiency while _decreasing_ BLEU.

It turns out these counterexamples also break other popular metrics (e.g. average tokenized sequence length).

My recent paper, Two Counterexamples to *Tokenization and the Noiseless Channel*, was accepted to LREC-COLING!We look at a recent metric,...

My recent paper, Two Counterexamples to *Tokenization and the Noiseless Channel*, was accepted to LREC-COLING!

We look at a recent metric, Rényi Efficiency, for *intrinsically* evaluating tokenizers. That is, how can we determine if a tokenizer will be good without having to train a full model (or "metric goes up implies BLEU goes up")?

We design two tokenizer families that break the metric.

https://arxiv.org/abs/2402.14614

Joint work with @zouharvi @s and https://twitter.com/chokkanorg

#nlproc #NLP

Some #HPC and #Julia internships at the Swiss National Supercomputing Center:https://jobs.ethz.ch/job/view/JOPG_ethz_sRMTIUMhsyaWMB4LVv

Some #HPC and #Julia internships at the Swiss National Supercomputing Center:

https://jobs.ethz.ch/job/view/JOPG_ethz_sRMTIUMhsyaWMB4LVv

This month in #Julia https://discourse.julialang.org/t/this-month-in-julia-world-2024-01/109549#JuliaLang

This month in #Julia

https://discourse.julialang.org/t/this-month-in-julia-world-2024-01/109549

#JuliaLang

A really nice list of Gaussian Splatting research and implementations: https://huggingface.co/spaces/dylanebert/research-trackerCompiled by...

A really nice list of Gaussian Splatting research and implementations:

https://huggingface.co/spaces/dylanebert/research-tracker

Compiled by https://twitter.com/dylan_ebert_ from #HuggingFace

#GaussianSplatting #ML