squigglysounds

Data scientist and musician. I like to make weird noises with computers.

[squigglysounds @squigglysounds]

@Bitwig The new beta does not connect to my existing jack server. Pretty basic workflows are broken. First time I've had a poor experience with a Bitwig beta release.

[squigglysounds @squigglysounds]

I really enjoyed this post about why linear loss mixtures don't always behave like you'd think they would. Although the examples hint at big models, this feels relevant for more everyday / small data modelling problems too.

[317070 @317070]

We wrote a blogpost on why so many machine learning algorithms are so hard to tune, and how an old Nips'88 paper shows how to fix it. It has some nice graphics to visually explain what is going wrong, and the proposed solution is simple to implement. https://t.co/fRd7aXJLaV

[squigglysounds @squigglysounds]

@OpenAI The people saying we badly need generative models that can attribute credit to the right examples from the training data are right. There is huge data laundering potential with a huge model like this.

[squigglysounds @squigglysounds]

@OpenAI I want to hate this result because it takes ML yet another step away from something you can do with a sane budget. And yet the results are very impressive.

[squigglysounds @squigglysounds]

@jordiponsdotme Specifically thinking of the difficulties GANs have with hair and other textures with both high-frequency content and complex structure.

[squigglysounds @squigglysounds]

@jordiponsdotme Just got through @jordiponsdotme's video lecture on upsamling artifacts. I think the frequency responses of upsamplers would be valid in 2D as well. Do we get these strong frequency peaks / notches in image synthesis too?

[squigglysounds @squigglysounds]

Big 👏👏👏 to @ismir2020 organizers, I could never have gotten this engaged with people's work at an in-person conference! Online videos ahead of time and chat rather than Q&A in a crowded room are both totally game changing improvements to the pre-2020 conference routine IMO.

[squigglysounds @squigglysounds]

This should be on the wikipedia for KLD. Super helpful intuition!

[Ari Seff @ari_seff]

KL divergence asymmetry

[squigglysounds @squigglysounds]

Besides impressive results, the implementation is high quality. Concise and really cleanly written code, good stuff!

[Jun-Yan Zhu @junyanz89]

Code for Contrastive Unpaired Translation (CUT): https://t.co/F4TxM8IGUF. Contrastive learning (instance matching) and adversarial loss (distribution matching) are complimentary for many conditional image synthesis tasks. (w/ Taesung Park, @rzhang88, Alexei A. Efros, ECCV 2020)

[Jascha Sohl-Dickstein @jaschasd]

Whitening and second order optimization both destroy information about the dataset, and can make generalization impossible: https://t.co/CuDeHxF90r We examine what information is usable for training neural networks, and how second order methods destroy exactly that information.

[David Braun @DoItRealTime]

New repo! DawDreamer is an audio-processing Python framework supporting core DAW features such as audio playback, VST MIDI instruments, and VST effects. Tested on OSX and Windows!

https://t.co/1ZJV2tMyWa

[Kyle Cranmer @KyleCranmer]

Very excited to share our newest paper
combining machine learning & physics. We develop normalizing flows that impose the elaborate of symmetry groups you find in fundamental particle physics.
It's a beautiful mix of math, ML, and physics
https://t.co/1keqEEEYM3

[squigglysounds @squigglysounds]

I have been hesitant to get on the GPT-3 hype train but I have to admit I'm impressed by this...

[Raphaël Millière @raphaelmilliere]

I asked GPT-3 to write a response to the philosophical essays written about it by @DrZimmermann, @rinireg @ShannonVallor, @add_hawk, @AmandaAskell, @dioscuri, David Chalmers, Carlos Montemayor, and Justin Khoo published yesterday by @DailyNousEditor. It's quite remarkable!

[Juan Nunez-Iglesias @jnuneziglesias]

Wow! This #SciPy2020 talk by Jim Pivarski on Awkward Array is ***awesome***. The power and speed of NumPy but on jagged nested data.

https://t.co/a4iml6IQfT

I hate to spoil it but they even provide a @numba_jit extension so you can write fast custom analysis code, too! 🤯

[squigglysounds @squigglysounds]

I enjoy this every time it comes up https://t.co/la0YjbV6UN

[squigglysounds @squigglysounds]

@drscotthawley I wanted to take a look at the blog post but the link 404s for me. I think it might be this one? https://t.co/C1052sOofV

[squigglysounds @squigglysounds]

2020 continues to be the year of the contrastive representation in ML. They are finding their way into more very hard applications.

[Daniel Yamins @dyamins]

1/7 A big problem with deepnet models of the brain is that they require training on huge supervised datasets. So even if they are approximations of neural responses in the "adult animal", the training process is a totally implausible model of learning in real visual development.

[thedotisblack @DavidMrugala]

Generative Portraits 1
Revisiting the code for some new visuals. It's always exciting to work with it.
#Processing #CreativeCoding #generativeart https://t.co/hs2GxB6xEe

[squigglysounds @squigglysounds]

@theshawwn Wild as this is since porn has a pretty systematic intended audience don't you think you could get into trouble biasing your datasets?

[squigglysounds @squigglysounds]

@theshawwn This idea is straight out of a scifi novel wow