nicholdav

@nicholdav@neuromatch.social

Pore over everything in my C.V.
But you'll still know nothing 'bout me.

Lapsed neuroscientist, budding biologist, mediocre programmer. Trying to make it easier to understand how animals communicate.

This profile is from a federated server and may be incomplete. Browse more on the original instance.

nicholdav, to random

Next @us_rse
neuro-RSE meeting this coming Mon 12p EST / 9a PST

  • Ryan Ly will give a talk on @NeurodataWB

  • @yarikoptic
    will update us on BIDS after the
    maintainers summit

We'll also discuss paths for RSEs in neuro

Reply/DM for invite! Boosts appreciated! (@jonny @elduvelle @emdupre @billbrod @arokem et al 🙏 )

nicholdav,

@jonny @us_rse @yarikoptic I'll add you to the invite!

nicholdav, to Neuroscience

Hi neuro friends: next Mon at 12p EST / 9a PST the neuro-rse affinity group of @us_rse will hold our monthly meeting

If you're not familiar with research software engineering (#rseng) it's just what it sounds like -- and you've probably noticed how crucial software is for neuro (and, like, all other fields)

Since we are US-RSE we are a bit US-centric but we just generally want to support anyone working with software for #neuroscience

We'll have lightning talks from Olivier Winter on how the #InternationalBrainLab does data infra and from @adamltyson on @brainglobe

Boosts appreciated; please feel free to reply and I can add ppl to the invite
@jonny @elduvelle @emdupre @patrickmineault etc 🙏

nicholdav, to random

Passing on from the @us_rse neuro-RSE affinity group that there will be a brainhack before cosyne

https://pre-cosyne-brainhack.github.io/hackathon2024/

Pass the word plz @jonny @elduvelle @neuralreckoning -- I don't find the organizers on masto but feel free to pls also correct me if I'm wrong

nicholdav,

@elduvelle @us_rse @jonny @neuralreckoning ofc, good idea to add the tag, done, thank you

nicholdav,

@elduvelle @us_rse @jonny @neuralreckoning close enough 🌝

nicholdav,

@elduvelle @CosyneMeeting @us_rse @jonny @neuralreckoning good catch, thanks -- added 🙂

jonny, to random
@jonny@neuromatch.social avatar

I know the answer to this is probably "thats exactly the opposite of what they're supposed to do," but the thought has been sticking with me - I am wondering if there is some kind of "semantic hash" algorithm, sort of like vector embeddings, that puts hashes nearby in hash space that are nearby in source space.

Usually you hash over undifferentiated binary, but say you were instead hashing over a graph where you wanted similar graphs (similar terms, structure) to have close but unique hashes, but still collision resistant in the sense that you couldnt trivially tell which terms were giving you proximity (and thus engineer a collision).

This feels like a thought that will get me fedi spanked by cryptographers

nicholdav,

@irenes @jonny @jonny sounds like you might be headed down this path already so sorry if I'm telling you stuff I'm already know, but it definitely seems like you are describing deep hashing with neural networks.
When you say "semantic hash, sort of like vector embeddings" I immediately think of deep hashing for info retrieval.
See https://dl.acm.org/doi/full/10.1145/3532624 (ofc there's always a recent "survey" paper on deep whatever :/ )

This paper doesn't include graphs per se but has some of the flavor of what it seems like you might want:
https://ojs.aaai.org/index.php/AAAI/article/view/6121
Not sure if you already know the deep hashing lit but if not the intro seems to ref most of the major players as you will see by comparing with that survey.

I think it will be hard to get any guarantees on whether or not you can engineer a clash with deep hashing though, because of the curse of dimensionality. Things are weird in high-D and it can be hard to enforce what's next to what. You can force a clash with adversarial methods, see for example https://github.com/anishathalye/neural-hash-collider

nicholdav,

@jonny @irenes
@jonny @irenes I hear you. I am not a crypto person so I can't fedi spank you but it does sound like "preserve local/global topology" and "collision-resistant hashing" are at odds with each other, a tough optimization problem.
Does sound like fun reading though 🙂

nicholdav, to random

Hello, I'm giving a talk at PyDataGlobal 12-1230 EST tomorrow on VocalPy, a core Python package for acoustic communication research
vocalpy.readthedocs.io/en/latest/

Virtual, cheap tix 🙂
pydata.org/global2023

Boosts greatly appreciated 🙏 @jonny @dingemansemark @danstowell

jonny, to random
@jonny@neuromatch.social avatar

OK i'm on the very first page of this from @dingemansemark and already i know i'm gonna love it. This is exactly what i was thinking about in the context of phonetic perception.

that if the modality isn't assumed to be error free, that communication was always interactive and imperfect, that we would study phonetics in a very different way as well - the haskins lab style psychoacoustic cue model that got imported into systems neuroscience (i think bc compatible with our behavioral methods, but digression) makes very little sense theoretically and empirically.

the capacity for repair in multiple dimensions - across time (asking for repetition) and frequency space (weighting different acoustic cues to accomodate for speaker/environment/etc. variation) - is definitive of spoken language and not captured by reduced synthesized pseudo-speech fragments. Perfect example of how reductionism can go beyond simplifying into inverting a problem.

curious to read more of this paper that starts with 'an inversion of perspective'

https://pure.mpg.de/rest/items/item_3530697_2/component/file_3547372/content
from
https://scholar.social/@dingemansemark/111475653592245864

nicholdav,

@jonny glad you're sharing this, it overlaps with this I'm reading today: "Overcoming bias in the comparison of human language and animal communication"

https://www.pnas.org/doi/10.1073/pnas.2218799120?TOC_v120_i47=

cites @dingemansemark

I need to read this now so I don't commit the sin of not being precise enough when I talk about turn-taking!

jonny, to random
@jonny@neuromatch.social avatar

I am re-reading the MATLAB package that caused me to quit being a neuroscientist and focus on how broken scientific infrastructure was (the first time), and i found my favorite bug of all time -

code for automated, trial-based behavioral neuroscience experiment:

  • query a database with hardcoded credentials every time you want to calculate reward/reinforcement size. reward is delivered once per trial max.
  • yes, hardcoded credentials use the password "password" and formerly exposed a root-level postgres user on public VCS.
  • query is structurally pointless: the query checks that a value hasn't changed - the class is named "constantReinforcement"
  • db connection has a 1s login timeout. connection always times out because db hadn't existed for ~6 years
  • if db connection takes longer than 0.3s, send an email to code author saying database access is slow
  • try to send email with an SMTP server without authentication, remarkably this works for awhile. SMTP server changes to require authentication.
  • SMTP connection times out.
  • default timeout is 100 seconds.
  • rewards in fact calculated between every phase of every trial, not once per trial
  • 5 phases in trial (6 phase transitions since trial transitions also count as a phase): pre-stimulus, stimulus presentation, pre-reward, reward delivery, post-reward

so our poor mice were waiting 600 total seconds, 10 minutes, to complete a single trial which should take ~1s. it took two weeks to debug the problem. this was one of about 2 dozen bugs of this nature that defined my first year or two of grad school.

since this I have seen far worse code in active use running live neuroscientific experiments that get published in the top-tier journals.

when i first started writing experimental code to fix this, one of the most famous ppl in my field told me not to waste my time because nobody really cares if the code is correct as long as it produces data. i stopped believing all papers and doing neuroscience shortly after.

nicholdav,

@jonny

> when i first started writing experimental code to fix this, one of the most famous ppl in my field told me not to waste my time because nobody really cares if the code is correct as long as it produces data.

Oh man I hear that

One of my heroes who wrote software that everyone in my subfield uses told me:
"Code is like underwear. Everyone's is dirty and no one wants to borrow yours."

This bothered me so much that I organized a session at a satellite about "songbird data science" where this person was present, and I quoted them (without naming them) and then said "so we've all been wearing your underwear and you didn't tell us it was dirty?"

This was in a little speech where I was trying to motivate people to work together, and instead I probably just wounded a person who had good intentions and made free software long before we had tools and infra like GitHub. A true hero developer (not that I think we need hero developers. I'd rather have a life.)

Uh so anyways I get where you're coming from

nicholdav, to random

Are you using neural network models to segment and/or annotate animal sounds?

You'll want to check out our poster HH22 comparing models that I present 1-5 today.

Sneak preview here: https://github.com/vocalpy/Nicholson-Cohen-SfN-2023-poster (I'll post a thread about it later).

Reposts + any feedback appreciated @jonny @dingemansemark @danstowell other segmentation fans 😼 @Iris @fedeadolfi

nicholdav,

@jonny 🙌 awesome, I saw you were here and was going to say we should meet up. Will be great to meet you in person

elduvelle, (edited ) to random
@elduvelle@neuromatch.social avatar

How’s your SFN poster preparation going?
#SFN23 #NeuroConf

nicholdav,

@elduvelle loved when they wanted me to upload a virtual poster weeks ago

Lol yeah right

nicholdav,

@elduvelle I really do want to make an effort. I wish they could have a system that's more flexible though, kinda like the way Neuromatch did flash talks. At SciPy conference we have people post virtual posters in the channel during the conference. Ofc, SciPyConf is not as big as SfN ... but Neuromatch was pretty dang big and managed to make something more fluid work!

nicholdav, to random

I made a PR fixing an example of vectorizing the Mandlebrot set in Numpy (adapted from a 2009 @neuralreckoning blog post!) to @rougier's excellent book "From Python to Numpy"
https://www.labri.fr/perso/nrougier/from-python-to-numpy/

and it got merged 😎​
https://github.com/rougier/from-python-to-numpy/pull/113#issuecomment-1761277546

Will this be my most impactful open source contribution? 🥲

nicholdav,

@jonny @neuralreckoning lol

"I have changed but one line, yet achieved enlightenment in the path I took to change it 🧘 plz merge"

nicholdav, to scipy2023

At @scipy2023 I had a convo with a friend about Python tools for tracking computational projects / experiments / provenance, like
http://neuralensemble.org/sumatra/
and
https://recipy.readthedocs.io/en/latest/

Are you using tools like these? Are there new ones I should know about? Have they all been displaced by ML/data sci industry tools like https://mlflow.org/ and https://dvc.org/?

@jonny @neuralreckoning @baratzz @elduvelle @emdupre

nicholdav,

@jonny @scipy2023 @neuralreckoning @baratzz @elduvelle @emdupre thank you @jonny for a thoughtful answer as usual, you are completely right that I am not thinking enough about the end state of a world where everyone uses these tools -- what I want is "git for experiments" but what I might end up with is "pubpeer for reviewing my terrible bash scripts" or something like that

I should play more with a wiki.
I want a blog I can write in MyST that is also a wiki 😕

jonny, to random
@jonny@neuromatch.social avatar

basically this is what I love music to be
https://youtu.be/O2F0oTqfL3E

nicholdav,

@jonny fuck yes this is so good

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • provamag3
  • InstantRegret
  • mdbf
  • ethstaker
  • magazineikmin
  • GTA5RPClips
  • rosin
  • thenastyranch
  • Youngstown
  • osvaldo12
  • slotface
  • khanakhh
  • kavyap
  • DreamBathrooms
  • JUstTest
  • Durango
  • everett
  • cisconetworking
  • Leos
  • normalnudes
  • cubers
  • modclub
  • ngwrru68w68
  • tacticalgear
  • megavids
  • anitta
  • tester
  • lostlight
  • All magazines
    •