Matthias Feys

Data geek @ml6team & @GoogleDevExpert for ML and GCP. Passionate about gaining insights from data - ML, NLP and Python

[apolinario (poli) @multimodalart]

Big release that got under the radar 📡📟: last week @ml6team released Fondant-25M: dataset of image-text pairs with a @creativecommons license

And that's just the tip: they are working on a 500M one 🤯

Blog: https://t.co/HgJ9BbHOLa

Dataset on 🤗 https://t.co/fgdcm888dx

[Niels Rogge @NielsRogge]

Data preparation is going to be the #1 differentiator when it comes to Foundation Models.

Hence I've been involved in a new open-source project aimed at building & sharing reusable components for data preprocessing: https://t.co/9I7qqnLrz5

1/2

[Jacob Kimmel @jacobkimmel]

building a company on large ML models is humorously closer to biotech than traditional software

- niche, shallow talent market
- long iteration times
- material capital requirements
- likely need an incumbent partner for distribution

[Apache Beam @ApacheBeam]

Discover how #ApacheBeam's multi-language pipelines allow you to utilize various I/O connectors and transforms without needing to reinvent the wheel to your language of choice in this post from @ml6team

https://t.co/E3EUCP7pWB

#datapipelines #python #java #expansionservice

[ML6 @ml6team]

October is going to be an eventful month for the @ml6team. Join us in one of the following events: https://t.co/HkIYsvYIzv
See you there ! #ml6 #accelerateintelligence #meetup #techtalks #events

[clem 🤗 @ClementDelangue]

Cool to see @ml6team on @huggingface! https://t.co/RqFO67oR5L

[Matthias Feys @FsMatt]

@WardVDriessche @YvesHanoulle @CoderDojoBe Zou inderdaad moeten lukken. DM gerust je mailadres zodat we kunnen afspreken wat mogelijk is!

[Matthias Feys @FsMatt]

@317070 Wow! That's a pretty sci-fi secret project indeed😀!! Awesome to see those 2 worlds meet!

[ML6 @ml6team]

Still haven't found the perfect valentine gift? Send your loved one a unique (free) love meme with your face embedded and remain in their heart always!

Link to the meme swap by Gener8 👉 https://t.co/jAmU5enPXb

#gener8 #generativeai #ml6 #ml6team

[ML6 @ml6team]

Today, the #Connexion framework got a new community-owned home! 🏠 Proud to continue the work of @ZalandoTech and take over the regular maintenance of the project as part of the spec-first organization. #opensource #Engineering #github

[ML6 @ml6team]

Today is marked as #WorldCancerDay
#ML6 always looks for projects that bring social value. Thanks to the support of @VLAIO_be & flanders.healthTech we will push the current technical capabilities of leveraging data & #ML to improve and personalize prostate cancer treatment plans

[Robin De Neef @RobinDeNeef]

Security doesn't have to look boring! 😴

Check out this blogpost on how we visualise our Cloud Armor logs and protect our #GoogleCloud infrastructure! 🛡️

[ML6 @ml6team]

#2021wrapped ✨ We are happy to have grown our team with 33 new joiners, and are now set to serve even more customers with great technology across our 5 international markets. Thankful for what 2021 brought us, and hopeful that 2022 will bring us much more of that. #ml6

[Matthias Feys @FsMatt]

Explainability as key enabler for more impactfull AI projects. Why & how illustrated with use case examples: https://t.co/qdU1Fd0LC3

Based on the work by @DhruvBatraDB
#xai #explainability

[Thomas Dehaene @TDehaene]

ByT5 by @GoogleAI opens a lot of cool doors 🚪, such as correcting common #OCR mistakes 💡. Check out our finetuned model and a demo on @huggingface spaces: https://t.co/d5RoIzSeFa

[ML6 @ml6team]

#OCR errors are a common side-effect when processing scanned documents 🔍🖨️. In our latest #NLProc blogpost, @simon_de_gheselle investigates the use of @GoogleAI’s #ByT5 to correct these errors 💡👆!
Full blogpost: https://t.co/49eXe2zYHw #ml6 #transformers #huggingface

[Matthias Feys @FsMatt]

@peeterskris For most projects we deploy, we can work with a serverless alternative such as Vertex AI, still using the same concepts & framework, but cheaper to run & easier to maintain. So the biggest downside is indeed the heavy architecture which isn't always needed.[2]

[Matthias Feys @FsMatt]

@peeterskris Kubeflow is still mainly Google sponsored & both David Aronchick & Jeremy Lewi (~creators) left Google, so this certainly has an impact. Otherwise the serverless options are getting much more popular & cheaper! [1]

[Erik Bernhardsson @bernhardsson]

I noticed a weird thing about writing that if you’re trying to prove a point and use too many arguments (beyond 2-3) then it ends up coming off as weaker in aggregate. Is this a real thing?

[Sergey Levine @svlevine]

To make an existing model more robust at test time: augment a single test image in many ways, finetune model so that predictions on augmented images "agree", minimizing marginal entropy. This is the idea behind MEMO (w/ Marvin Zhang & @chelseabfinn): https://t.co/UjO6oJIXs8

🧵>