Papers of the day   All papers

Overton: A Data System for Monitoring and Improving Machine-Learned Products

Comments

Julien Chaumond: Apple, are you trying to bankrupt us? At ~500 MB per average model download, 90k hits (Apple's IP addresses are the 17. block) translate to ~45TB of downstream bandwidth… which starts being costly for us. Your new ML platform Overton (https://arxiv.org/abs/1909.05372) ⤵️ https://t.co/JS7tnRfTjt

22 replies, 547 likes


Hazy Research: Congrats @vincentsunnchen !!! Ideas used in Overton (https://arxiv.org/abs/1909.05372). Apple is awesome for ML! Side note: submitted Overton Arxiv paper myself (it's been a while)... remembered my password from grad school...victory for all faculty, and subject of today's group meeting.

0 replies, 37 likes


Nirant: Overton, a data system for doing model lifecycle management for ML& NLP @Apple -- builds on PyTorch-Transformers So much better explanation of motivations and things they thought about than Ludwig by Uber. https://arxiv.org/pdf/1909.05372.pdf

0 replies, 33 likes


bayo adekanmbi: AI is getting better! This is an “auto ML” tool but a highly scalable one. With this, we can improve Machine learning products by focusing on schema/data and leave the coding part to the system! Indeed, we get more when developers focus on higher-level tasks & not coding 👍👍

0 replies, 32 likes


Ben Lorica 罗瑞卡: Overton, @Apple's machine learning platform ("support engineers in building, monitoring, and improving production machine learning systems") cc @HazyResearch https://arxiv.org/abs/1909.05372 https://t.co/uEPsOcStDc

1 replies, 28 likes


arxiv: Overton: A Data System for Monitoring and Improving Machine-Learned Products. http://arxiv.org/abs/1909.05372 https://t.co/lv3vAovMNk

0 replies, 20 likes


Hacker News: Apple downloads ~45 TB of models per day from our S3 bucket : https://twitter.com/julien_c/status/1173669642629537795 #Apple #S3 Comments: https://news.ycombinator.com/item?id=20987292

1 replies, 19 likes


arXiv Daily: Slice-based Learning: A Programming Model for Residual Learning in Critical Data Slices https://deepai.org/publication/slice-based-learning-a-programming-model-for-residual-learning-in-critical-data-slices by Vincent S. Chen et al. including @hazyresearch #ComputerVision #MachineLearning

0 replies, 19 likes


vincent sunn chen: For more... * Paper: https://arxiv.org/abs/1909.06349 * Code/Tutorial: https://www.snorkel.org/use-cases/03-spam-data-slicing-tutorial * In production (Overton): https://arxiv.org/abs/1909.05372 With @Wu_Sen, Zhenzhen (Jen) Weng, @ajratner, @HazyResearch, @SnorkelML, and many more!

0 replies, 19 likes


DataScienceNigeria: Watch out for OVERTON from Apple. It automates AI system lifecycle by providing a set of novel high-level abstractions. Actually, you can go CODE-less to build deep-learning-based applications without writing any code in frameworks like TensorFlow Read:https://arxiv.org/pdf/1909.05372.pdf https://t.co/DHSHh5FDFM

0 replies, 18 likes


Hazy Research: @NandoDF @karpathy @SnorkelML @Google @Apple here are some... @GoogleAI https://ai.googleblog.com/2019/03/harnessing-organizational-knowledge-for.html @Apple : http://arxiv.org/abs/1909.05372 @IBMDataScience : https://arxiv.org/pdf/1812.06176.pdf @intel https://ajratner.github.io/assets/papers/Osprey_DEEM.pdf @StanfordMed : https://www.nature.com/articles/s41467-019-11012-3 more at https://www.snorkel.org/ 2/2

1 replies, 15 likes


Corey Quinn: And this is why there's a "requestor pays" setting on the bucket. Added bonus, it's cheaper for everyone since Apple's @awscloud discount is almost certainly better than most others'. :-)

0 replies, 12 likes


Carl Carrie: Ode to Overton, Apple’s machine learning lifecycle of building, monitoring, and improving production machine learning systems #ml pipeline https://arxiv.org/pdf/1909.05372 https://t.co/eAC8WUWH8c

0 replies, 12 likes


Nirant: @huggingface @Apple In the Overton paper from Chris Ré et al at Apple: Under Page 7/Section 2.4: "Major Design Decisions and Lessons" PDF link: https://arxiv.org/pdf/1909.05372.pdf

0 replies, 9 likes


arXiv CS-CL: Overton: A Data System for Monitoring and Improving Machine-Learned Products http://arxiv.org/abs/1909.05372

0 replies, 8 likes


Hazy Research: @soumithchintala @ID_AA_Carmack We failed with declarative for ML long ago ... recently gotten one \eps used (Overton/Apple, https://arxiv.org/pdf/1909.05372.pdf) similar to @w4nderlus7's awesome Ludwig https://uber.github.io/ludwig/. IMO declarative helpful when many types of users and model coding not main challenge, c.f. SQL

0 replies, 7 likes


Scott Manley: So glad that Apple really supports our scientists - Overton: A Data System for Monitoring and Improving Machine-Learned Products https://arxiv.org/abs/1909.05372

0 replies, 6 likes


Mark Wilcox: If storage services like this were paid for directly in bitcoin you would just make money when people failed to retain information they’d previously downloaded

1 replies, 5 likes


The Institute for Ethical AI & Machine Learning: Apple has released a paper where they describe their approach to monitoring and improving machine learning products through a system which they have named "Overton". https://arxiv.org/abs/1909.05372 https://t.co/KOwZLkW4E7

0 replies, 5 likes


Hacker News 250: Apple downloads ~45 TB of models per day from our S3 bucket https://twitter.com/julien_c/status/1173669642629537795 (http://news.ycombinator.com/item?id=20987292)

0 replies, 2 likes


Hacker News Feed: Apple downloads ~45 TB of models per day from our S3 bucket https://twitter.com/julien_c/status/1173669642629537795

0 replies, 1 likes


ni sinha: A nice set of lessons on software engg of DL systems from Overton. First class support for weak sup pipelines, unify prototype with deployment, supporting both core platforms.

0 replies, 1 likes


Content

Found on Sep 16 2019 at https://arxiv.org/pdf/1909.05372.pdf

PDF content of a computer science paper: Overton: A Data System for Monitoring and Improving Machine-Learned Products