Data Science Digest

2021-06-30 13:05:59 The FLORES-101 Data Set: Helping Build Better Translation Systems Around the World

Building on the success of machine translation systems like M2M-100, Facebook AI has open-sourced FLORES-101, a many-to-many evaluation data set covering 101 languages from all over the world, to enable researchers to rapidly test and improve upon multilingual translation models like M2M-100. In this article, you’ll delve into its basics.

https://bit.ly/3hja98z Open / Comment

2021-06-29 19:01:57 Consistent Instance False Positive Improves Fairness in Face Recognition

In this paper, Xingkun Xu et al. propose a false positive rate penalty loss, a novel method to mitigate face recognition bias by increasing the consistency of instance False Positive Rate (FPR). The method requires no demographic annotations, allowing to mitigate bias among demographic groups divided by various attributes.

Paper — https://bit.ly/361fHiQ
Code — https://bit.ly/2UKAqoF Open / Comment

2021-06-27 10:00:06 DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification

In this research, Yongming Rao et al. propose a dynamic token sparsification framework to prune redundant tokens progressively and dynamically based on the input. A lightweight prediction module can estimate the importance score of each token given the current features. The module is added to different layers to prune redundant tokens hierarchically.

Web Page — https://bit.ly/3dgESlq
Paper — https://bit.ly/3xWf5XJ
Code — https://bit.ly/3jlUtUK Open / Comment

2021-06-26 10:00:05 MLOps Toys
The platform is a collection of MLOps projects by category, including data versioning, training orchestration, feature store, experiment tracking, model serving, model monitoring, and explainability.

https://bit.ly/3xV97GF Open / Comment

2021-06-25 17:00:08 #DataScienceMemes Open / Comment

2021-06-25 12:36:38 Ingestion and Historization in the Data Lake

In this video, Alexey Grigorev, the founder of DataTalks.Club, hosts Illia Todor, Data Engineer, to talk about ingestion and historization of data in the data lake.

https://bit.ly/3zWKaMP Open / Comment

2021-06-24 12:48:10 ​​Data Science Digest — 24.06.21

The new issue of DataScienceDigest is here! The impact of NLP and the growing budgets to drive AI transformations. How Airbnb standardized metric computation at scale. Cross-Validation, MASA-SR, AgileGAN, EfficientNetV2, and more…

https://bit.ly/3qnuy0u

Join @DataScienceDigest Open / Comment

2021-06-23 09:30:04 AI Can Now Emulate Text Style in Images in One Shot — Using Just a Single Word

In this article, the engineering team of Facebook AI presents TextStyleBrush, an AI research project that can copy the style of text in a photo using just a single word. With this AI model, you can edit and replace text in images. The team hopes to spur dialogue and research into detecting potential misuse of this type of technology, so make sure to contribute.

https://bit.ly/3wU3t7I Open / Comment

2021-06-22 10:00:05 How Airbnb Standardized Metric Computation at Scale

The engineering team of Airbnb reveals the design principles of Minerva compute infrastructure. Minerva is a single source of truth metric platform that standardizes the way business metrics are created, computed, served, and consumed. The article features the link to the first post on Minerva. Check it out, too!

https://bit.ly/3zLDv8g Open / Comment

2021-06-21 10:00:09 Comparing Test Sets with Item Response Theory

In this paper, Clara Vania et al. use the Item Response Theory to evaluate 29 datasets using predictions from 18 pretrained Transformer models on individual test examples. Quoref, HellaSwag, and MC-TACO are best suited for distinguishing among state-of-the-art models, while SNLI, MNLI, and CommitmentBank seem to be saturated for current strong models.

https://bit.ly/3xHDkJj Open / Comment