Artificiality Bites đź’Š Issue #13
Hello Human! This is a new issue from my weekly newsletter, holding a tiny compilation made of interesting articles from last week, projects, tutorials and tools; all related to Data, Artificial Intelligence and adjacent topics. Bon appetit!
đź“ť Interesting articles this week
- Measuring Gendered Correlations in Pre-trained NLP Models (Google AI)
9'
Researchers have found that NLP models can behave in an undesired way in many applications, such as professions correlating with one gender more than another. This article explore Google's approach to evaluate and mitigate unintended correlations in pre-trained models. - Fast reinforcement learning through the composition of behaviours (DeepMind)
26'
A major limitation in Reinforcement Learning is that agents currently require an insane amount of training data and time. This article tells how we could use the knowledge acquired in previous tasks to learn a new task more quickly, like you humans do. - Nemo: Data discovery at Facebook (Facebook Engineering)
8'
Facebook engineers share its experience building its own data discovery solution. - How AI is powering a more helpful Google (Google)
8'
How Google is applying AI into its products to improve searching. - Microsoft explains how it improved automatic image captioning
4'
Microsoft launched a new computer vision service claiming it can generate image captions that are, in some cases, more accurate than human. Read more at Microsoft Blog.
🔧 Tutorials
- Structural Time Series (Cloudera Fast Forward)
44'
How to model time series with some underlying periodic patterns by breaking down it into several components. - What Color Is This, Part 2: The Computational Parts
13'
A case study on how to extract the colors in clothes pictures using computer vision.
📦 Repositories
- timoschick/pet
Python implementation of Pattern-Exploiting Training (PET), a semi-supervised training procedure for NLP which outperforms regular supervised training, various semi-supervised baselines and even GPT-3 despite requiring 99.9% less parameters. - nyu-mll/jiant
Jiant is a NLP research library designed to facilitate large-scale, replicable, configuration-driven experiments, allowing multi-task training and transfer learning. It’s been recently upgraded to version 2.0. - lightly-ai/lightly
Lightly is a computer vision framework built on top of PyTorch for training deep learning models using self-supervised learning. It can be used for applications such as nearest neighbors or similarity search, transfer learning and data analytics.
🎓 Courses / Events / Books
- CS 685: Advanced NLP (University of Massachusetts Amherst)
Lesson recordings and additional material from the course taught by Mohit Iyyer at UMass (Fall 2020) is being publicly shared. - Putting ML in Production (Made with ML)
A hands-on course on MLOps / End-to-End Machine Learning, starting this week. - Heroes of NLP đź“ą (Andrew Ng)
A deeplearning.ai interview series featuring Andrew Ng in conversation with industry and academic experts. Four videos available so far. - Open Source Tools & Data for Music Source Separation đź“•
An interactive book exploring modern open-source tooling and datasets for running, evaluating, researching and deploying music source separation approaches (expect some Deep Learning!).
🚀 Extra bits
👉 Newsletter en Español
đź‘‹ See you next week!