A selection of articles on machine learning: cases, guides and studies for May 2020

We continue to collect for you the most interesting news and tools from the field of machine learning, written in an accessible language.

Jukebox

Earlier this month, OpenAI aroused great interest in the ML community by posting the source code for its project called Jukebox. This tool using machine learning algorithms allows you to generate compositions of popular artists. On the network you can already find examples of tracks generated by users, there are very unusual combinations.

AR Copy paste

French developer Cyril Diagne introduced an AR application that takes pictures of objects, removes all unnecessary background from pictures and (using U ^ 2-Net) transfers the result to programs running on the computer. For example, the author shows how, using the application, you can quickly select and add illustrations to a presentation. You can already look at the code and sign up for early access to the application, which is currently under development.



Pose Animator

An open source tool for web animations with which you can animate characters drawn in SVG. The tool is based on two other libraries Facemesh and PoseNet , which use a webcam to capture movements. The article shows how the tool was created and demonstrates how to use it.



Galaxy zoo

A case study on how to successfully combine crowdsourcing and machine learning to quickly process complex information. The Galaxy Zoo platform has combined these two approaches to study the evolution of galaxies by classifying millions of images. The material describes how to mark only the data that will best help improve the existing model.

DistilBERT

Startup Hugging Face shares its experience in creating a public API with which you can optimize the performance of NLP models on Node.js.

Clinical Trial Parser

Most clinical trials fail to recruit enough participants. This is due to the fact that people without honey. Education do not always understand the selection criteria and the details of research. Facebook introducedopen source tool that aims to solve this problem.

GrokNet

Facebook has announced a universal computer vision system designed for ecommerce. She is able to identify the attributes of goods in different categories, from auto to home decor. The article describes how the model was trained on seven datasets, and what difficulties it encountered. A series of videos also show how this model will help the platform change online trading.

S2IGAN

It seems we promised not to include materials that contain links to empty repositories in the collection, but I want to make an exception. S2IGAN is a framework that translates speech into images. Using a dual encoder, a model has been created that generates images using a voice description. The authors of the study promise to post the code soon, but for now they have to be content with examples of generated birds.



Consistent Video Depth Estimation

An algorithm is presented that, using a convolutional neural network, determines the depth of individual frames and restores a geometrically consistent depth for the entire video. This helps to cope with a number of limitations, for example, when the image is unstable due to shaking. You can apply a similar technology in various fields, for AR-effects or autopilot cars. The repository is still empty, but the authors promise to share the source code.

That's all, thanks for watching!