A selection of articles on machine learning: case studies, guides and research for April 2020

We continue to select publications that help lower the threshold for entering the ML sphere. As before, it primarily compiles open source tools, pre-trained models, and high-level APIs.

Blender

Facebook AI introduced the largest open - source chatbot Blender . It is based on a model with 9.4 billion parameters, which makes it the largest system of existing. This is the first open-domain chatbot that combines a variety of conversational skills in one system: it is able to express empathy and show individuality. According to the authors of the study, Blender is in all respects better than Meena, which Google announced earlier this year , but has not been demonstrated.

Background matting

Machine learning is often used to remove background from images. If with static images the results have long been very good, and there are even whole services like remove.bg, now this can also be achieved for video - the chromakey is no longer needed! A free tool has appeared that does an excellent job of removing the background from the video. For the desired result, you need to take a background picture without a person, and then machine learning algorithms do their magic. The results are amazing, in the publication you can see the benchmarks.



De-occlusion

Existing machine learning algorithms are able to analyze only the visible parts of objects. This leads to an incomplete interpretation of the scene.
Now the open source framework has appeared, which is able to complement hidden from view fragments of objects in the image. The tool is based on a model trained without the involvement of a teacher.



TensorFlow Profiler

Performance is a key factor in machine learning research. The faster the model is trained, the more iterations can be made, reducing overhead. This is very important in industrial development. However, it is not always clear what should be optimized, and it takes time to search for narrow necks. Now for TensorFlow there is a set of tools designed to solve this problem.

Quant noise

Modern machine learning models are becoming more voluminous and contain millions of parameters. However, there is an urgent need to run these models on weak devices. In an attempt to resolve this contradiction, an open source tool has appeared that provides maximum compression of models with virtually no loss in performance. In the future, this will allow you to run applications locally on mobile devices and IoT-chipsets.

TensorFlow Lite

However, there are already many models that are optimized for weak devices. This publication shows how, having no experience in machine learning, using TensorFlow Lite to assemble a complete product. The number of pre-workout models in the repository is constantly replenished, so nothing prevents you from creating mobile applications now, in the heart of which there will be machine learning models.

Style transfer

Remember the Prisma application, which transferred the art style to the photos of users using a neural network? You won’t surprise anyone now, but now you can assemble your own prism for Android and iOS. The publication describes how this technique is optimized for TensorFlow lite to be supported by not-so-powerful mobile devices. By the way, in April, the same feature appeared in the Google Arts & Culture app.



Bonus:

Stanford University posted an open access lecture course in 2018, which can now be viewed on Youtube .

Instead of a conclusion:

Based on the libraries from the March selection, we put together a small project that allows you to control the web interface using an ordinary webcam, and talked about it in detailon Habré. Perhaps the tools in this collection will also inspire some readers to come up with a solution to some actual problem. It will be great to read about this. In the meantime, that's all, thank you for your attention!