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

Research work in the field of machine learning is gradually leaving the university laboratories and from the scientific discipline becomes applied. However, it’s still difficult to find relevant articles that are written in plain language and without a billion footnotes.

This post contains a list of English-language materials for January, which are written without undue academicism. In them you will find code examples and links to non-empty repositories . The technologies mentioned are in the public domain and do not require heavy-duty iron for testing.

The articles are divided into four types: announcements of open source tools, cases of using machine learning in commercial development, ML research, and practical guidelines for using PyTorch and TensorFlow.

Announcements of open source technology for machine learning

LaserTagger

The article briefly discusses the functionality and architecture of LaserTagger ; a tool that allows you to generate text in various ways - to combine and split sentences, as well as retell them in other words. Unlike seq2seq-models, it re-uses words, which ensures high speed of text generation.

Reformer

To process serial data — text, music, or video — requires a large memory resource, since this data depends on the surrounding context. Unlike Transformer, a popular architectural model for neural networks, Reformerconsumes much less memory to hold context - only 16GB. This is enough to work with large text data sets - Reformer, for example, was able to process the text “Crime and Punishment”. You can also use the Tool to “finish” the missing image fragments.

Wav2letter @ anywhere

Most real-time speech recognition frameworks use recurrent neural networks, while Wav2letter @ anywhereuses convolutional acoustic model. The framework is written in C ++, and is easily integrated into other programming languages. The creators managed to increase throughput three times and provide computational efficiency on not very powerful processors. The article provides benchmarks and links to models that will help repeat the learning outcomes.

Polygames A

framework that teaches artificial intelligence to play strategy games, even completely unfamiliar ones. Unlike other similar technologies, this model does not learn from numerous examples of successful game management, therefore, it does not require data sets.

HiPlot

Toolto visualize research data. Allows you to evaluate the effectiveness of hyperparameters, such as learning speed, regularization and architecture. The article provides snippets of python code and examples of their visualizations. HiPlot is supported by Jupiter Notebook.

Machine Learning Guides

TensorFlow.js The open

source library allows you to run machine learning models without leaving the JS ecosystem. It works in all current browsers, the server side on Node.js, and mobile platforms on React Native.

Creating a custom TFX component

Anyone can use the TFX platform to create ready-to-use machine learning pipelines . The platform offers a ton of standard components, but there are times when they are not suitable. This article talks about how to create a fully custom pipeline for working with TensorFlow.

Speed ​​up NVIDIA Dali with PyTorch

The article discusses some ways to optimize the use of DALI and create a process that uses only the CPU (without a GPU) and speeds up the machine learning process on PyTorch by 4 times.

Version Control of Data and Models for Rapid Experimentation in Machine Learning This

article describes how to create and use versioned data sets for a reproducible machine learning process. As an example, using Git, Docker, and Quilt, a deep neural network is created for object recognition using Detectron2, a PyTorch-based system that uses algorithms for object recognition.

Active Transfer Training at PyTorch

Transfer training is a process when a machine learning model created for one specific task adapts to another task. Usually it’s enough to re-train the last few layers in an existing model. The advantage is that this requires minimal labeling of data by humans. This article provides guidance on applying transfer learning to active learning, which aims to identify data patterns that need to be labeled by people.

Cases from commercial development

Creating a smart music library with TensorFlow

Developers of China's leading music streaming platform share their experience using the framework. They tell in detail what instruments were used, how they trained the model to determine the genre and style of music for subsequent recommendations. Separate the voice from the music, for example, to output the lyrics of a song using text recognition. Since some of the content on the platform was recorded by users without special sound recording equipment, the authors learned to eliminate interference and background noise.

Dog Training with Machine Learning

Obedient dogs that know the team are more likely to find new owners and leave the shelter. However, not all shelters have the ability to train dogs. The goal of the project is to train the car to understand whether the dog has completed the “Sit down” command, the dog barks or whines, etc.

Training a vacuum cleaner A

Chinese manufacturer of smart vacuum cleaners has trained devices to recognize socks and wires scattered across the floor. Case talks about the difficulties of creating a unique data set from photographs and using various machine learning algorithms.

Create a smart camera using TensorFlow Lite on the Raspberry Pi

The camera can detect a smile on a person’s face and automatically take a picture. She can be controlled by voice commands. For development, we chose the Raspberry Pi 3B +, with a gigabyte of RAM and a 32-bit operating system. TensorFlow Light has established itself as a framework for mobile and IoT devices. The article demonstrates the algorithms of the recognition system for smiles and voice commands.

Research

Neural network for solving mathematical equations

Facebook has developed the first neural network that uses symbolic logic to solve integral and differential equations of the first and second kind. It used to be that machine learning was not suitable for such a task, since mathematics requires absolute accuracy, not approximate. The authors of the study used neural machine translation. The equations were perceived by machine as text sentences, and solutions as their translation. For this it was necessary to compile syntax and prepare a set of training data from more than 100 million pair equations and their solutions. Thus, the study shows that neural networks can be used not only to distinguish between patterns.

AI Habitat: ultra-precise navigation AI in the space of

Facebook has developed a large-scale DD-PPO reinforcement learning algorithm that effectively copes with navigation in the space of residential and office premises using data from the RGB-D camera, GPS and compass. Robots trained with DD-PPO achieve their goal with 99.9% accuracy. The article contains impressive videos.

Can you trust the uncertainties of your model?

Published code and research results, which focuses on the problem of covariant shift, when the data for training the model do not coincide with the test sample. The authors conducted an extensive comparative analysis of the most advanced models, providing them with datasets from text, pictures, online advertising and genomics. The goal was to determine the accuracy of popular methods. In benchmarks, the ensemble method proved to be the most reliable.

Thank you for the attention!

Before you was a trial release, any recommendations are welcome.

Source: https://habr.com/ru/post/undefined/