January 19, 2018 at 11:16“Knowledge Day” for AI: TOP30 of the most impressive machine learning projects for the past year (v.2018) published
- Transfer
To select the TOP 30 (only 0.3%), over the past year, the Mybridge team compared nearly 8800 open source machine learning projects.
This is an extremely competitive list, and it contains the best open source libraries for machine learning, datasets, and applications published between January and December 2017. To give you an idea of the quality of the projects, we note that the average number of Github stars is 3558.
Open source projects can be useful not only to scientists. You can add something amazing on top of your existing projects. Check out the projects you might have missed last year.
Caution, under the cut a lot of pictures and gifs.
1. FastText
fastText is a library for teaching word representations and classifying sentences, which allows you to organize the automatic assignment of categories to arbitrary text using machine learning methods. [11786 stars on Github]. Courtesy of Facebook Research .
[ Muse : Multilingual Unsupervised or Supervised word Embeddings, based on Fast Text. 695 stars on Github]
2. Deep Photo Style Transfer
Code and data for scientific work Deep Photo Style Transfer [9747 stars on Github] . The approach to the transfer of the photographic style from one image to another with the successful suppression of distortion and preservation of photorealism in a variety of scenarios is described, including the transfer of the features of the time of day, weather, season and artistic changes. Merit Fujun Luan, Ph.D. at Cornell University.
3. Face Recognition
The “World's Easiest” Face Recognition API for Python. The model has an accuracy of 99.38% in the benchmark Labeled Faces in the Wild . It also offers a simple tool that allows you to recognize faces from images in a folder using the command line. Developer - Adam Geitgey [8672 stars on Github] .
4. Magenta
Generate art and music through machine learning [8113 stars on Github] .
5. Sonnet
Sonnet is a TensorFlow-based machine learning library for building complex neural networks. [5731 stars on Github] . Courtesy of Malcolm Reynolds of Deepmind
6. deeplearn.js
deeplearn.js is the WebGL-accelerated JavaScript open source machine learning library from Nikhil Thorat of Google Brain.
7. Fast Style Transfer in TensorFlow
Fast style transfer with TensorFlow [4843 stars on Github] . Logan Engstrom of MIT.
Add the styles of famous artists to any photo in a split second! You can even create videos.
8. Pysc2: StarCraft II Learning Environment [3683 stars on Github] , provided by Timo Ewalds from DeepMind
9. AirSim
AirSim is a simulator for unmanned aerial vehicles, cars and other vehicles created on the Unreal Engine. It is an open source platform for physically and visually realistic simulations. The goal is to develop a platform for AI research and experiments with deep learning, computer vision and stimulated learning systems for autonomous vehicles. [3861 stars on Github] . Developer - Microsoft Shital Shah
10. Facets
The power of machine learning is related to its ability to study patterns in large amounts of data. Understanding your data is critical to creating a powerful machine learning system. The Facets project offers two reliable types of visualization that help you understand and analyze datasets: Facets Overview and Facets Dive.
Visualization is easily integrated into Jupyter notebooks reports or web pages ( Polymer web components, backed by Typescript code).
[3371 stars on Github] . Courtesy of Google Brain
FACETS OVERVIEW Report Example
11. Style2Paints
AI-coloring of images [3310 stars on Github] , can colorize according to a specific color style, create your own style for drawing, or convey the style of an example illustration.
12. Tensor2Tensor
The authors of the scientific work “One Model for Learning Everything” from the Google Brain Team group asked a natural question: “Can we create a unified model of deep learning that will solve problems from different areas?”
Google did this and opened Tensor2Tensor for general use, code published on github . [3087 stars on Github .
In a scientific article, they describe the architecture of MultiModel - a single universal model of deep learning that can simultaneously learn tasks from different domains.
MultiModel Architecture
In particular, researchers tested MultiModel simultaneously on eight data sets for verification:
- WSJ Speech Recognition Enclosure
- ImageNet Image Database
- Base of ordinary objects in the context of COCO
- WSJ Parsing Base
- English to German Translation Corps
- Reverse Previous: German to English translation corpus
- English to French Translation Housing
- Reverse Previous: French to English translation corpus
More details here .
13. Image-to-image translation in PyTorch (for example, horse2zebra, edges2cats, and so on)
[2847 stars on Github] . Courtesy of Jun-Yan Zhu, Ph.D at Berkeley
14. Faiss
Faiss is a library for efficiently looking for similarity and clustering vectors [2629 stars on Github] . Quite often, programmers and specialists in the field of data science are faced with the task of finding similar user profiles or selecting similar music. Solutions can be reduced to converting objects to vector form and finding the closest. Read more on Habré.
Given the first and last image, the algorithm calculates the “smoothest path” between them from the YFCC100M (95 million images). Taken here .
15. Fashion-mnist, Han Xiao, Research Scientist Zalando Tech
Fashion-MNIST [2780 stars on Github] is proposed as a replacement for the MNIST database (short for “Mixed National Institute of Standards and Technology”), since MNIST is too simple. Fashion-MNIST has the same image size and structure for training and testing.
MNIST is a voluminous database of handwritten numeral samples. The database is a standard proposed by the US National Institute of Standards and Technology for the purpose of calibrating and comparing image recognition methods using machine learning primarily based on neural networks. The data consists of pre-prepared examples of images on the basis of which training and testing of systems is carried out. The database was created after processing the original set of black and white samples with a size of 20x20 pixels NIST. The creators of the NIST database, in turn, used a set of samples from the US Census Bureau, to which were added test samples written by students of American universities. Samples from the NIST set were normalized,
The MNIST database contains 60,000 images for training and 10,000 images for testing. Half of the training and testing samples were taken from the NIST training kit and the other half from the NIST training kit.
Numerous attempts were made to achieve a minimum error after training on the MNIST database, which were discussed in the scientific literature. Record results were indicated in publications on the use of convolutional neural networks; the error level was brought to 0.23%. The creators of the database themselves have provided several test methods. In the original work, it is indicated that the use of the support vector method allows to achieve an error level of 0.8%.
Fashion-mnist
16. ParlAI
ParlAI is the foundation for learning and evaluating AI models on a dataset from a variety of dialogs [2578 stars on Github] . Courtesy of Alexander Miller of Facebook Research
17. Fairseq: Facebook AI Research Sequence-to-Sequence Toolkit [2571 stars on Github]
Facebook AI Research (FAIR) team has published impressive results on the implementation of a convolutional neural network for machine translation. She claims that fairseq, a new tool, works 9 times faster than traditional recurrent neural networks, while it is only marginally inferior to them in accuracy.
18. Pyro: Deep universal probabilistic programming with Python and PyTorch [2387 stars on Github] . Courtesy of Uber AI Labs
19. iGAN
Interactive image generation [2369 stars on Github] .
20. Deep-image-prior
Image recovery using neural networks without training [2188 stars on Github ]. Courtesy of Dmitry Ulyanov, Ph.D at Skoltech
21. Face classification and detection from the B-IT-BOTS robotics team
Real-time face detection and emotional + gender classification using fer2013 / IMDB datasets [1967 stars on Github] .
Gender Classification Accuracy (IMDB): 96%.
Accuracy of classification of emotions (fer2013): 66%.
22. Speech-to-Text-WaveNet by Namju Kim from Kakao Brain
End-to-end speech recognition in English using DeepMind's WaveNet and tensorflow [1961 stars on Github] .
23. StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation [1954 stars on Github] . Yunjey Choi at Korea University
24. Ml-agents: Unity Machine Learning Agents [1658 stars on Github] . Courtesy of Arthur Juliani, Deep Learning at Unity3D
Unity Machine Learning Agents allows researchers and developers to create games and simulation environments for machine learning using the Unity Editor using the easy-to-use Python API.
25. DeepVideoAnalytics [1494 stars on Github] . Courtesy of Akshay Bhat, Ph.D at Cornell University
A platform for searching and analyzing visual data.
26. OpenNMT: Open-Source Neural Machine Translation in Torch [1490 stars on Github] .
27. Pix2pixHD: [1283 stars on Github] . Ming-Yu Liu at AI Research Scientist at Nvidia
Pix2pixHD is designed for photorealistic synthesis or conversion of high-resolution images (e.g. 2048x1024). It can be used to turn semantic label cards into photorealistic images or to synthesize portraits using the face tag map.
28. Horovod: Distributed training framework for TensorFlow. [1188 stars on Github] . Courtesy of Uber Engineering
29. AI-Blocks [899 stars on Github]
A powerful and intuitive WYSIWYG interface that allows anyone to create models for machine learning.
30. Deep neural networks for voice conversion (voice style transfer) in Tensorflow [845 stars on Github]. Dabi Ahn, AI Research at Kakao Brain
The goal of the project is to transmit the style of voice or to transform someone's voice into the voice of a particular person. Work on this project was aimed at converting the famous English actress Kate Winslet into the voice.
Disclaimer The
materials provided above are for research purposes only. Using the results to achieve illegal goals may result in criminal, administrative and (or) civil liability. The author is not responsible for such incidents.
