Bill Vorhies, editor-in-chief of Data Science Central and who has years of experience in data science and business analysis models, has written that the most important developments in artificial intelligence and deep learning over the past year are not technology but the shift in business models – all giants will Its deep learning IP open source. Needless to say, the "open source wave" is a trend that can't be ignored in the field of artificial intelligence in 2016, and the most popular project is Google's deep learning platform TensorFlow. The following section starts with TensorFlow and takes a look at the 2016 AI open source project, and finally statistics on Github's Top 50, the most commonly used deep learning open source project.
1. Google's second-generation deep learning engine TensorFlow open source
In November 2015, Google open source deep learning platform TensorFlow. In April 2016, Google introduced distributed TensorFlow. Now, TensorFlow has become one of the most popular deep learning platforms in the industry.
2. Google open source world's most accurate language parser SnytaxNet
On May 13, 2016, Google Research announced that the world’s most accurate natural language parser, SyntaxNet, is open source. Google open source further. According to reports, the accuracy of the language understanding of Google's model trained on this platform exceeds 90%. SyntaxNet is an open source neural network framework running in TensoFlow that provides a natural language understanding system. Google has disclosed all the code needed to train new SyntaxNet models with user's own data, and Google has trained to analyze English texts. The model Paesey McParseface.
Paesey McParseface is based on powerful machine learning algorithms that learn to analyze the linguistic structure of sentences and explain the function of each word in a particular sentence. In this model, Paesey McParseface is the most accurate in the world, and Google hopes it will help researchers and developers interested in automatically extracting information, translation, and other applications in natural language understanding (NLU).
3.Google launches Deep&Wide Learning, an open source deep learning API
On June 29th, 2016, Google launched Wide & Deep Learning and opened the TensorFlow API. Developers are welcome to use this latest tool. At the same time, open source also implements Wide & Deep Learning as part of the TF.Learn application programming interface, allowing developers to train their own models.
4. Google open source TensorFlow automatic text summary generation model
On August 25th, 2016, Google opened up a model for extracting text information and automatically generating abstracts in TensorFlow. It is especially good at long text processing, which is very useful for automatically processing massive information. The most typical example of automatic text abstraction is the automatic generation of headlines for news reports. In order to do a summary, the machine learning model needs to be able to understand the documents and extract important information. These tasks are very challenging for the computer, especially in the document. Increase in length.
5. Google open source image classification tool TF-Slim, defines the TensorFlow complex model
On August 31, 2016, Google announced the open source TensorFlow advanced software package TF-Slim that enables users to quickly and accurately define complex models, especially image classification tasks. Since the release of TF-Slim, TF-Slim has achieved significant growth. Many types have been added to the network layer, cost function, and evaluation criteria. Training and evaluation models have also had many convenient routines. These means that you don't have to worry about the details when you run large scale operations such as reading data in parallel or deploying models on multiple machines. In addition, Google researchers also produced the TF-Slim image model library, which provides definitions and training scripts for many widely used image classification models, all written using standard databases. TF-Slim and its components have been widely used within Google, and many upgrades have also been integrated into tf.contrib.slim.
6. Google open source large-scale database, 1 billion + data, explore the RNN limit
On September 13, 2016, Google announced the open-source large-scale language modeling library, which was called "Exploring the RNN Limit" when it was published in February of this year. Now, it is even more attractive to see the open source. . The research test has achieved excellent results. In addition, the open source database contains about 1 billion English words, 800,000 words, and most of it is news data. This is a typical industry study. It was only done by big companies like Google. This open source should also play a role in machine translation, speech recognition, and other areas as the author hopes.
7. Google open source TensorFlow diagram says to generate models that can really understand images
On September 23, 2016, Google announced the latest version of the open source graphics system Show and Tell on TensorFlow. The system uses an encoder-decoder neural network architecture, with a classification accuracy of 93.9%, and can generate accurate new graphs when encountering new scenes. Google said that this shows that the system can truly understand the image.
8. Google open source large database, including 8 million + video
On September 28, 2016, Google announced on its official blog that it has open-sourced a video database containing 8 million Youtube video URLs, and the total video duration has reached 500,000 hours. Also released is a video-level tag extracted from a collection of 4800 knowledge maps. This database has a significant increase in size and coverage compared to existing video databases. For example, the more famous Sports-1M database has only 1 million Youtube videos and 500 sports categories. According to Google’s official blog, Youtube-8M represents almost exponential growth in the number and variety of videos.
9.Google releases Open Images image dataset with 9 million annotation images
On October 1, 2016, following the release of 8 million video data sets the day before yesterday, Google also released the image database Open Images, which contains 9 million annotation data and more than 6,000 kinds of label types. Google wrote in the official blog that this is closer to real life than ImageNet, which has only 1,000 categories. For those who want to train computer vision models from scratch, this data is far enough. In December, Google also open sourced a script for the Open Images parallel download tool, which was faster than 200 M in 5 days.
10.DeepMind open source AI core platform DeepMind Lab (with papers)
On December 5, 2016, DeepMind announced the open source of its AI core platform, DeepMind Lab. The DeepMind Lab uploads all the code to Github for researchers and developers to experiment and research. DeepMind Lab's platform integrates several different AI research areas into one environment to allow researchers to test AI agent navigation, memory, and 3D imaging capabilities. It is worth mentioning that these codes also include AlphaGO's code, Google hopes to increase the openness of AI capabilities, allow more developers to participate in AI research, observe whether other developers can challenge and break DeepMind's current record.
Facebook Open Source: Implementing Ideas1.Facebook Open Source Go Engine DarkForest
6 months ago, Facebook opened its Go engine DarkForest. The training code is now fully released. Github link: https://github.com/facebookresearch/darkforestGo.
2. Facebook's opentext text classification tool fastText, without deep learning can also be fast and accurate
On August 19, 2016, the Facebook AI Lab (FAIR) announced the openText text analysis tool fastText. fastText can be used both for text categorization and vocabulary vector characterization. The accuracy of text classification is comparable to some commonly used deep learning tools, but it is much faster in time - model training time is reduced from days to a few seconds. In addition to text categorization, fastText can also be used to learn vector representations of words. Facebook claims that fastText is much better than the most advanced morphological characterization tools like Word2vec.
3. Facebook open-source computer vision system deepmask, understand the image from the pixel level (with papers and code)
On August 26, 2016, Facebook announced deepmask, an open-source computer vision system, that the system can "understand objects from the pixel level." Facebook hopes that open source will accelerate the development of computer vision. However, Facebook does not use these tools in its own products, and it is open source before it is implemented in specific applications. It is somewhat different from what is commonly referred to as “open sourceâ€. In this regard, Yann LeCun, head of FAIR, Facebook's artificial intelligence team, once said that it is precisely because of FAIR's foundation that it is not subject to the company's short-term benefit research that can actually advance the development of artificial intelligence technology.
4.Facebook open source AI training and testing environment CommAI-env
On September 27, 2016, Facebook announced the opening of an AI training and testing environment, CommAI-env, to set up agents in any programming language. According to reports, the CommAI-env platform is used to train and evaluate AI systems, especially AI systems that focus on communication and learning. Unlike OpenAI Gym, which uses reinforcement learning from playing games to playing games, Facebook's CommAI-env focuses on communication-based training and testing. This is to encourage developers to better create artificial intelligence that can communicate and learn. It should be the company’s 10-year plan. Facebook also stated that CommAI-env will continue to be updated and will organize competitions to promote the development of AI.
In terms of the AI ​​test environment, Facebook also has open sourced CommNet, a model for neural network-based agents to better interact with each other and achieve cooperation. This model is supported by CommAI-env. In December, Facebook also open-sourced TorchCraft to bridge the gap between Torch and StarCraft in the deep learning environment, allowing researchers to use controllers and write smart agents that can play Starcraft games.
5.Facebook Jia Yangqing introduced Caffe2go, mobile phone can run neural network
On November 8, 2016, Caffe author and Facebook researcher Jia Yangqing published an article on the official website describing the new machine learning framework Caffe2go and stated that it will be partially open source in the coming months. Caffe2go's smaller scale, faster training, lower computing performance, and ability to run on mobile phones have become core technologies for Facebook's machine learning.
OpenAI
1.OpenAI Launches Agent Training Environment OpenAI Gym
The establishment of the nonprofit OpenAI, which was founded at the end of 2015, broke the pattern of giants such as Google and Facebook occupying the AI ​​field, but its founder and Tesla CEO Musk has repeatedly published artificial intelligence threat theory. What is the purpose of Muske’s founding OpenAI? On May 4, 2016, OpenAI released the artificial intelligence research tool set OpenAI Gym, which is used to develop and compare reinforcement learning algorithms, analyze OpenAI Gym or find out the real motivation of Musk.
2. Another open source: OpenAI introduces the deep learning basic framework
General Purpose Servo Motors,Cnc Lathe Electric Motors,Cnc Lathe Servo Motor,Servo Motor For Sewing Machine
Zhejiang Synmot Electrical Technology Co., Ltd , https://www.synmot-electrical.com