2017 has been a good year for AI, machine learning and especially the branch of deep learning gave many of the breakthroughs that have impacted the most during 2017, from game winners to art that competes with humans.
- DeepMind's AlphaZero managed to beat the champions of Go, chess and shogy.
After winning the world champion of Go, AlphaGo a year later, it was updated and converted into AlphaZero, which managed without human supervision, only with the basic rules of the game, to learn to play chess like an expert in just 4 hours; and he proved it by beating the AI Stockfish, until that moment the best chess player based on AI. - Universe openAI was endorsed by great partners
Universe is a free platform where developers can train an AI agent using reinforcement learning in different environments like websites, apps, and games. Launched in December 2016, the platform gained traction in 2017, attracting partners like EA, Valve, and Microsoft Studios by taking advantage of the opportunity for Universe's AI agents to browse and learn from their games. - Sonnet and Tensorflow Eager join open source systems
After Google released its open source machine learning library Tensorflow in 2015, it was followed by Magenta (a research project exploring the role of machine learning in the art and music creation process) also from Google. Facebook AI released a year later PyThorch, a python deep learning platform that supports dynamic computing graphics. In 2016 Google also introduced Tensorflow Eager. In 2017 and through its subsidiary DeepMind, Google launched Sonnet, an open source system that makes it easy for developers to build neural network modules. - Facebook and Microsoft joined forces to enable combinable AI systems
Tech giants created ONNX (Open Format for Representing Deep Learning Models) that allows models to be trained on one system and transferred to another by inference. - UNITY allowed developers to easily create intelligent agents in games
One of the world's leading game developers Unity Technologies created Unity Machine Learning Agents. A platform that allows AI developers and researchers to use simulations and games as customizable environments where they can train intelligent agents using evolutionary strategies, deep and reinforcement learning, among others. - Machine learning platforms as a service are everywhere
More and more companies have joined the race to build in-house machine learning platforms and deep learning centers of excellence. Uber has Michelangelo, Facebook has FBLearner Flow, Twitter has Cortex. Capital One and other forward-thinking companies outside of the core technology environment have also established their own centers of excellence in machine learning.
Amazon, Microsoft, IBM and Google decided to open their APIs to try to democratize the use of AI and share it with companies that do not have the talent or the ability to create their own. - The GAN variety continued
In January 2017, a team of AI researchers published a paper on Wasserstein GANs (WGANs), a substantial improvement on the traditional GAN (generative adversarial network). Which spawned a host of new GANs, from BEGAN up to CycleGan and progressive GAN which was also joined by progressive GAN training, It was this latter approach that allowed Nvidia to generate high-resolution fake photos of celebrities. - Recurrence and convolution are not necessary if you have attention.
Natural language processing tasks such as speech recognition and machine translation have historically been addressed with neural network architectures with memory components, such as LSTM. A groundbreaking article, "Attention Is All You Need," proposed a new model, the transformer, which dispenses with the expensive stuff like recursion and convolution to achieve next-generation performance in machine translation tasks, at least for English and German. Although more research is needed to see if the Transformer architecture holds up across all use cases, the post generated tons of feedback in the community and was ranked as the 4th most popular post of all time on Arxiv (Online Archive owned by Cornell University). - AutoML has made life easier for data scientists and machine learning engineers.
Most developers often find machine learning much more "difficult", as the process of building and training models (and then deploying them to production) is too complicated and time consuming.
Platforms that automate the processing of large amounts of data from sources to train machine learning, ranging from data cleaning and preparation, to model parameter search and optimization, to deployment and scaling, have helped make the “hard” part of machine learning one less hassle for these practitioners. With platforms like Google's AutoML, Amazon's SageMaker or DataRobot, these tasks have been greatly streamlined.
Source: topbots
Keep reading: