12 AMAZING DEEP LEARNING BREAKTHROUGHS

Last year many AI engineers aimed to create a real artificial intelligence system. A real machine that thinks is what every AI geek wants. Machine learning is the most successful application of artificial intelligence. For example, machine learning is able to win a human Go player.

Last year was the year of artificial intelligence. The most amazing AI applications are:

1. DEEPMIND’S ALPHAZERO

AlphaGo is a deep learning system, and it has a new version called AlphaZero. AlphaZero was trained to play chess in just four hours. AlphaZero was able to trounce Stockfish (which is the top AI chess player) in a 100-game match.  

2. OPENAI’S UNIVERSE GAINED TRACTION

OpenAI aims to build a “friendly” AI system. In fact, Universe is a free platform for developers that want to train an AI agent using reinforcement learning across different environments like websites, apps, and games. Universe was launched in 2016.  

3. SONNET & TENSORFLOW

TensorFlow (which was launched by Google) is an open-source library for machine learning. Then, it was launched Magenta, which is an AI platform for creating art and music. In 2016, PyTorch is another python deep learning platform that was launched by Facebook. PyTorch supports dynamic computation graphs. Then, it was launched another deep learning platform that is called Tensorflow Eager by Google. Then, Sonnet was launched by Google. Sonnet is an open-source framework for deep learning. Developers can use Sonnet with ease to build their neural network.

4. FACEBOOK AND MICROSOFT TEAMED UP TO ENABLE AI FRAMEWORK INTEROPERABILITY

Facebook and Microsoft developed the Open Neural Network Exchange (ONNX), which is a deep learning open format. ONNX can be used to build deep-learning models, and transfer the deep learning models to another inference.  

5. UNITY FOR BUILDING INTELLIGENT AGENTS IN GAMES

Unity allows you to build ML-Agents, which is a platform to leverage Unity simulations and games as customizable environments. Unity can be used to train intelligent agents using algorithms such as evolutionary strategies and deep reinforcement learning.

6. MACHINE LEARNING AS A SERVICE (MLAAS) PLATFORMS

In 2017, it was launched many Machine Learning as a Service platform (MLaaS) in order to make it easy to developers without machine learning skills. So, it was launched Google Cloud AI. Amazon launched some NLP platforms such as Lex, Polly, and Alexa Skills Kit. Microsoft and IBM also launched their own machine learning services. 

There are many startups that launched their own machine learning platforms and deep learning platforms. In fact, Uber launched its own platform called Michelangelo. Furthermore, Facebook launched its platform called FBLearner Flow. Additionally, Twitter launched its own platform called Cortex.  

Furthermore, Capital One and other companies have also set up their own Center of Machine Learning Excellence.

7. THE GAN ZOO CONTINUED TO GROW

In 2017, it was developed an improved GAN (generative adversarial networks) training, which is called WGAN. WGAN is a new GAN algorithm that improved learning stability, addressed mode collapse, and improved debugging. Then, it was developed new GANs such as BEGAN, CycleGan, and Progressive GAN flourished. In fact, NVIDIA and the progressively training GANs were used to generate high-resolution facial photos of fake celebrities.      

8. NLP    

Natural language processing (NLP) tasks such as speech recognition and machine translation use deep learning techniques such as long-short-term memory (LSTMs). It was proposed a new neural network architecture on the paper “Attention Is All You Need”. The new neural network architecture proposed is called Transformer, which dispenses with computationally expensive aspects such as recurrence and convolution to better performance on machine translation tasks than other state-of-the-art ANN architectures.

9. AUTOML

AutoML is a platform to automate machine learning. Indeed, AutoML is used for data cleaning and preparation, to model parameter search and optimization, and to deployment and scaling. Google launched AutoML (in alpha) to automate machine learning. Additionally, Amazon launched a similar platform, which is called SageMaker. There are other platforms called DataRobot, H2O.ai’s Driverless, and open source Python solutions like TPOT. 

10. HINTON’SCAPSULE NETWORKS

Geoffrey Hinton (deep learning pioneer) developed a new network architecture, which is called Capsule Networks. Capsule networks are better than the traditional convolutional neural networks. 

 11. QUANTUM & OPTICAL COMPUTING ENTERED THE AI HARDWARE WARS

Google launched a new version of Tensor Processing Units (TPU), which is designed specifically for deep learning. IBM and Google will launch solutions for quantum computing. In fact, it was found that parallel computing for deep learning can be done by switching from an electrical computing paradigm to a phototonic one.

 12. ML SYSTEMS TOOK CENTER STAGE 

Cathy O’Neil (author of Weapons of Math Destruction) said that we need to drop our blind faith in big data. Fei-Fei Li (Stanford Professor and Chief Scientist of Google Cloud AI/ML) expanded AI4ALL, which is an educational non-profit AI training. Kate Crawford and Meredith Whitaker starts AI Now, an interdisciplinary research organization dedicated to studying the social implications of AI. Actually, Crawford said that many challenges facing machine-learning systems today. Crawford wants to prioritize ethics, fairness, and safety.

Bigdataguys consults fortune 500 companies on using blockchain technologies, artificial intelligence, & data science to enhance their operations. We’re a full-service machine learning, artificial intelligence & deep learning firm.

Contact us Bigdataguys offers deep learning consulting and bootcamps.