What Were the Major Advances in Artificial Intelligence or Machine Learning in 2017?

Suffice it to say, 2017 was an incredible year for artificial intelligence. We also saw the technology begin to make some baby steps from the “hype” and Hollywood sci-fi realms into more mindboggling and even creepy applications. Artificial intelligence is yet to permeate our lives but we are already seeing some really interesting developments in this field. In 2017, a lot of AI or machine learning applications came into fruition. Breakthroughs continue to roll in and even though we are still far from seeing anything close to artificial general intelligence or utility in our everyday lives for some of these groundbreaking innovations, they are quite notable developments and they reveal the potential of artificial intelligence to radically shape our world in the future.

Below is an overview of some of the main advances in artificial intelligence in 2017:

AlphaGo Zero Beat Top Champs in Go, Chess, and Shogi Without Human Knowledge

This was the biggest artificial intelligence news in 2017. Developed by DeepMind, a subsidiary of Google, AlphaGo Zero learned to play the Go game in four hours and the most remarkable thing is that it didn’t do so by watching humans play or analyzing data but by mastering the rules/mechanics of the game and then proceeded through 5 million games of self-play to fully master the game.

Even more remarkable is the fact the AlphaGo Zero learning process was similar to what humans typically go through when learning and mastering the game. The artificial intelligence system even developed new strategies: it rejected short term goals and developed the level of strategic thinking typically seen in top-level human players. It even went further and created new moves that have never been seen before in any human gameplay. This was an artificial intelligence system displaying a high level of genius-learning a game from the ground up without human help and building its own sets of strategies to reach the highest level of mastery.

It was an important breakthrough because it demonstrated the power of deep learning and reinforcement learning. Reinforcement learning is a type of machine learning where the algorithm is only in possession of the rules and the mechanics of its environment but not the data. The original AlphaGo was the classic experiment in Machine Learning; it learned its moves through classic machine learning by observation and data analysis- from more than 30 million moves of expert human gameplay. In contrast, AlphaGo Zero learned all moves from scratch without any observation and refined itself beyond any human expertise by playing against itself.

Advancements in Generative Adversarial Networks (GANs)

The generative adversarial networks or GANs have continued to post impressive results, especially when it comes to their applications in the image space. This has been on the heels of the Wasserstein GAN that radically enhanced the traditional GAN and enabled the flourishing of various types of GANs such as progressive GANs, Conditional GANs, BEGAN and CycleGAN.

The most spectacular production out of the progressive GAN over the past year was Nvidia’s production of fake but photorealistic celebrity images using artificial intelligence. The GANs used to consist of two neural networks. The first network is called the generator and it generates celebrity images based on analyzed data. The second part of the GAN is the discriminator that assesses the output from the generator and identifies the anomalies. These are then corrected by the generator and the result was a high-resolution generation of fake celebrity images as seen below.

Facebook and Microsoft Launched ONNX AI Model/Platform Interoperability

Two of the world’s biggest tech giants-Microsoft and Facebook-announced a joint open standard project that will facilitate the sharing of deep learning models between various artificial intelligence frameworks. This will be the Open Neural Network Exchange (ONNX). The ONNX open source platform will also make it possible to train models in a single framework and seamlessly transfer or deploy them to other machine learning frameworks for ease of applicability. The new project was also backed by other partners such as Huawei, Nvidia, Intel, AWS and Qualcomm.

DeepMind Launched the Sonnet Open Source Framework

2017 saw the release of the Sonnet open-source framework by Google’s AI subsidiary DeepMind. Sonnet will provide developers with a platform where they can create neural network components. It joins other AI platforms that have been released in the recent past such as Google’s TensorFlow Eager and Facebook’s deep learning platform PyTorch, a python-based DL machine learning library that provides support for dynamic computational graphs.

The Rise of Machine Learning as a Service (MLaaS)

Machine learning has been touted as the most successful and most practical implementation of artificial intelligence. As companies move to embrace artificial intelligence or machine learning, the growth of companies that offer machine learning platforms as a service to the less technically endowed companies has been rising expeditiously. In turn, many tech companies have emerged that offer users machine learning platforms or APIs as a service, thus helping bridge the gap between the aspiration and implementation of AI.

Some of the major tech companies in the AI realms have been making a big splash for their MLaaS platforms aimed at democratizing the artificial intelligence space and taking it more mainstream. The aim of MLaaS is to help companies that lack AI or data science talent to still leverage the power of the latest artificial intelligence breakthroughs in their enterprises through an API.

Some of the biggest MLaaS companies include the following:-

Amazon Machine Learning
Azure Machine Learning
Google Prediction API
BigML
ForecastThis
Deep Cognition
DeepAI
Ersatz Labs Inc

AutoML Simplified Machine Learning

Training high-quality performant machine learning models for your unique use cases is generally a difficult process. It involves a lot of data munging and other iterative processes.

With AutoML, developers with limited machine learning expertise are able to leverage certain tools to train top quality custom machine learning models. AutoML increases the accessibility of machine learning models to non-machine learning experts.

Universe Gained Traction in 2017

Universe is a free software platform that allows for the training of “friendlier” general artificial intelligence through a process of reinforcement learning in disparate environments like websites, games along with various other applications. While the platform was launched in 2016, it began to see solid traction in 2017.

Geoffrey Hinton Unveiled the Capsule Networks

One of the leading global deep learning experts Geoffrey Hinton unveiled a new neural network architecture called the Capsule Networks in 2017 as an alternative to Convolutional Networks. The Capsule Networks are a vast improvement over the convolutional neural networks which tend to have some weaknesses such as the ease with which they can be tricked by intentional adversarial attacks and images that are conceptually similar. The Capsule Networks are expected to have a massive implication on deep learning, particularly with regards to computer vision.

The First AI Citizen

This was not necessarily a major artificial intelligence breakthrough but more of a quirk. Saudi Arabia granted citizenship to an AI robot created in Hong Kong called Sophia. The robot is not a general artificial intelligence product; it is more within the scope of narrow AI applications with functionalities such as facial recognition, voice recognition and natural language processing.

Serious Ethical Debate on Machine Learning Systems

For the first time, a serious debate on the ethical questions surrounding the development of artificial intelligence took centre-stage. Many industry professionals and leaders convened forums around the world to discuss the place of AI in our society and the ethical issues around its evolution. One of the most consequential technology leaders of our time Elon Musk has sounded off a nightmarish warning that advances in general artificial intelligence could lead to the creation of an immortal dictator.