AI predictions for 2019

Author: Dr. Zunaid Kazi
Chief Technology Officer

In November of 2018, I had the opportunity of attending the AI Frontiers conference in San Jose. This conference brings some of the leading minds of AI to showcase frontier AI technologies. My one key takeaway from the conference is that this may be the last time I am able to make any reasonable predictions of where AI will go next year. And I did so well last year.

AI progress is accelerating at such a rate that any predictions we might make as well be as off the mark as those 19th century predictions about the 20th century.

Before putting my prediction stakes into the ground for 2019, I can share with you one prediction from my AI in the 2020s article from next year: AGI is coming.

Coming back to firmer footing, here are my top five AI predictions for 2019.

It’s all talk

Image recognition and generation have been “solved”. Natural Language Understanding and Natural Language Generation too will be solved in 2019.  A conversational agent on your phone will pass the Turing Test.  AI assistants will be ubiquitous and embedded everywhere: your phone, your car, your house. From the simple tasks you do with Alexa or Siri today, you will be engaging in a normal conversational mode with smart agents to solve more complex tasks. We got an early preview of this with Google Duplex.

It’s about the carrot and not the stick

Even though the potential of RL is immense, RL research had mostly been confined to academia partly due to complex algorithms and lack of mature tools. 2018 saw the re-emergence of RL in the form of deep learning particularly assisted by the amount of compute available today (there has been a 300,000 increase in neural network compute in the past six years!). DeepMind’s AlphaZero and OpenAI’s Dota bot have shown the immense potential of deep reinforcement learning agents. In 2019, attention will start switching from supervised learning to reinforcement (and semi-supervised) learning. 2019 will be the year of Reinforcement Learning (RL).

Seeing is not believing

Generative AI has gotten so good that very soon people will be able to create fake videos of people that will cross the “uncanny valley” and be indistinguishable from the real thing. We will see a national or political crisis precipitated by a deep fake video.  Someone will create a video of a public figure saying something incendiary to manipulate a political process that will go viral and have an even more onerous impact on the fabric of democracy than the fake news of the recent past.

Time to update that resume

Can machine learning be used to automate machine learning? The answer to that question is, yes. AutoML in 2018 showed that you can use machine learning techniques to select algorithms and tune model hyperparameters. Incremental advances to AutoML will allow for solving standard supervised learning out of the box, You won’t need to be a Machine Learning Engineer to do machine learning in 2019.

Living on the edge

There are billions of microcontrollers (MCUs) being shipped every year. These are low cost and low power devices that are mostly embedded in IoT devices and usually underutilized. More and more researchers are now adding AI to these edge devices. By moving the compute and storage capacities nearer to where the data is will result in smarter and faster decisions. Machine Learning will increasingly move from the cloud to the edge. This development will be further accelerated as more vendors develop specialized machine learning processros designed for the edge.