Making AI live up to its hype

For many people, artificial intelligence (AI) is a magical black box. Data goes in, algorithms do their thing and out comes a result.

But is that result reliable? Can it be trusted to support a critical business or public policy decision that could make or break a company, or impact millions of lives?

These are the kinds of questions being tackled at the University of Ottawa Faculty of Engineering’s School of Electrical Engineering and Computer Science.

It begins and ends with data

Prof. Herna L. Viktor examines the issue where databases and AI intersect.

“A machine-learning algorithm is only as good as your data,” she said. “We cannot just blindly run algorithms and take the output at face value. It is very important to look at the characteristics of the data and to consider issues such as noise, underrepresentation and bias. All of these will influence the final results.”

For example, a recent tech talent recruitment engine used by Amazon had to be abandoned, since the underrepresentation of women in the field led to the machine learning algorithm only targeting male candidates. As another example, the use of machine learning techniques to detect rare or orphan diseases are hindered by the small pool of patients, which may lead to their symptoms being overlooked.

Data is constantly evolving and growing. It’s been estimated that, by 2020, 1.7MB of data will be created every second for every person on earth. The challenge is to create algorithms with the sophistication to not only ingest such staggering amounts of data, but to recognize and sort good data from bad, and be able to explain how a conclusion or output was reached. Providing that kind of transparency and accountability has always been a key challenge in machine learning.

“With AI beginning to play such a large role in decision-making, there is a need to make the results of learning interpretable and to explain to domain experts exactly how we arrived at a decision,” Prof. Viktor said.

Those domain experts could be in healthcare, in emergency preparedness, or in an IT vertical such as cloud services.

Prof. Viktor and her team have already been working with public and private sector partners on projects such as trending shifts in patterns of illness and disease for more proactive healthcare planning, predicting extreme weather events and detecting cybersecurity threats.

5G and beyond

Prof. Melike Erol-Kantarci, Director and Founder of uOttawa’s Networked Systems and Communications Research (NETCORE) Lab, is researching deep learning and machine learning as it applies to the efficient operation of 5G wireless networks.

5G is heady stuff. Data rates on a 5G device could be up to 1,000 times faster than current 4G (LTE) devices. Ultra-HD movies that may be too frustrating for most people to even attempt to download could be ready to watch in mere seconds.

With that kind of connectivity and speed comes a dramatic increase in network complexity.

“Where there are hundreds of cellular base stations today, there will be thousands,” Prof. Erol-Kantarci said. “These will have to compensate for interference from other stations and users. Current AI models are not capable of ensuring efficient resource allocation and managing initial and fair access for wireless devices.”

The first 5G products will hit the market in 2019 with full network deployments to follow. AI enablement is crucial. Prof. Erol-Kantarci points out that this isn’t just about that smartphone in your pocket. 5G is a crucial standard to support the increasing connectivity of society as a whole – the Internet of Things.

“Canada is investing a lot in autonomous driving and that needs the connectivity that will come with 5G,” she said. “This also applies to any ‘Smart’ initiative – SmartHealth, SmartGrid or SmartCity. Wherever there is a demand for higher speed wireless connectivity and ultra-low latency, it falls under what we do at uOttawa to enable future networks with AI.”
 
 Learn more about uOttawa’s Faculty of Engineering at engineering.uOttawa.ca.