October 28, 2020
Artificial intelligence (AI) is at the forefront in changing the world and the way we live our lives.
Self learning machines help us execute everyday, mundane tasks.
Companies across industries are all jumping on the AI-bandwagon, but there’s a lot of hype around AI.
So what is fact, and what is fiction? In these series, our founder Svetlana Kordumova will be talking with industry and academic experts who have deep knowledge of AI to understand better what AI is and how it will continue to change the world.
This article is an excerpt of her interview with Cees Snoek, Professor in AI, Computer Vision, Deep Learning at the University of Amsterdam.
I am a full professor in computer science at the University of Amsterdam where I lead the Video and Image Sense lab.
When I entered university in 1996, the Internet was becoming mainstream.
It was an exciting time, a lot of technological breakthroughs happened during that period. I found that very interesting.
During my studies I came in contact with automated video recognition and I just got hooked.
I wanted to understand how it is feasible that a machine can interpret and translate a stream of pixels to human-understandable information.
So I made it my master thesis topic, and just kept going.
Went through the academic job hierarchy, visited labs at Berkeley and Carnegie Mellon, was head of R&D at a University startup, gained industry experience at Qualcomm and eventually decided that I wanted to be a full-time university professor.
For me that is clearly the rise of deep learning with convolutional neural networks.
A technology that was around for years, but could only reveal its true power when large-scale labeled datasets became available together with highly parallel GPU computation devices.
The victory by Alex Krizhevsky in the ImageNet 2012 competition marked the end of an old-era and the start of a new one.
Due to the success of deep learning, the divide between academia and industry is blurring.
The best minds often have industry affiliations at the research labs of big tech companies.
At the same time these companies are supporting many university researchers with funding.
Naturally the research agenda of the two gets mixed. In a way academia has to reinvent itself and accept that deep learning from many examples just works and is no longer interesting for university research.
Academia should focus on the really hard fundamental problems and explore other unpopular but innovative AI methodologies.
Among PhD graduates I see only a small percentage with the ambition and drive to found a startup.
It is more common that fresh PhDs join a startup to spearhead the R&D efforts. For a startup you need a blend of skills, many of them are not learned during PhD.
Technology skills are certainly not the most important.
Interestingly, I see a lot more startup initiatives among our AI MSc student population.
Today’s AI is working very well in some specialized or highly-constrained domains.
It is tempting to extrapolate this success to exciting challenges like self-driving cars and curing cancer.
AI will certainly contribute to these challenges, but I don’t expect we will see a complete solution in the next decade.
Self-driving cars are likely to happen on highways, but I consider it unlikely to be able to drive in the city centers of Mumbai or Jakarta.
Throughout its history AI has suffered from over-estimating the time it would take for breakthroughs to happen.
Herbert Simon, one of the founders of the field, predicted in 1957 that AI would need a decade to beat a human chess player, while in reality it took us forty years.
Al has no compassion.
A critical human skill in many jobs, whether it is teaching children, caring for elderly, or judging a criminal offence.
AI has the potential to transform any industry, COVID-19 is forcing many industries to be creative and innovative earlier than they might have planned.
Today’s e-commerce is still somewhat simplistic, essentially a digital catalogue in which it is easy to browse and find recommendations.
Basically emphasizing information retrieval and recommendation techniques.
More advanced AI-powered computer graphics and computer vision can enrich the customers’ product experience, for example by getting an impression of the fabric of a dress or shawl, or to have a virtual clothing try-on using a 3D representation of your own body.
There are many, the ones that I find most interesting are learning from limited examples and long-range reasoning, especially when the two come together in making sense from video streams.