Five key trends for AI in health in 2021 and beyondFive key trends for AI in health in 2021 and beyond
Having a background in computer vision, and having recently attended the “Future of Health-Inspired AI” conference, I’d like to break with tradition and make my predictions a little more specific
November 30, 2020
Predicting key trends is a lot like taking on the role of a traditional fortune teller – the predictions are at their most believable when they are kept frustratingly vague. I can easily predict that “The Internet of Medical Things will revolutionize health care” but as that forecast does not give a date for this, or specify what “revolutionize” actually means, the prediction can never be evaluated, although spotting trends sounds great and makes the author sound very smart.
AI learns by feedback loops, where an error in a prediction is taken into account to improve the next prediction. Having a background in computer vision, and having recently attended the “Future of Health -Inspired AI” conference, I’d like to break with tradition and make my predictions a little more specific. Just as with AI, this is a chance to improve my own forecasting skills, and to let you as a reader accurately assess my qualities as a forecaster. To that end, I’ll even give you a date and a way to evaluate if I was right.
With that in mind, here are my five key trends:
1. Public anxiety around Artificial Intelligence continues to grow.
In the nineties there was a great deal of talk about making machines intelligent by developing “rule based engines”. An example would be “If a patient recently traveled to the tropics, and today the patient has a high fever, it is possible that the patient suffers from Ebola.” These ideas are well understood by the public – however, these rule engines never worked very well, and became unmanageable with increased size or complexity.
In contrast, today’s artificial intelligence works as a “black box.” It is trained on data, and from that data it derives its own set of rules, often incomprehensible to us humans - engineers and the general public alike. The AI black boxes outperform any other technology known today and are becoming more and more widely used, but because their output cannot be explained and can often end up amplifying our own human biases, “algorithm” is increasingly becoming a dirty work with wholly negative associations. The public is more cautious about AI than ever even as applications and adoption skyrocket. There is hope though: technologies are being developed for making AI explainable which will hopefully combat this.
How to validate this forecast: In 2021, at least ten Artificial Intelligence startups focusing on determining bias and explaining artificial intelligence will receive over $250M funding.
2. Computer vision applications will improve recognition accuracy thanks to new “Transformers” neural network architecture.
The recent star of the show is the artificial intelligence model GPT-3, which stands for Generative Pre-trained Transformer 3. GPT-3 can create anything that has a language structure – which means it can answer questions, summarize long texts, or translate languages. Very recently, it transpired that these Transformer networks not only produce excellent results on texts, but that they can also be employed to recognize what is going on in images and video.
How to validate this forecast: In the year 2022, Transformer networks will outperform the better known deep convolutional networks on all academic image and video datasets.
3. Increased proliferation of computer vision-based object detection which incorporate reasoning over time.
A good example of this is AI applications in the medical field that look after the well-being of patients and improve over time. Here, the AI software needs to recognize if the patient is in bed, getting out of bed, moving into the bathroom, and making sure that the patient gets back into bed. Today, this is one of the most complex applications of computer vision technology, as a single recognition error early on ruins all subsequent reasoning over time. Further down the line we can expect to see this extrapolated to a huge range of body language recognition applications beyond the medical field.
How to validate this forecast: In the year 2023, Amazon and Microsoft offer this functionality in their computer vision cloud.
4. Where traditional businesses are moving en-masse to the cloud for their computing needs, the AI community does the opposite by moving out of the cloud towards a variety of edge devices.
The big variety of edge devices with specialized AI hardware start with your mobile phone. It extends to personal assistants such as the Amazon Echo, and further still to devices like your doorbell which can contain network hardware to recognize who is outside your home. Smart security cameras are managed through mini personal computers with GPU hardware (conveniently called IoT Edge Devices by their marketing departments). Telecoms operators will enable AI processing close to your home in their 5G broadcast towers. And so on. At its current rate of proliferation the AI computation hardware in your personal home stands to become a real tangled mess in the years to come.
How to validate this forecast: If something becomes a mess for you, it can become a financial opportunity for someone else. Watch out for the stock price of specialized AI hardware vendors. These will have surged by 100% by 2024.
5. Europe will not win the AI race against China and the United States.
Despite Europe’s ambition to be competitive in the AI space, this ambition will remain out of reach. Aside from broader issues of funding, this is because of Europe’s clear definition of GDPR limiting the time that data can be stored. The GDPR regulation is absolutely necessary to govern what happens to data, but the Chinese and American equivalents are vaguer and more permissive, respectively. Our European regulation is often misinterpreted by data processing officers in the strictest way possible, just so that they are safe. While this benefits individuals, in practice, this means European companies can store their training data for just a month. Meaning that European startups’ AI software can become “forgetful” as the data on which they train cannot be reused a year or more later. In contrast, Chinese and American AI systems will become more sophisticated over time as they can memorize much more data.
How to validate this forecast: In the year 2026, take an AI ETF index tracker. The number of Chinese and American companies in it will be larger than the number of European companies.
Dr Harro Stokman is CEO of the computer vision startup Kepler Vision Technologies