If you want to use Artificial intelligence in your start-up, never base your business on state-of-the-art models. Always try to find good partners for developing your AI. Try not to promote cognitive laziness and only use AI if it actually solves a real world problem that customers will pay to have solved. These were central take home messages when four speakers from academia, institution and start-ups shared their key insights at a 4th of November Brew Your Own webinar for entrepreneurs as well as researchers and students with an interest in entrepreneurship.

State-of-art AI is not always best for you

The first speaker at the event, Christina Lioma, is a professor of Machine Learning at the University of Copenhagen, Department of Computer Science. She warned against assuming that state-of-art AI is always best in start-ups, because these tend to be overly complex for the smaller set up of start-ups. The more complex an AI is, the more data you need in order to train it.

Always remember, that models are cheap. It’s the data that is expensive. The best Machine Learning model is the simplest one that can learn from your data, so start with a reliable model that you know well, and then look into how you can improve the data. After all: A Machine Learning algorithm will also learn from noise and mistakes in the data. If that happens, your output results could cost you both money and credibility”: Christina Lioma, professor, Department of Computer Science, University of Copenhagen.

Partners, regulations and funding

You train AI’s by letting them analyze large amounts of data, so the technology is not the only complex aspect. Things like traceability and quality of data as well as transparency of the analyzing algorithm can determine whether the AI-solution is ethical or even legal. Jens Nedergaard is a Senior Relations manager at DigitalLead, the new national Danish cluster organisation for the digital industry. To tackle these added layers of complexity, he suggests that start-ups find co-creation partners in academia or public authorities.

No one alone has the competencies to handle all these aspects of complexity. We can help you find relevant partners in both public, private and academic institutions. We can help you set up the consortium. We can even help you find national or international funding for your project”: Jens Nedergaard, Senior Relations Manager, DigitalLead.

Overburdened and under-trained

For many applications, AI systems will most likely collaborate with human experts but human experts take a long time to train. Niels Kvorning is CEO of the start-up Melatech. Launching in March 2020 his AI powered DermLoop platform will train doctors to tell the difference between benign moles and the malign melanomas, which signify mole-cancer. In due time the technology will likely be able to diagnose melanomas and the most common differential diagnoses on typical body sites, but Melatechs’ products will only unlock that capability for expert clinicians. They want to avoid cognitive laziness and delayed diagnostics due to inertia caused by a low level of diagnostic confidence among clinicians.

Right now, doctors are overburdened and under-trained for this task. We believe that we can reduce training time from six years to as little as six or eight weeks, because we can display thousands of cases before the doctor sees the first real patient. We can give immediate feedback, unlike real life where you have to wait weeks for an answer from the pathologist, and finally we can tailor the teaching to individual learning patterns. None the less the final call will always have to come from a responsible and competent physician”: Niels Kvorning, CEO and founder, Melatech.

Ignore the rare. Manage the ordinary

You need a lot of good quality data to train an AI. Laying your hands on these is a general challenge for AI start-ups. For Jonas Moll and his VitalBeats start-up, the answer was pacemakers. Doctors monitor these for signals that a heart attack is coming on. However, keeping track of 100,000 heartbeats per patient per day to find the anomaly is… Impractical. In the US alone every 40 seconds, someone suffers a heart attack.

We knew that cardiac doctors and nurses have a problem. They are overwhelmed by pacemaker data. Artificial Intelligence is very good for automating routine tasks, so we were able to build a solution that provides real value to them. Many medtech companies are thinking about how to use AI for diagnosing rare diseases or predicting infrequent events. I would turn that on its head. Use AI to take care of the things that eat up most of the doctors’ time. Use it to manage the usual”: Jonas Moll, CEO and founder, VitalBeats.

What are Brew Your Own and who plans them?

Copenhagen Science City planned and executed this “Brew Your Own” event in collaboration with University of Copenhagens’ DIKU Business Club, which is an outreach organisation for the Department of Computer Science and Biopeople, Denmarks Life Science Cluster, which is a network organization bridging bioscience business and academic research. Brew Your Own events are aimed at entrepreneurs interested in co-creating with academics as well as at researchers and students interested in starting their own business or working for a start-up. The events provide inspiration, information and motivation.