Startup Neurolytics tries to demonstrate with technology and science whether applicants are suitable for a certain position. In January the company became the MT/Sprout Startup of the Year. How exactly does their ingenious system work?
The latest Startup of the Year winner uses technical and scientific insights to enable companies to hire great people for the best-fitting roles. The startup founded by Belén Hein, Marnix Naber and Felix Hermsen is developing video assessments for this. Their customers can screen their own people and applicants for characteristics such as stress resilience, motivation and their match with the team- and company culture.
The Neurolytics system uses a combination of artificial intelligence and insights from neuroscience. The company’s goal: Enable companies to build better teams and offer applicants a fair chance, regardless of age, gender or ethnicity. The company has already raised half a million euros and has eight employees. We ask assistant professor at Utrecht University and Neurolytics’ head of R&D, dr. Marnix Naber to explain the technology in more detail.
Have you built this machine learning system yourself and how exactly does this work?
“Yes we have. Our technology is a combination of open source and proprietary algorithms, all strung together by Neurolytics. It consists of our unique codes. We use a computer vision and artificial intelligence system to detect detailed information through a webcam. We then analyze the applicant’s behavior to determine psychological characteristics. This particular part of science is called psychophysiology. We use machine learning to collect all data and to build ethically responsible models, so that we can measure psychological characteristics in a scalable way.”
Let’s suppose I apply via your online portal. What will you be looking at?
“You can run the scan online via your internet browser. You will receive a link, that will lead you to our online platform. There you’ll watch company videos, answer interview questions and perform a few cognitive challenges. In this process we can add time pressure, or unexpectedly change tasks very quickly. This gives us the opportunity to analyze what happens to the candidate. We receive behavioral data that correlates with the mental state or intentions of the person. We see it immediately when someone gets stress from particular elements or is less intrinsically motivated for a certain topic. For example, if you’re not great with numbers and don’t like math, we will see your frustration and heart rate increase.”
How does that system actually learn to recognize facial expressions?
“We don’t necessarily focus on specific facial expressions, but more on facial muscle movements. These insights can be combined with certain emotions, but we mainly look at unconsciously controlled changes, such as eye movements. We have set up a controlled database, which contains many candidates who have taken our scans. They have completed questionnaires, so that we can better analyze what is happening with their behavior. It took us a year and a half to build that database and we are still building it out.”
How many people are already in that database?
“Hundreds, and with every test we do, the number increases. This way we continuously obtain more relevant information about behavior. The more data we get, the more accurate our system becomes.”
Some faces are more sad or happy face than others. For example, someone can laugh very little, but still be very enthusiastic. How do you prevent issues in this area?
“The scan starts with a baseline measurement for each person. With every assignment we look at the changes in the face compared to the baseline measurement. If your neutral face looks a bit more negative, then that is your baseline measurement. Then you have a somewhat gloomy face as a baseline measurement. Others can have a very happy baseline, that is also possible. Also for them, we look at the changes compared to the baseline measurement. So someone’s natural emotional state doesn’t matter in our approach.”
Which psychological models do you work with? In other words, how do you translate an expression into behavioral qualities?
“Our models are bases on psychophysiology. This involves the activation of the central nervous system. Among other things, your movements and behavior are points that we analyze. We want individualized measurements, so we look at many different muscle movements and other data points. If you only look at holistic emotions, it will produce less insightful results. This is because someone can smile, while he or she actually feels a different emotion. In addition to the facial analysis, the scientific questionnaires we use are vital to get the best insights. We relate those to our video data. “
How do you avoid bias and privacy issues?
“Everything is GDPR-proof. We have spent a long time and a lot of effort to take away any concerns on this. All models are set up scientifically, we inform candidates of the full process and applicants give permission before they start the scan. We also continuously check our verified datasets for bias. Fortunately, so far we have not found any bias in the datasets. And if we would find any in the future, we can take out the bias and adjust the models. “
This article first appeared here.
Download The Paper
Are you curious about our product and want more in-depth information?
Leave your mark as an HR leader and hire the right motivated candidates. We’re redefining the candidate experience and excellence in hiring.
*By downloading this paper, you consent to receive news from Neurolytics.