Maciej Zelwert - Senior AI/ML implementation engineer at EY

Kornelia Trzęsowska: Tell me something about Yourself.

Maciej Zalwert: I have joined EY 7 months ago and I am a part of Artificial Intelligence Build Center here in Wrocław.

I have done my Master Studies in Finance at Wrocław University of Economics. I am passionate about Artificial Intelligence and broad aspects of risk. I have started my journey with AI during my studies. In my free time I like jogging. I do it quite often and, surprisingly, it turned out that since the New Year I have managed to reach 380km of running.

KT: What is most interesting for you in the concept of risk? How do you perceive it?

MZ: Risk is very hard to predict. We do our best to have control over everything, but these are only efforts and don’t always work. It is very hard to predict risk especially when it comes to machine learning that heavily depends on varied data.

KT: How did your adventure with Artificial Intelligence start?

MZ: During my studies we tried to create an AI that would help to manage the economy resulting in high economic

growth. Then, after graduation I worked in market risk in the department of financial risk in a few big banks. I was involved in projects like creating a deep neural network to identify financial fraud and identifying outliers in terms of pricing financial instruments.

KT: How would you describe AI?

MZ: It is an attempt to replicate the biological neurons in the decimal system. You can find examples in the intelligent self-parking car systems or face recognition tools. It is a method to resolve problems that can help humanity in some tasks or even to replace the human’s work. That’s why we call it intelligence. Some argue if AI is really an intelligence or only a complex algorithm used to find some dependencies of data. Maybe with the era of quantum computers we will have the next AI revolution showing its real possibilities and power.

KT: What exactly do you do with AI at EY?

MZ: In my role I am responsible for building an AI model to discover the scope of an audit in an external company so in fact to help auditors to do their job. They conduct audit in a company, but the AI may help them to find the areas they should concentrate on and to detect the most significant accounts. It will help them adjust their methodology to the company and its unique case.

Another internal project was to capture the risk in AI. This is more interesting because over the years of developing artificial intelligence, almost no one ever focused on thorough research on the risks it can pose. The computation power is way faster than before and since several years we can use AI on a larger scale. There are a lot of models, we use them, but no one has focused on the risks of these models before. To illustrate with a simple example, let’s think of a model for so-called self-driving cars. There is always the probability that AI will not pick up something, like a passerby on the street. Fortunately, the probability is really very very low, but it should be measured in a proper way. It is a quite unique topic that’s why we decided with the team to write a publication about it.

The typical example of usage of our unique model is for example — in the company we have a model that removes water-marks (from invoices and other documents). This piece of paper is scanned by the system and then takes numbers from it. Sometimes it may happen that the model instead of recognizing digit 1 recognizes digit 7 — in other words instead of a million we have 7 million. The company would have to pay 7 million dollars, which is obviously a problem for us. Invoices are checked manually by our colleagues, but it is very expensive and time consuming. Our Prediction-at-Risk (PaR) model can capture the statistical risk of such a situation for most artificial intelligence models.

KT: Tell us something more about it!

MZ: In the world of science, people come up with different state-of-the-art models, technological innovations and give it to the world. Our paper is very innovative, because no one has focused on it before. This can be attractive because EY can use this concept in auditing artificial intelligence models from other companies for the risks and consequences of what may go wrong. Our solution is already implemented internally. We provide statistical proof that this AI solution will be operational. We want to publish our paper in one of the scientific journals, but first it must be officially checked, i.e. pass the so- called score system to make sure it meets the quality standards to be published under EY affiliation.

KT: Why is AI worth getting to know?

MZ: The topic of AI is impossibly interesting and broad, because it’s vigorously developing. It is without any doubt, the future. It gives us countless possibilities, so if You like programming and innovations — it is definitely an area for you.

KT: Where should we start then?

MZ: YouTube and Udemy are full of different videos about Artificial Intelligence. For the beginning of the practical part of AI I highly recommend the Scikit Library in Python3. You will find there many simple tools possible to use in AI. It will help You write some simple algorithms from the area of machine learning. Worth checking is also PyTorch — the most popular framework to write the deep neural networks.