Hello can, thank you for being here with us today!
Please, tell us a little bit about yourself:
Hello everyone, thank you for having me!
My name is Can Karabey, I am originally from Germany, and I moved to the Netherlands 3 years ago to study Applied Mathematics at Inholland University of Applied Sciences.
At the beginning of my 3rd year of studies, I held my internship at MM Guide, where I worked in the research and development department. Here I worked on an image and classification system based on deep learning. Currently, I`m doing a minor Android development while also working part-time as a data scientist at Capte in Amsterdam.
WHEN AND HOW DID YOU GET IN CONTACT WITH MM GUIDE?
AND WHY DID YOU CHOOSE OUR COMPANY FOR YOUR INTERNSHIP?
At the beginning of summer 2021, I asked one of my lecturers for help since it was an extremely complicated period, and it was quite difficult to find an internship. Luckily, this lecturer was part of a committee alongside Bastiaan Slop (MM Guide’s Business manager), and he gladly referred me to him. After an interview with both Bastiaan Slop and Israël Jamlean (MM Guide’s operations manager), we agreed that we were a good match, so I started my internship on September 1st, 2021.
MM Guide was the best choice for me because the company is not focused on only one product. Since MM Guide works with different clients, it was a far more exciting choice since it would have allowed me to follow different projects. Furthermore, I heard positive feedback from other students who previously collaborated with MM Guide, and this was for sure something that caught my attention.
WE ARE HAPPY TO HEAR THAT OUR REPUTATION PRECEDES US!
I HAVE A TRICKY QUESTION FOR YOU: WHY DO YOU DO WHAT YOU DO?
This is indeed an interesting question!
As expected as it might sound, what I want is to have a positive impact on the world. While other people want to become a doctor or something similar, I think that the tech industry is often overlooked, despite all the positive impacts it has generated over the years. Many exciting things are being developed as we speak (like algorithms that can detect tumors through MRIs) that can have a profound influence on people’s lives. This is one of my primary drives and what keeps me interested in this field day after day!
THIS IS INDEED A GREAT ANSWER!
WHAT ARE YOU WORKING ON AT THE MOMENT?
Right now, I am developing a mobile app with other students. Before I tell you more, I admit its use is quite a niche one, but its development is really interesting. One of my lecturers is part of a LARP (Live Action Role-Play Game) organization, and the app we are creating will provide the players with several tools. They will be able to upload their character sheets directly into the app, send direct messages to other players and receive updates regarding what is happening in the game universe. This app makes the whole game experience more in line with the current technologies, aiding both the players and the organizers to optimize the whole gameplay.
I also have a part-time job as a data scientist for a start-up in Amsterdam. The company offers IoT solutions for electric vehicles, mainly for electric buses. My main task is to elaborate the data to optimize several aspects of the project: range estimation models, smart charging schedules, and so on.
INTERESTING!
BACK TO YOUR INTERNSHIP WITH MM GUIDE, COULD YOU TELL US A BIT MORE ABOUT IT?
Sure thing!
The main task of my internship was to create an automated quality assurance tool using AI technologies. To do so, we started with preliminary research to figure out the best options and technologies. We also took into consideration previous solutions adopted by other companies for quality assurance automation and image specification systems. We then listed the possible approaches and did some basic tests to figure out what was possible to achieve with each of them, how they perform, how difficult they would be to implement, how scalable they are, and what kind of resources they would require.
After this initial testing phase, we decided that using a deep learning approach was the most suitable solution. The other options we evaluated were the OpenCV library (the most used computer vision library) and traditional machine learning algorithms like a support-vector machine or decision tree using the scikit-learn library.
However, we quickly found some impediments with these two solutions: OpenCV is not user-friendly nor customizable to our standards, so it was quickly discarded. Traditional machine learning algorithms are not able to make use of graphic cards for calculations so they aren’t scalable enough. Deep learning algorithms, on the other hand, have higher accuracy scores on image classifications compared to the other solutions and can utilize the GPU to do matrix calculations which speed up the process, and this is the reason why we chose this approach.
The standard for image classification tasks is a convolutional neural network, which we also utilized. This solution gave the best results in extracting information from images. We used the Python library keras (which is built on Tensorflow) to train the neural networks.
The next step was to get data, create the model and optimize it.
THANK YOU FOR THE DETAILED EXPLANATION!
HOW WOULD THIS SOLUTION HAVE HELPED THE CLIENT?
DID YOU ACHIEVE SIGNIFICANT RESULTS?
To better understand the reason behind this project, you should know that it was developed to establish a preliminary contact with ANCAP, an Italian porcelain factory. One of the pain points of ANCAP`s production line is the number of people assigned to the quality assurance of the products. Reducing this number means that the employees can be relocated to other areas of the implant, where their knowledge would be more impactful.
If properly implemented, our solution would have reduced the number of people required for quality assurance by circa 83%, bringing the number of people required for this task down to one, just for double-checking. Another advantage of this solution is that a computer does not need to rest and can run 24h nonstop. This would have been more efficient in terms of both time and money consumption.
To answer your second question, we indeed achieved some interesting results while we also faced a very complex challenge. Training this model requires a lot of images. And when I say a lot, I mean it. Sadly, the amount of data we received was not enough to reach a production-ready accuracy (100% is not realistic). However, by using different techniques including data augmentation (AKA creating slightly different images from other images) we managed to artificially generate more data for training our model. At the end of the process, we reached an accuracy of 95%. Not too bad for a model trained on a small amount of data!
IMPRESSIVE!
WHAT WAS THE MOST VALUABLE NOTION YOU LEARNED DURING YOUR INTERNSHIP? AND WHICH WAS THE HARDEST CHALLENGE TO OVERCOME?
I think the most relevant notion I have learned is how to tackle a project like this: how to research, how to test, and how to deal with a lack of data. I would say that solving problems related to data science and deep learning is what I will carry with me for my whole career. Furthermore, I had a chance to learn more about teamwork and communication and these two aspects will be relevant in any future work environment. I also think that this project gave MM Guide a better idea of how a similar project can be approached. The company now knows what to expect, what can go wrong, and what the requirements are.
As I mentioned earlier, the hardest challenge was the lack of data. Data is an essential part of machine learning, so I had to find some workarounds to solve the issue. The lack of data also meant that I was not able to implement the project as planned, and this issue influenced the report I was writing for my university. Luckily, I was able to change the route and create a neural network to test on different scenarios. I used the neural network to detect rotten plants, cracks in marbles, and cracks in tires. In all the scenarios the accuracy was over 90%, meaning that what I developed could have been used in a real-life scenario.
Of course, the acceptable accuracy threshold depends on what you are using the model for. For ANCAP, the current accuracy threshold is close to 100% since the quality assurance is done manually by humans. It would be quite hard to reach the same accuracy level with an automated machine, but I think that with the right amount of data it would be possible.
THANK YOU CAN!
ONE MORE QUESTION: HOW WOULD YOU DESCRIBE YOUR EXPERIENCE WITH MM GUIDE AS A COMPANY?
I think the best term to describe it is pleasant. The whole environment of MM Guide is calm and organized compared to other realities. Despite the fact that I was in the office only for a short period each week due to the COVID situation, the atmosphere there was great. I really got along with everyone and there was always the chance to make small talk as well as ask for ideas or suggestions.
I think people are what makes MM Guide unique and extremely valuable. I found everyone very competent and professional, and this really contributed to my experience. Everyone was always eager to share their knowledge with me and I never felt left alone or overlooked.
On a side note, I would like to emphasize how the guidance of Israël Jamlean helped me to properly draw a plan for this project. His support was constant and essential for creating a clear path on what to do and where to move on next.
I honestly think that MM Guide is a great environment for a student to learn and I will absolutely recommend it.