Tverdov Konstantin | AIcrew Stories

Welcome back to our new series of AIcrew Stories! 🎉

👋 In the third episode of AIcrew Stories, Vrushank from AIcrowd met up with Tverdov Konstantin, one of the runner-ups of Round 3 of the Food Recognition Challenge.

✏️ Who is Tverdov Konstantin?
💪 Getting Started with the Challenge
🔬 His Approach
🚧 Encountering Obstacles
👨‍💻 Konstantin’s Advice For You
📈 Next Steps

✏️ Who is Tverdov Konstantin?

Tverdov Konstantin, like Mark, is a student of the Moscow Institute of Physics and Technology. He is part of the Department of Innovations and High Technology. He is also currently working with a company on ML related tasks.

During his undergraduate studies, Tverdov took various courses in machine learning. This piqued his interest in the field. He shifted from doing programming to Machine Learning as he found it more interesting. He says he finds the field of ML more fun because “you can see your solution working’.

💪 Getting Started with the Challenge

At work, he dabbles with food and nutrition-related assignments that use Machine Learning. While those tasks are quite different from the challenge, they provided the initial motivation for him to sign up and take part. Recounting the story he says, “I don’t exactly remember how I got started with the challenge. While being interviewed for my current job, someone shared the Food Recognition Challenge as an additional resource. In November, I noticed the launch of a new round. So I participated in the challenge to explore new ideas”.

🔬 His Approach

Tverdoc said he didn’t try a very unique effort because of the lack of time. He was also competitive and wanted his model to be close to the top submission. He used MM-detection for object detection and from there he tried several models. At first, he used the htr_r101_fpn model and refined it. He worked on augmentations for tweaking contrast and brightness. Those things helped improve the solution. He also created his own bounding boxes from masks rather than using the masks from annotations.

🚧 Encountering Obstacles

He found the submission phase of the challenge most difficult! Konstantin elaborated by saying, “It was a bit hard to submit. Not because of the docker but because of the lack of information on failure. Not knowing why the submission was failing made trouble-shooting difficult. I understand that having logged on to a private test set can lead to leakage and cheating. But when you spend a day refining the solution and it does not work, it is confusing.” We took his feedback and implemented a better submission log with a description of the failure.

Konstantin is yet another member of the popular ODS community of Data Scientists. As someone who is a part of an active & collaborative community, Tverdov had some insight for us. He suggests a large-scale sharing of ideas and approaches of the previous winners of challenges. “Not knowing how the winners of previous rounds got the score makes improving them difficult”, he says. He adds, “Often I didn’t know if I was bringing in a new solution or reworking a previous idea”. We will work on this wonderful feedback — creating a collection of shared information for the future competition!

👨‍💻 Konstantin’s Advice For You

Konstantin received an AP of 0.506 with only 23 entries. So we asked him to share his secret sauce. His top advice to participants is to practise! “Take part in more competitions. Try to read and learn from public information and solutions created by others. The field is developing so read a lot of good articles and resources, learn from them!”, he adds.

📈 Next Steps

Konstantine said he thoroughly enjoyed the challenge! At the time of the interview, he was preparing for round 4 of this challenge. Due to falling ill in the last few weeks of round 3, he wasn’t able to experiment with new approaches. He is looking forward to implementing those ideas in round 4 to see how they perform on the new dataset.

When not coding machine learning solutions or busy with schoolwork, Konstantin enjoys outdoor activities. In summer, he enjoys riding a bike around the city of Moscow. In winter, he visits a mountain ski resort in Sochi with his family. He enjoys reading, especially non-fiction books. He recommends “Surely You’re Joking Mr Feynman” by Richard Feynman. This book by a famed physicist is his favourite. You can see other projects by Konstantin here.

Check out episodes one and two for AIcrew stories here.