Gaurav Singhal | AIcrew Stories
Welcome back to the AIcrew Stories — our blog series that shares inspiring stories, insights, and tips from top AIcrowd participants. This is the fifth chapter in the Food Recognition Challenge series, find past blog posts over here. ✨
Today we’re having a chat with one of our most persistent participant who, after several tries, has won round 4 of the Food Recognition Challenge.
We talk with Gaurav Singhal, who won the fourth round all the while finishing his Masters’ thesis, working at his day job and running his own mentorship program.
Keep reading to know Gaurav’s story on how he got started in the field of Data Science, used the AIcrowd community to write his research thesis and dealt with failure.
As always, do let us know what you’ll like to read next by dropping a comment or tweeting us @AIcrowdHQ.
Gaurav Singhal started his career as a software developer. He has extensive experience in python development, mentoring and ERP development. After working as an ERP expert for almost 4 years, Gaurav wanted to branch out and pick up different skills. After his work hour, he started exploring different development language — initially playing around with JAVA and finally focusing on Python. Through my experience and connections, I also started my own mentoring company — code alphabet. Around this time, he developed an interest in Machine Learning.
To pursue his interest in ML, he enrolled on a Master’s program in Data Science at Otto Von Guericke University in Germany. He currently works at Oviva AG. Before he moved to Germany, Gaurav lived in Jaipur.
🤖 Gaurav’s Machine Learning Journey
We wanted to know more about Gaurav’s journey in the field of Machine Learning. He said, “During my Masters’ degree, I was involved in NLP projects at university. Then I joined a company that works in healthcare. The Food Recognition challenge addresses an important problem in the healthcare industry. By understanding what items and portions a person is eating, we are better able to guide nutrition and prevent diseases like obesity. For this purpose, I use ML tools both for university projects and for my work.”
His Masters’ degree is about the theoretical aspects of Data Science. Gaurav said that he chose to study at Otto Von Guericke University as Germany has an affordable education system and great research infrastructure. It also had a booming data science industry and that aligned will with his goals for the future. he wishes to further his ML journey by focusing on research and publish more papers.
Gaurav still sees himself as someone who is slowly transitioning into ML research. “I am currently collaborating with Mohanty on writing this paper. My company has a strong research culture and they encourage innovative thinking”, he says. Gaurav wishes to contribute his work and see them integrated into the product, slowly shifting into the research community.
🦾 Getting Started
Gaurav first started with the Food challenge back at end of October, early November. Around that time, he was also looking for my Master’s thesis topic. At the same time, he was also working at a company. He says, “I was managing this competition, my thesis and my day job. For my work, I had to perform portion size estimation, for that we didn’t have the data. Although my work isn’t around food portion, this challenge helped me. Through this challenge, I am able to submit my thesis as well.”
Talking about his submissions, Gaurav recounted, “In the third round, I tried a lot of approaches — but they were in the wrong direction, I guess. I tried the detectron2 Facebook API model. I even reached 3rd or 4th position using it. But then Eric rocketed to the top and I got shifted to the fifth position.”
Gaurav showed resilience for round 4 and started from scratch. He described his experience, “I tried a new method. With a fresh new approach that I hadn’t used before. I used MM-detection along with series of experiments and finally reached the first position.”
🥇 The Winning Method and overcoming obstacles
Gaurav says that when he started, he was new at instance segmentation so he tried a basic mask and this approach did not give him a good result. After this, he started exploring other approaches like detectron2 and MM-detection. “I did a rudimentary analysis between these two models and found that detectron2 is better in terms of computing and performance, so I went ahead with it”, says Gaurav. According to Gaurav, the challenge was to understand the flow of this library. “I tweaked other hyperparameters and despite that, I was scored at the score of 44%. After this, I tried different ideas for improving the ground truth data. With this, I reached 3rd or 4th position but soon I noticed Mark and Eric jumping ranks and taking a top position”, he added.
Seeing the performance of the participants Gaurav started to wonder about what the other participants were using. “This made me wonder that there is probably some approach or tool out there that I am unaware of and missing”, he adds. Gaurav decided to tap into the community and chat with one of the top participants. He says, “When round 3 got over, I had a chat with Eric, getting to know him and how did he approach the problem. We were sharing ideas on how each of us attempted to solve the challenge and I realised that our approach was very similar except for one thing. I found out that detectron2 does not provide some libraries and architecture that MM-detection provides. So for round 4, I tried MM-detection with a few hurdles”. Gaurav decided to turn the failure of round 3 into a plan of action for round 4, “It felt like I had wasted the last three months but I wanted to improve my solution using another approach.”
For round 4, he tried various architectures of MM-detection and finally, he read literature on A2RS(?). “I thought it would perform faster. It didn’t, but the performance was more accurate”, he says. To further refine this model, he purchased Google Colab and performed the experiment on it. Due to the limited compute available and the time it took to train the first model (7 days) he trained with the best parameters he had from the masked(?) HTC architecture. With a lot of tweaking and improvement, Gaurav was able to reach a score of around 53% on the leaderboard. The score was still not good enough. The top participants were doing better, so he read a few more papers and learnt how ensembling of model can improve results. “I tried the approach but it was time-consuming. From a research perspective, it is an exciting avenue and I will try to publish in that direction. But for this competition, I was running out of time and couldn’t go further in this direction”, he says. To counter this, he stuck to the basic ensembling of the model by removing the existing accurate predictions and bounding boxes based on some inferences. “I tried bayesian statistics to remove the bounding boxes — this helped my research but not so much for the competition”, he concluded.
👨🏽🔬 How did the challenge help your thesis?
“The challenge helped me get the data,” says Gaurav. “The other segmentation data had laboratory settings, which nobody wants to use. The other dataset focuses on a specific type of cuisine, or had standardised background”, he adds. Elaborating on the uniqueness of this data, Gaurav says,” In reality, we will just click the image of food on the table without setting it up as we will post on Instagram. For practical usage, training models on pretty pictures won't be very helpful. It won't be able to do prediction on real pictures. For this reason, the AIcrowd data was very helpful.”
But data was not the only benefit of participating in this challenge. Gaurav emphasised the need for having a collaborative research community. “For thesis, motivation is a key ingredient”, he says. “If I were to conduct research by myself, it would have been exhausting. Reading up on previous work and papers, how they can be improved. Through this competition, I was able to compete with so many participating trying out various methods. Without the competition, I wouldn’t have explored so many different methods and approaches to solve the problem. I would have built on a basic model I initially created rather than trying out other architectures that exist.”
Gaurav would like to see more such research challenges on the platform. “The community is very active and interactive. Challenges were the dataset is used for next 4-5 years would be of great value to the domain”, he adds.
💻 What do you think about AI’s future?
“AI will surely be used in the public domain and help everyone a lot. But explainability remains the biggest challenge”, says Gaurav. “Trust in AI model is required for it to be adopted widely by the population. To trust a technology, we must know how it is built and its functionalities. So, I believe, explainability will be the main area that researchers will focus on”, he adds. He details his experience in the field of NLP and shares the example of OpenAI GPT4. “It is a famous model that uses a huge infrastructure but it lacks explainability. People are unable to explain the results, intuition is there but openness in output is lacking. The model also requires a lot of computing and is not easily accessible”, he adds. Gaurav finally concludes by saying, “AI will surely aid people but it needs to earn people's trust before being adapted.”
🔮 What’s next for Gaurav?
Gaurav’s current focus is to publish top-quality research papers in reputed publications and conferences. This would help his PhD applications. “I want to pursue my doctorate in an elite university at a top research lab and for that my application and work needs to stand out. The research track that I am on would help me reach that goal”, he says.
When he is not juggling research, challenges and work, Gaurav can be found playing the guitar, cycling or swimming. He particularly enjoys cycling in Germany as there are dedicated tracks for it and people are respectful o cyclists. Back when he was in India, he made a trip to a different part of the country every two months or so.