Rohit and Shraddhaa | AIcrew Stories
Welcome back to AIcrew Stories — our blog series that shares inspiring stories, insights, and tips from top AIcrowd participants. This is the fourth chapter in the Food Recognition Challenge series, find past blog posts over here. ✨
Today we’re having a deep-dive chat with the veteran participants of the Food Recognition Challenge. Rohit & Shraddhaa have consecutively won three rounds of this challenge! 🤯
Keep reading to learn methods from these winners and get the motivation to further your Machine Learning journey. Comment to let us know which participants and challenge you’d like to read about next!
Rohit Midha and Shraddhaa Mohan are students at the SSN College of Engineering. They have been active participants on AIcrowd’s platform since its early days. Between the two of them, they have participated in more than 40 AIcrowd challenges and puzzles. “We visited AIcrowd when ImageCLEF and NeurIPS challenges were live", says Shraddhaa
"Our professor suggested that we could solve challenges on AIcrowd as a part of a course project”, she adds. “Solving AI and ML challenges was more fun than the usual software projects that students end up doing during their coursework”, Rohit added. “We were also getting started with computer vision and exploring classification challenges, so these challenges caught our fancy and we started solving them for educational purpose”, he said.
💪 Getting Started with the Challenge
While round 3 was not Rohit or Shraddhaa’s first rodeo, the journey from being young students exploring Computer Vision to becoming multiple rounds winning seasoned ML problem-solvers is inspiring.
We asked them to share their motivation for participating in this challenge on AIcrowd. Shraddhaa told us, “We initially participated in the Food Recognition Challenge as we wanted to explore instance segmentation. We had just participated in Google’s OpenImages challenge. Despite the lack of compute power and constrain of datasets we could utilise to train our model, we ranked 36th in the challenge. This was a positive boost for us. When we found out about the Food Recognition Challenge which had a dataset we could compute, we took it up!”
Shraddhaa added, “In the beginning, it was definitely an educational thing. The first few submissions were difficult to make as we were new to the submission format. But once AIcrowd released a baseline, making submissions became a breeze! We knew that the baselines were of great value. As a beginner, there was no way we could have made our submission without that extra help.”
Rohit and Shraddhaa created and shared their own solutions as baselines for the next round. They did this with the hope that this might help the AIcrowd community members improve their solutions and scores.
🔬 Their Approach
We asked Rohit and Shraddhaa to guide us through their journey of making submissions for the challenge.
“It was definitely an incremental journey”, says Rohit. “From not knowing much about instance segmentation to refining our solution and winning the challenge”, he adds.
“The first step was to explore the data”. Much like Rohit, one of the top participants Eric Scuccimarra had the same advice for participants starting a challenge.
“Shraddhaa made a visualisation for the dataset that highlights which classes are present in the train set and not present in the validation set. Through this experiment, we decided which classes we want to train our models on”, says Rohit. Shraddhaa adds to this, “Some classes had a lot of data, some classes didn’t so this helps us decide which classes we wanted to focus more on”. She goes on, “For prototyping, we build our model for only 176 classes instead of the total of 273 models. Eventually, we trained all the classes but identifying important classes help us improve result on weaker classes as well”.
They employed an interactive improvement system to make sure their model was always ahead of the others. “As we go through the rounds, we use the submission from the previous round as the base. We improve on it, build it for the increased dataset. That’s probably what helped us”, said Shraddhaa.
Rohit elaborated on how they trained various models. “We tested which one of our models gave what score and compared these scores across various classes to see their performance. As I said before, we tried to figure out important classes that have more data or are more frequently occurring. From these classes, we interpreted which models are doing well on these specific classes”. He goes on to explain how different rounds impacted this process, “For Round 2 the timeout was less strict which allowed us to submit an ensemble of models. We emsembled in such a way that the model would utilise the stronger classes and perform better on weaker classes”.
🚧 Encountering Obstacles
“Annotation was surely one of the biggest challenges we faced”, replied Shraddhaa. She went on to elaborate, “In some instances, only the rim of the glass is annotated as water and in some cases, the whole glass of water is annotated, this causes some confusion for the model”. Rohit agreed, “Yeah, sometimes only one french fry was annotated and in some cases, a plate of french fries. Consistency in an annotation can be helpful”. He continued, “Since the test dataset is private, we don’t know what factors are of importance so the inconsistency in annotation was a challenge”.
👨💻 Shraddhaa and Rohit's Advice For You
Shraddhaa shared her pick first, stating that “consistency, persistence and patience” are three important things required for success. “You’ll need these qualities when you’re dealing with setbacks and failure”, she adds.
Rohit added, “Baseline is an important place to start. Without Nikhil’s baseline for the early round, we would not have been able to make a submission”. His second advice was to encourage smart usage of the model rather than brute force. “It is important to use different models but it is also necessary to have the understanding of why you are trying out a specific model. Reading up about various parameters and models before using them can be very helpful and might give you an edge. This will also help you in fine-tuning it for your data”.
📈 Shraddhaa and Rohit on benefits of participating in challenges
Rohit said that in addition to the obvious improvement in his Machine Learning skills, the biggest change has been a refined understanding of documentation and how different models work.
Shraddhaa further added, “In a competitive environment, when other people start catching up you look for ways to improve your model. If I was solving this challenge by myself, I would not constantly improve after reaching a particular level of accuracy and stayed satisfied by my model. These challenges force you to improve your solution and encourage innovative thinking”.
“Yes, this also made me explore new models and approaches for the various dataset. I tried to understand why certain models work better and look under the hood”, Rohit added.
👩🔬 Their AIcrowd Research Fellowship Experience
After winning one of the previous rounds, Rohit and Shraddhaa worked with AIcrowd as research fellows. We asked them to retell the story of how that happened.
“After winning in the previous round, we got a chance to attend and present at AMLD conference. It was our first research conference so we were really nervous but we had a very good time. We even asked the people in the front row seats to click a picture of us presenting”, Shraddhaa tells us. “Soon after that, Mohanty reached out to us, offering the fellowship and that’s how our partnership with AIcrowd started, it was a great learning experience.”
“We got a chance to collaborate and work with a lot of people we wouldn’t have interacted with otherwise. We worked with individuals from WHO and Firmenich”, said Rohit.
To this Shraddhaa added, “we wouldn’t have participated in different challenges such as Learning to Smell, but having an exposure to them through the fellowship motivated us to look into the solution from an educational point of view”.
Rohit further said, “The fellowship also improved my documentation skills. Looking at Mohanty’s codes I have started writing better notebooks, explaining every function before using it”.
🔮 What’s next for them?
With a strong set of experience in the world of AI and ML, what are Rohit and Shraddhaa’s next plans? Rohit has wrapped up his internship at Goldman Sachs and is focusing on finishing the coursework for this semester. Shraddhaa, too, is focused on her final term exams. At the time of this interview, they were preparing for campus placements. They also participated in Round 4 of the Food Challenge and finished in the third position on the leaderboard.
🎮 How do they have fun?
Much like many of us forced indoors due to the virus, Rohit and Shraddhaa, have fallen into the video-gaming and Netflix wormhole. “After online classes, my friends and I hang out on Discord and play Valerian together”, said Rohit. “Yes, we all play Valerian till odd hours in the morning", added Shraddhaa. I have made more friends online through this game than I did in my pre-COVID social life”, she says. Shraddhaa shared how being in quarantine for too long has reduced the threshold of what TV shows she’ll watch on Netflix. “I recently watched Emily in Paris, it is so bad but I had fun watching it. We usually re-watch The Office episodes and I have also gotten back into watching anime”, she said. “Our life outside of AI and university is just TV shows and video games”, concluded Rohit. Which, if we say so ourselves, is a great way to live!