AIcrowd Blitz 02-16May'20 - Wrapup Summary
By
debalb
Following blog is a summary of the AIcrowd Blitz competition organized from 02-16May’20.
INTRODUCTION ⚡
The first ever AIcrowd Blitz competition has recently concluded after an exciting two weeks of tough competition for the top three positions.
The competition was organized between 02-16May’20 and saw participation from more than 350 participants with 3.5K+ submissions. The majority of users were from India (260) while others were scattered from different parts of the world.
MOTIVATION 🏫
In case you are wondering what all the buzz is about, AIcrowd Blitz is an event where participants are challenged to solve a bunch of carefully curated Machine Learning problems within a fixed timeframe of 15 days.
AIcrowd Blitz was started with a vision to serve two purposes simultaneously, firstly to ease the entry of newcomers into the field of Machine Learning as the first step is usually the toughest, and secondly to provide a medium for the experts to practice and test themselves while hopefully learning new concepts.
Through this competitive environment, we want to induce the art of solving open real world problems in the community and hope that we can harbour collective learning along with fun, excitement and growth of the whole community.
TASK 📑
In this recently concluded version, the participants were given the following 5 problems to solve, using any platform and programming language of their choice:
The Tasks were made to strike a balance between the target audience which includes both newcomers and experts.
The problems were curated from classical Machine Learning categories like image classification, predictive analysis etc with difficulty levels ranging from beginner to intermediate.
In order to help fresh participants, starter kits were provided for each problem containing a sample, ready to submit code.
PRIZES 💰
Though our aim is to incentivise through learning, a competition is never complete without exciting prizes for the worthy winners 😉 !
For the top 3 participants, cash prizes of INR 7K, 4K and 3K were announced respectively.
LEADERBOARD 🥇
The competition turned into a battleground with the active participation of the contestants, especially in the last few days where more than 1K solutions were submitted and the top of the leaderboard saw frequent changes almost like a game of musical chairs !
In the end, the top-3 positions in the leaderboard stood as under:
sawar_sagwal secured the coveted 1st place by top scoring in 4 out of the 5 challenges (the only participant to achieve this feat).
2nd place winner gaetan_ramet was not very far behind, closely followed by team Hard_Drive_Corrupted (consisting of Anchit and Pulkit) grabbing the 3rd place.
While it was seen that quite a few participants top scored in more than one challenge, in the end what helped one climb to the top of the leaderboard was the overall rank in all 5 challenges.
HIGHLIGHTS 💎
We reach out to some of our participants about what were the key highlights for them in this challenge.
“Highlight for me, is the diversity of the challenges ... I especially like the fact that the Blitz is composed of 5 different challenges that are very diverse in terms of what they require and what type of model you can apply or what type of data you are dealing with. So this is a very interesting and very challenging problem. You deal with diverse data and solve a lot of different problems in a short time like this.”
Team Hard_Drive_Corrupted:
“As a highlight, I think, was how towards the end, both of us were really scared because the guys on the 4th or 5th position were really catching up to us and we weren't making any improvement. I think the last 2 days were really thrilling because we were training models, again and again, to see if there is any improvement that we can do or something like that. I think that was the highlight and it really gave us a lot of joy whenever we improved our score and saw our rank on the leaderboard go up.”
The starter kits released for each of the 5 challenges proved to be quite popular and much appreciated by the participants in general.
“I think the whole thing was quite well made. The notebooks and starter kits and basically explanations about the data and all this, I didn't find too many problems with them. I think you had one or two updates in some of the challenges but you guys are quite reactive on that, and that's great. I'm well experienced in Machine Learning, but even if you are not, it's possible to get a decent score in these challenges, given the resources that you provide.“
To quickly cover a few interesting statistics, the Diabetic Retinopathy challenge had the highest number of solutions submitted (1400+), followed by Plant Disease Recognition (700). Poker Hand turned out to be comparatively less popular, possibly due to an algorithmic approach to the solution.
WINNING STRATEGIES 🏆
So what does it take to get that winning solution?
Who better to answer this question than our top 3 winners.
“In most of the challenges, what helped me the most was that I looked at the data set thoroughly. For the Food Challenge I used an Ensembling method where I trained two different models, one was the DenseNet model and another was the EfficientNet model.”
“I quickly realized that, for Poker Hand and Rubik’s Cube Orientation, it was not necessary to go into Deep Learning or even Machine Learning oriented models, but something very algorithmic for the Poker Hand was quite obvious in my opinion. ”
Team Hard_Drive_Corrupted:
“We really searched the internet on how to attack these kinds of problems, as both of us are new to this type of problem. We saw different techniques, like in FOODC, there was something called Test Time Augmentation, wherein at test time you augment the images and you take the prediction and you do a majority voting on it. So things like that, even Progressive Training wherein you train at a lower scale, then you train again at a higher scale.”
FEEDBACK 😇
In an effort to keep improving our events, we reach out to a few participants and ask them for suggestions, which they would like to see in future editions of the Blitz.
“it would have been better if only a part of the test set was released and some part was released on the last day of the challenge. I thought that I was, sort of, fitting my model on the test set by training the model, submitting on the platform and trying to get the best score. It would have been better if only a part of the test set was released upfront.”
“On some of the challenges you can achieve like 100% performance quite easily. For example on the Mini Leaves, you will quite quickly get to 0.99 something and then maybe it's a bit also luck based. If you are first, second or third place, everyone is very close together. It can be interesting to have challenges where the top scores end up being a bit lower.
I think it would be nice to have a general discussion channel, not only for one per challenge, but also one for the whole Blitz, for general announcements etc.”
We also asked them about the kind of problems which they would like to see in the future events.
“I would like some simpler version of the GANs and the Object Detection challenges. I would also like some challenges which include traditional Computer Vision approaches, not necessarily the Deep Learning and Machine Learning approach, where you can sort of use mathematics and then tackle the problem”
“It could be interesting to have something either from Reinforcement Learning or Natural Language Processing topics, since these were not really covered.”
FUTURE ROADMAP 🔮
Though the competition is over, we don’t want the excitement to dip. Learning must go on !
Hence we have launched AIcrowd Practice, a bunch of practice challenges designed With flavour similar to AIcrowd Blitz but with a wider range of problems to let you practice and play the field.
In addition, we are working on our first ever live webinar, where the winning teams from this edition of the Blitz will share their knowledge with participants by showcasing their ideas and solution. To make the event more exciting, they will be answering live questions from the audience.
So while we all wait for the next edition of Blitz, we have a lot going on, sharpen your blades, practice a lot, soak up in the events and gear up before we meet next !
We hope you enjoyed participating in this edition and would love to hear your feedback.
Do share your suggestions at hello@aicrowd.com and stay tuned for more such exciting events coming up soon.
Before we close, we want to share with you this beautiful visualization created from code submissions in the first ever AIcrowd Blitz! It shows how the participants competed to solve the 5 exciting problems.
https://www.youtube.com/watch?v=EOVqnS1jSjo
Comments
You must login before you can post a comment.
You may also like...
debalb
marcel