Dennis Tsaregorodtsev | AIcrew Stories
🎉 Welcome back to AIcrew Stories!
As you know, the AIcrew blog series shares the real-life machine learning journey of our winning contestant. Our aim is to inspire more participants and share winning tricks and resources.
🙋🏼♂️ Let’s meet Dennis
Dennis originates from Yekaterinburg, situated in the Ural mountains in Eastern Russia. Before setting onto his Machine Learning journey, he worked as a Linux System Administrator at an ISP.
AI Blitz ⚡ 8 was Dennis' first challenge on the platform.
🚗 Dennis' Machine Learning Journey
Dennis's journey started due to a conversation with friends regarding the upcoming potential and advancement that AI promises. Motivated by this discussion, Dennis enrolled in Basics of ML scourses on Udemy. Once he found himself proficient in the basics, Dennis decided to enrol in the CS231n course offered by Stanford.
With instructors like Fei Fei Li, founder of ImageNet, and Andrej Karpathy, Director of AI and Autopilot Vision at Tesla, Dennis was motivated to delve deeper into the field of Computer Vision.
After experimenting with Deep Learning for Computer Vision concepts, Dennis started exploring and implementing the architectures and models introduced in the course on the CIFAR Datasets.
💪🏼 Getting Started with the Challenge
Dennis came across the Blitz challenge on AIcrowd through ODS discussion forums. Seeing how the AI Blitz 8 challenge resonated with his own niche, he decided to give it a shot. Denis enjoyed the Blitz format as each puzzle of increasing complexity challenged him to explore new methods for sub-domains of computer vision.
🐱👤 His Submission
Denis found the F1 Smoke Elimination puzzle most challenging of all the 5 problems. The untouched, open-ended task pushed him to come up with a creative solution. Before the launch of Blitz 8, the challenge organiser Shubham, created several iterations of this puzzle to make it an interesting problem for all.
The challenge winning submission saw Dennis use GCANet paired with PyTorch. This architecture utilizes an end-to-end gated context aggregation network to essentially dehaze the images.
With the abundance of available data, augmentation wasn't something that held up Dennis. He took up the time to explore more efficient CNN architectures for feature extraction. This exploration led him to Efficient Nets, which served as a defining factor for some of his other task-winning submissions.
👨🏼🏫 Dennis’s Advice for You
The thing that made Dennis's approach truly stand apart was the amount of research he took up for each of the problems. Dennis strongly advises his fellow participants to read research papers and explore GitHub repositories. Additional reading and experimentation help in training the best models.
What do you think of Dennis’s approach? What stories do you want to read next? Let us know in the comments!