My Experience at OffNote labs [Azhar]
Intro
I was like any overly ambitious undergraduate student looking for opportunities to gain some valuable real world experience during the summer break when through some stroke of luck I stumbled onto a LinkedIn post about self-supervised learning and an opportunity to conduct independent deep learning research by Dr Nishant Sinha via OffNote labs. Through some fortuitous events I was able to land a meeting with Dr Nishant where we discussed unsolved deep learning problems and research. Being inclined towards pursuing research in deep learning I knew right away that this was a very unique opportunity to work on really hard research problems with the added advantage of being mentored by someone with a great deal of expertise in deep learning and research in general.
The Problem
With my interest in deep learning for computer vision and OffNote Labs’ goal of democratizing and making AI accessible to the masses in mind, we decided to work on scene text detection and recognition for low resource Indic languages with the moonshot goal of producing a single model capable of handling multiple languages continually (i.e., learning to detect and recognize a new language without forgetting about previously learned languages). However, this problem is not trivial as neural networks are susceptible to forget learned knowledge when the input changes due to a phenomena named catastrophic forgetting, this still remains the greatest weakness of deep learning models.
The Process: Read, Experiment, Fail, Repeat
Hurdles
Being an uninitiated researcher only just taking baby steps towards doing actual research, I was in for disappointment when my initial attempts fueled with naive optimism at tackling the problem failed miserably. The humbling experience coupled with advice from Dr Nishant taught me an important lesson, that is to make any sort of progress towards a research goal one has to accept the fact that there will be a lot of setbacks and the more quicker and comfortable you become in dealing with these setbacks the easier it will be to navigate hurdles thrown at you in the future.
Learning
Apart from dealing with failure I was also able to pick up a few vital skills like
- critically assessing the space of possible solutions to try because there is always an associated sunk cost.
- Picking the right papers and reading them the right way
- being disciplined in organizing experiments
- Succinctly conveying my thoughts and ideas
- the ability to rapidly implement ideas in code from scratch.
Additionally the constant support and mentorship through discussions helped produce valuable insights, these insights coupled with many more iterations of reading, ideating, experimenting refined my initial efforts into something worthy of being published at CODS-COMAD 2021, a premier conference.
In parallel, weekly virtual meetups presented an opportunity to share my work with other aspiring researchers and hone my presentation skills. Invited talks AMA style by deep learning experts from Facebook, Target, to name a few, gave me rare insight on how deep learning is used in industry.
In Conclusion, my experience at OffNote has given me a taste of what research truly is at it’s core — an exercise in creation and refinement of ideas.
