August 26, 2023
Last Friday marked the conclusion of my remarkable summer internship at TikTok. Also the reason why I was so irregular in writing a newsletter every Saturday for the last 3 months. Throughout summer, I was working on enhancing model explainability for an embedding model, which pretty much took all of my time.
The goal of my internship project was to bridge the gap between transformer model’s complex internal processes and our human understanding by providing a visual representation of its perception. We don’t question explainability when the interpretability is high, but we resort to explainability when the results do not meet our expectations despite checking all the boxes in model training. Model explainability is like the airbags in your car. Nobody values them unless there is a car crash. It is like the insurance you cash in after the crash. Explainability is important for us in order to ensure that the models are not making decisions based on a data point that we wish to make ambiguous. We need to make sure that the model is labelling an image as train only because it sees a train in it and not because of railway tracks. Here’s a good use case for model explainability: Google Vision Racism
I worked on the explainability of a specific architecture model based on transformer. If self-attention is the heart of transformers, you could say my project involved giving the model eyes. I delved into both gradient and attention insights to unveil the inner workings of the black box that powers machine learning models. Through this experience, I came to realize the pivotal role explainability plays as a tool. As we continue to develop more customized models for specific use cases, the need for explainability becomes increasingly vital to ensure our trust in these models. And of course the best part of the work was getting to see TikTok videos the entire day, well it was a part of the job so don't complain 🤌
And lastly but not the least, the experiences I gained and the skills I honed would not have been possible without the guidance and support of the remarkable people I had the privilege to work alongside. I am truly thankful for each individual I met and collaborated with, as they have left an indelible mark on both my professional and personal growth. I look forward to carrying these cherished memories and lessons with me in the next chapters of my career. Cheers!
X link to the internship experience
Being in the bay area also gave me the opportunity to explore San Francisco. Attaching a small gallery showcasing the ‘good’ part of SF.
ALWAYS DAY 1
Cheers!
Great blog post about internships! You've captured the essence of the internship experience exceptionally well. Your insights into the value of hands-on learning are spot on. Keep up the great work in sharing such informative content!