I was the only undergraduate on a four-person core development team that participated in the Microsoft Global 1 Million Celebrity Face Classification Competition (MS-Celeb-1M), the world’s top ML competition with the largest dataset for both large-scale and low-shot classification. I not only helped to lead an intern undergraduate team to preprocess data but also contributed to the core development team by training models. For example, I worked on training 121-layer Densely Connected CNNs, experimented with a novel enforcing scheme for the standard Softmax objective function, and created a new structure, Conditional Dual-Agent Generative Adversarial Network (CDA-GAN) to generate more realistic synthesized faces along with neural-style transfer from my previous research into the low-shot problem. In the end, we achieved the first place in all subcategories of the competition.
These experiences not only prepared me for graduate school and allowed me to familiarize myself with lab work; they also pushed me to the frontier of CV. Powered by massive graphics processing units from industry, I pushed the limits of large data classification, and gained new insight into the difference between industrial and academic research. Seeing the power of academic theories applied at the industrial scale encouraged me to dig deeper into academic ML research to contribute to new directions. I was inspired by this work and now want to create more robust, generalized models that could be less limited by data and computational power.
