- Introduced a new technique called anonymized training with optional anonymized inference and tested it using two datasets. The results demonstrated that this method significantly improves the robustness of QA models when compared to the standard methods described in the paper.

- Replicated results from the paper: Semi-Supervised Semantic Segmentation with Cross-Consistency Training (https://arxiv.org/abs/2003.09005) - Improved the result of the paper with GAN network and proposed other method of Temporal Ensembling.

– Extracting image/text embeddings to determine if two products are the same by their images and products’ description.

Team project – Designed a sentence-based emoji suggestion using fine tuned BERT model. – Deployed the model as a REST API with FastAPI and a web view chatting-box application is built for demonstration.

– Applied UNet architecture for a semantic segmentation task with 9 image classes. Trained the model from scratch with 1000 images data, and achieved about 62.15% mean Intersection of Union on the test dataset (top 10 in class).