Aditi Jain

SiMT models struggle to maintain translation quality while keeping latency low. In my work, we used an auxiliary language model to augment the training of the decoder model. Under this notion of target adaptive training, generating rare or difficult tokens is rewarded which improves the translation quality while reducing latency. The predictions made by a language model in the decoder are combined with the traditional cross entropy loss which frees up the focus on the source side context. This training method significantly improved BLEU scores in the English-Vietnamese and English-German language pairs, especially in the early latencies which are more difficult to translate for.

The text, image and voxels of the objects of the Text2Shape dataset are embedded in a common space using a trimodality contrastive loss. I experimented with introducing different augmentations to the text and image encoding modules, and implemented methods like SimCSE on sentence embeddings with augmentations such as word-dropout and switch-out. Compared improvements in sentence embeddings by using pretrained models like BERT, GLoVe, and SentenceBERT.

Replaced traditional hash functions with a random projection-based approach to exploit patterns in the data by projecting each data point onto a set of vectors. The vectors themselves were divided into bins (to map to indices of the bit array) and the best vector set for a particular dataset was selected after some iterations. This method showed a significant reduction in the memory footprint of a BF while maintaining the desired false positive rate (FPR). I also found upper bounds for size and FPR for the proposed method with extensive mathematical grounding. Awarded the Undergraduate Research Award for the project and laid the groundwork for a VLDB publication.

Language Model Based Target Token Importance Rescaling for Simultaneous Neural Machine Translation
Aditi Jain, Nishant Kambhatla, Anoop Sarkar, Association for Computational Linguistics (ACL) International Conference on Spoken Language Translation (IWSLT) Workshop 2023. [DOI]

Investigating the Quality of DermaMNIST and Fitzpatrick17k Dermatological Image Datasets
Kumar Abhishek, Aditi Jain, Ghassan Hamarneh, arXiv preprint [Under Review]