Responsibilities include:

• Creating synthetic training data using Text-to-Speech pipeline to improve Automatic Speech Recognition task. The main goal is to reduce the performance gaps across age-groups, making the model perform more evenly across different demographics.
• Reduced the performance gaps in ASR task across age-groups up to 13 percents by: Doing literature review to looking for the right dataset to create synthetic data from. Analyzing speed and pitch features distribution on real data to make sure the generated augmented data simulate closely the real data.Reduced total training time of significantly for the Distill-FiD pipeline using a variety of optimization techniques including half-precision training, parallelizing data processing, fused Adam optimizer and others.

Responsibilities include:

• Surveying literature and implementing strong models such as Distill-FiD for Open Domain Question Answering task. Working on tuning and accelerating models (running on NVIDIA’s DGX A100 and Selene supercomputer) to minimize computational costs.
• Reduced total training time of significantly for the Distill-FiD pipeline using a variety of optimization techniques including half-precision training, parallelizing data processing, fused Adam optimizer and others.

Responsibilities include:

• Working on distilling strong, powerful, but heavy transformers-based language model into more compact, fast-runningmodel while still maintaining reasonable performances.
• Proposed a novel approach to take advantage of large models’ performance (RoBERTa, XLM) to be used in a real time ondevices with small computing capability.

Responsibilities include:

• Working with material scientists, implementing text mining algorithms for BNL’s material science literature database.
• Built a search engine using BERT-based contextualized embeddings specifically for material science literature research, received highly rated feedback from material scientists.

Responsibilities include:

• Working with psychologists, exploring the novel idea of applying NLP tools into psychology research, using text embeddings to analyze personality questionnaires.
• Proved validity of the proposed hypothesizes, with correlation up to 0.421 (significant in psychology field). Paper accepted at Findings EMNLP of 2020.

Responsibilities include:

• Assisting in analyzing customers’ needs and designing system solution
• Implementing and deploying Salesforce Sales Cloud system

Responsibilities include:

• Proposing novel method segmenting brain images using combined Gaussian Mixture Model and Deep Learning approach.
• A jump of 8.6(/100) points to beat state-of-the-art methods and therefore accepted to WACV 2017.