May 2020 – Jan 2022 · 1 yr 8 mos

• Developed complex document Question & Answering (Q&A) methods, specifically supervised and weakly supervised tableQ&A as well as neural retrieval systems that can be transferred to novel domains utilizing controlled question generation. Improved performance for tables with real-world complex structures beyond Wikipedia tables. Slated for GA in Q4-2021. Currently extending to a hybrid table-text settings like HybridQA benchmark.
• Co-lead to build WatsonNLP, a unified NLP library for all IBM products. Introduced multilingual support for 14 languages including Hindi and curated datasets and morphology rules to build models across the NLP stack: sentence splitting, tokenization, parts of speech, dependency parsing, Named entities, Sentiment for Hindi. Improved performance of named entities for low resource languages with few-shot transfer learning methods.
• Build speech-to-text RNN-transducer models for Hindi and code-switching for Hindi-English (Hinglish). Developed transformer-transducer approaches for code-mixing augmented with code-mixed text generation.
• Lead multiple scientific collaborative projects with four AI faculty members and 8 students at IIT-Bombay as part of IBM's AI Horizons program which has resulted in several publications at top AI conferences. Jointly conducted a special session at Interspeech 2021 on multilingual and code-mixing speech models.

Sep 2017 – Oct 2020 · 3 yrs 1 mo

• Conceived the Cognitive machine vision (CMV) agenda with uses customized Mask R-CNN, YOLO and traditional vision models (Hough transforms) for four pillars: detect-count, measure, track and segment.
• AgriVision: uses CMV for agriculture applications over images and videos from mobile apps and drones
• IndustryVision: built from CMV for warehouse inventory management and cloud cover prediction for solar farms.
• The toolkit was used for several POCs for various agri and energy clients of IBM. Some modules are integrated into Weather.com offerings for various agri-vision applications.

Oct 2014 – Sep 2017 · 2 yrs 11 mos

• Payment stream analysis with sub-sequence mining with modified frequent sequence mining methods to identify recurring behaviors (sub-sequences) from enriched payments time-series. Advanced finance modeling for micro-finance customers combining banking data with soft Spatio-temporal signals.
• Helped build a multilingual Covid chatbot that answers frequently asked questions, demoed and used by ICMR, Govt. of India, and received wide news coverage in 2020.
• http://researcher.watson.ibm.com/researcher/view.php?person=in-samarth.b

Jun 2011 – Sep 2011 · 3 mos · West Virginia

• I am fortunate to be a visiting scholar to WVU for the summers of 2010 and 2011 advised by Prof. Afzel Noore and Prof. Arun Ross. The project titled An Evolutionary Approach to Adaptive Fusion in Multi-modal Biometrics is based on quality assessment of each modality. Specifically, I worked on incorporating quality in the multi-modal biometric system to improve reliance in adverse military conditions. The work is supported by DoD, USA.

Jun 2010 – Aug 2010 · 2 mos · West Virginia, USA

• Multi-modal biometrics in dynamic context switching environment. Project supported by Dept. of Defense, USA.

Aug 2009 – Sep 2015 · 6 yrs 1 mo · New Delhi

• For my dissertation work, I specialize in machine learning and computer vision techniques with applications towards face and ocular recognition in unconstrained environment. Some of the key research problems that have been my focus during my PhD are:
• Face liveness detection in videos against vulnerabilities to spoofing attacks in 2D face biometric systems based on analysis of micro-expressions. The technique provides robust state-of-art results with low computational complexity.
• Face image quality assessment methods that indicate the usability of a biometric sample to uniquely identify a person in large scale biometrics. The techniques are used for sample rejection and differential processing in multi-modal biometric systems.
• Image texture analysis for face and ocular biometrics recognition for on-the-move and at-a-distance personal identification. The proposed approach is resilient to capture distances of up to 8 meters.
• Face recognition for newborns is a challenging problem that can help stem the high number of swapping and abduction cases, prevalent in many parts of the world, including India. The learning based approach is faster and cheaper than other alternatives such as DNA testing and is more reliable than RFID bracelets.
• Face recognition via derived social context: Social context derived from consumer photos
• is employed to improves face recognition performance to aid album tagging and photo management
• in social networking sites.
• Analysis of Fingerprints of Indian Masses for Universal Identification: The feasibility of using
• fingerprints as a biometric modality despite challenges such as capture inaccuracies were analysed. The studies performed as PoC, and further lead to quality standards of the Unique Identification Authority of India (UIDAI), Government of India.