Amrita Saha  Amrita Saha photo       

contact information

Software Engineer, Research
India Research Laboratory, Bangalore, India




Research Interest

Amrita has been working both on various core problems in Machine Learning like kernel learning, structured prediction, unsupervised graph learning, as well as various applications in Machine Learning, Natural Language Processing and Information Retrieval over the last six years. Recently she has been keenly interested in Deep Learning for various problems in Representation learning, Natural Language Processing and Image Processing problems, especially exploring multilingual and multimodal applications.For the past four years she has been working as a Research Software Engineer at IBM Research, India, prior to which she did her Masters in Computer Science from Indian Institute of Technology, Bombay. Previous to this, she had worked extensively on Wireless Sensor Network Security during her Bachelor Course in Information Technology and have had a few publications in that area.

Overall she has co-authored 10 conference and workshop papers in various eminent international conferences and 3 arxiv articles in the area of Machine Learning/natural language processing and Information retrieval. Other than this, from her past work on Wireless Sensor Networks, she has 5 conference papers and 4 journal articles in various internationally acclaimed conferences and journals. Her overall citation count, combined over the two research domains is 80. Other than this, she also has numerous patents in various applications of text/image analytics.

Work Experience [2012 - Current]

Amrita has been working at IBM Research Lab, India (Bangalore) as a research software engineer for the last four years in the Cognitive Text and Services team, which works on a variety of text, and image analytics based solutions catering to a large domain of applications. Apart from her individual research interest in various text/analytics problems, she is part of the core team at IBM, which is developing a Computational Argumentation Framework where machines can argue and debate with humans over any open-ended topic of controversy.



Research Projects at IBM

IBM Debating Technologies: A Grand Challenge [2013 - 2016]

The goal of this Grand Challenge project is to develop technologies that can assist humans to reason, make decisions, or persuade others. When it comes to decisions, opinions or points of view, there are no "right" or "wrong" answers. Successful arguments are based on evidence, constructed with perspective, and delivered persuasively. This multi-year cognitive computing project explores scenarios where there isn't only one answer and demonstrates the ability to generate pro and con arguments (consisting of claim and evidence statements) and speeches for controversial topics and will show other elements over time.

Role in the Project: As a leading person on the module of Pro-Con analysis from the India Research Lab, who has been working since the inception of the grand challenge, owned the following modules:

  • Topic-Based Stance (Pro/Con) Classification of Claim statements w.r.t. the debate motion statement
  • Topic-Based Stance Classification of long Evidence passages w.r.t the debate motion
  • Semantic Relationship (e.g. Consistent or Contrastive) identification between open-domain topics
  • Unsupervised learning of a graph of semantic relations (consistent or contrastive) over Wikipedia Headers (which involves learning of topics in the Wikipedia headers and the semantic relation between the learnt topics)



IBM Visual Linguist: A Picture is worth a Thousand Words [2014 - 2015]

The goal of this project is to understand open domain images, which involves understanding entities and actions and backgrounds in the image and generating a natural language caption crisply describing the salient factors in the image.

Role in the Project:

  • Language Model and Corpus-Co-occurence based action/attribute prediction in image
  • Label Aggregation Framework: Worked on a probabilistic graphical model based inference framework for a taxonomy-grounded aggregation of scores from multiple pre-trained classifiers.
  • Implementation of various building blocks for the Image-Understanding system
  • One of the key contributors of a Visual Search application for e-commerce (especially fashion) using the above image-understanding system (which was shortlisted in the top-9 out of over 60 submissions for the IBM Cognitive Hackathon, 2015)

IBM Cognitive Fashion [2015 - 2016]

Fashion is a multi-billion dollar industry with social and economic implications worldwide. With the advent of modern cognitive computing technologies (data mining and knowledge discovery, machine learning, deep learning, computer vision, natural language understanding etc.) and vast amounts of (structured and unstructured) fashion data the impact on fashion industry could be transformational. With this motivation this project is aimed at a mix of various applications including multimodal question-answering/dialogue-systems/recommender system/representation-learning/cross-domain retrieval.

Role in the Project:

  • Modeling Multimodal dialogue systems that are enriched by structured sources like knowledge bases and catalogue data and unstructured sources like free-form description of products
  • Building interactive demos in various applications like multi-modal dialogue systems, cross-modal retrieval and visual search for different clients in the fashion/jewellery/e-commerce domain
  • Co-organized a workshop on "Machine Learning Meets Fashion" at the international conference of Knowledge Discovery and Data Mining 2016


Research Projects at Indian Institute of Technology, Bombay

Optimal Feature Induction from an Exponential-sized Feature Space using Hierarchical Kernels (M.Tech Project, under Prof. G. Ramakrishnan, May 2011 – Jan 2012)

  • StructRelHKL:Rule Ensemble Learning using Hierarchical Kernels in Multiclass Structured Output Spaces for Activity Recognition :  Generalizing Rule Ensemble learning using Hierarchical Kernels to Multiclass Structured Output Spaces and Extending StructRelHKL to the case of complex First Order Activity Recognition Settings
  • DisjunctiveRelHKL:Learning Optimal Optimal Disjunctive Projections for Dimensionality Reduction in Max-margin Classification – Achieving a non-parametric Dimension Reduction integrated with Classifier training to build an optimum model 

Inductive Logic Programming for learning interpretable features (M.Tech Project, under Prof. G. Ramakrishnan and Prof. Ashwin Srinivasan, Setp 2011 - May 2012)

  • Studying which class of First-Order Relational Features are sufficient for building good models in Statistical LearningEmpirical results suggest that the entire set of first-order features from definite clauses are not necessary for statistical learning and a more constrained class of features actually suffice.
  • Optimal Resource-Bounded Search for Inductive Logic ProgrammingStudying the applicability of the “theory of optimal search” in directing a resource-bounded heuristic search for first-order logic clauses, conducted by Inductive Logic Programming Systems. 



Positions of Organizational Responsibility at IBM Research

  • ●  Co-organized a first-of-a-kind workshop on Machine Learning meets Fashion, at Knowledge Discovery and Data Mining conference, 2016

  • ●  Served as PC-member and in organizing and reviewing committee for workshops in several internationally acclaimed conferences like KDD and VLDB.

  • ●  Organized the regular department meetings at IBM Research Lab for the Cognitive Technology and Services Team, for over two years

  • ●  Handled miscellaneous responsibilities like hiring researchers and research software engineers and mentoring under-graduate students from various universities interning at IBM Research Lab over the last few years and training annotators to collect relevant data for various text/image analytics applications