Biography
I am an AI Researcher working on conversational AI and OpenGPT-X at Fraunhofer IAIS in Germany. My major research interest centers around investigating and comprehending the reasoning and planning aptitudes displayed by the new generation of foundational language models. Additionally, I explore diverse novel applications and evaluation systems leveraging large language models.
I successfully obtained my Master’s degree from the Department of Computer Science, Saarland University in Germany. Before that, I served as a research student at the Max Planck Institute for Informatics, focusing on ad-hoc neural information retrieval systems and natural language processing. Additionally, I had the opportunity to work as a visiting student researcher at the Laboratory for Computational Social Systems, IIIT Delhi.
I completed my Master’s thesis titled “Retrieval augmented generative task-oriented dialogue systems for faithful knowledge grounding” with the Chair of Computer Science and Computational Linguistics and Fraunhofer IAIS. I was fortunate to be advised by Prof. Dr. Vera Demberg and Prof. Dr. Jens Lehmann. At my previous research position (Immersion Lab), I worked with the Database and Information Systems Group where my work was supervised by Dr. Erisa Terroli. During my Masters, I have also worked as a research assistant at German Research Center for AI, Saarbrücken, and at Fraunhofer IZFP, Saarbrücken.
My interests are in building robust machine learning models for scalable and reliable AI systems, spanning the areas of natural language processing, information retrieval, knowledge bases and knowledge graphs-based reasoning, and graph machine learning. Currently, I am involved in supporting the training of foundational LLMs like OpenGPTx and TrustLLM, with a particular focus on investigating their trustworthiness and exploring dimensions of reasoning, planning, and calibration/alignment.
I am currently looking for potential fully-funded PhD positions to work on challenging and interesting problems in my topic of interest. I am also open to R&D full-time and research residency positions where I can utilize my creative problem-solving and programming skills.
- Machine Learning
- Natural Language Processing and Generation
- Information Retrieval (Retrieval-augmented Generation, RAG)
- Science of foundational language models
- Graph Machine Learning and Complex Networks
- Fair and trustworthy AI Systems
MSc, Department of Computer Science, 2022
Saarland University, Saarbrücken, Germany
BTech, Electronics and Telecommunication Engineering, 2016
IIIT Bhubaneswar, India
Skills
5+ years experience in coding with Python and ML libraries
Experience through relevant coursework, projects and research work
Experience through industry projects, related course and research work
Experience through industry projects, related course and research work
Experience in writing big data algorithms and visualizations
Experience in working on industrial and academic projects
Education
Graduate student in the department of computer science with a strong focus on machine learning and information systems.
- Grade: 1.7 / 1.0 (German Scale)
- Thesis : Retrieval augmented generative task oriented dialogue systems for faithful knowledge grounding (at the Chair of Computer Science and Computational Linguistics) - supervised by Prof. Vera Demberg and Md. Rashad Rony.
Undergraduate student in the department of electronics engineering with a focus towards signal processing.
- Grade: 8.4 / 10
- Thesis: Performance improvement of MIMO based Free Space Optical links: Diversity and Variable Aperture techniques - supervised by Dr. Bijayananda Patnaik, IIIT Bhubaneswar
Experience
Studying social networks for exploring structural and behavioural properties using graph mining and NLP.
- Supervisor: Dr. Tanmoy Chakraborty
Research student working in the topics of neural information retrieval systems, with a focus on health domain.
- As part of my research work, I worked towards the objective of enhancing clinical information retrieval in health forums, and therefore improving patient-centric community QA and browsing experience.
- Wrote a set of optimal heuristic functions that maximizes the relevancy scores for a labelled dataset by training a snorkel classifier that classifies a given query-document pair as relevant or irrelevant. Later, these functions will be extended to classify the unlabelled set of query-document pairs, followed by re-ranking
- Additionally I explored neural ranking models, and conducted several experiments for evaluating modern deep IR techniques based on : a. Type of deep matching models - Representation based vs Interatcion based b. Type of matches : Semantic matching vs Relevancy matching c. models that perform reranking in multi-stage ranking architectures vs learned dense representations that attempt to perform ranking directly.
- Supervisors: Dr. Erisa Terolli, Dr. Patrick Ernst
Worked at the department of Algorithm, Signal, and Data Processing.
- Implemented LSTM Autoencoder for automatic audio defect detection for rotating machineries.
- Research and development of algorithms and performing theoretical analysis in sparse/manifold models using machine learning techniques with applications to 3D image segmentation..
- Supervisor: Prof. Osman Ahmed
Worked as a part of the project on MMPE: A Multi-Modal Interface for Post-Editing Machine Translation.
- Implementation and testing of various automatic machine translation evaluation metrics.
- Assited on developing the web-based application for post-editing experiments to assist in user studies.
- Supervisor: Dr. Satanu Pal
Worked as part of the Data Science team, with a focus on Information Systems.
- Developed a toolkit (Match and Merge) for automating Enterprise Asset Management processes. Match and Merge is a user-initiated intelligent asset master record recommendation tool that recommends potential files to be merged into a master record based on the user’s query or specifications. The tool supports an exploratory local and global analysis of the master record. The toolkit is developed using Klein web services and uses Elasticsearch from the ELK stack.
- Mentored interns on developing a Master Data extraction tool. Together we developed a custom named entity recognition system for enterprise master data assets using classical feature engineering and distant supervision based inductive rule mining.
Worked as part of the DataOps Analytics team and Cloud Infrastructure R&D team.
- Worked on association and temporal data mining problems to mine and associate frequent error patterns in log data leading to infrastructure based incidents.
- Worked on NLP based text analytics for development of NIA, AI Platform of Infosys. Played a major role in developing Exploratory Data Analytics and Predictive Analytics web-applications using RShiny.
- Worked on a R&D project by Infosys Labs and Infosys Security Group: Mining and classification of Malicious, spam and phishing texts using machine learning and deep learning approaches.
- Supervisor: Mercy Peter
Recent Publications
Projects
Case-study of Offensive Language Classification from Online Tweets
Here we try to analyze different approaches of online tweet classification into offensive and non-offensive categories in the presence of emoticons. We try to enhance the tokenization process to capture better sparsity in the presence of non-textual corpus.
Clustering based machine learning application using Flask
This is an unsupervised learning based lustering web tool that has been used to cluster microstructures in Steel based on their physical properties that have been extracted using classical image processing algorithms. Files can then be uploaded in .xlsx or .csv format. The tool supports five different clustering algorithms in combination with dimensionality reduction.
Data Augmentation using Feature Generation for Volumetric Medical Images
In this work, we propose using U-net and ACGAN as a learning framework for feature generation of medical images followed by classification to validate the quality of generated features.
Enhancing clinical information retrieval on health forums using neural ranking models
Since health forums become a rich source of information to people with medical conditions discussing treatments, doctor’s opinions, side-effects to complex-drugs, while also sharing personal background medical information in a community question-answering framework, we develop a neural search engine on top of such health forums by exploring the state-of-the-art neural ranking models. We first write a set of optimal heuristic functions that maximizes the relevancy scores for a labelled dataset by training a snorkel classifier that classifies a given query-document pair as relevant or irrelevant. Later, these functions are extended to classify the unlabelled set of query-document pairs, followed by re-ranking using neural re-rankers.
Music Information Retrieval For Genre Classification
We implement k-nearest neighbors, Gaussian Mixture Model, Multi-class SVM, Convolutional Neural Network, and Convolutional Recurrent Neural Network to classify the following four genres- Dark-Forest, Hi-Tech, Full-On, and Goa. We further extract 30 temporal features using a Long Short Term Memory based Auto encoder from individual frames, and augment them with the frame-level audio features, which is a novel contribution in this work.
Score Me if You Can
Study on Robustness of Automated Essay Scoring Systems to Out-of-domain and Adversarial Inputs summary
Recent and Upcoming Activities
Attended the Max Planck Intersectional Symposium on Computing and Society (January 2021).
Attended the 2020 Neural Information Processing Systems (NeurIPS ) (December 2020).
Miscellaneous
- Outdoors: I hike. I play football. I swim. I photograph. And I imagine.
- Indoors: I bake. I cook. I hack electronics. I play piano. I play Go. I read. And I imagine.
- I am an Otaku. And I imagine (そして私は想像します)
Phonologically, my name is pronounced as ‘Soumy’ Ranjan Sahoo | ‘सौम्य’ रंजन साहू | ‘ସୌମ୍ୟ’ ରଂଜନ ସାହୁ
Contact
- mail[dot]soumya10[at]gmail[dot]com
- Saarland University, Saarbrücken, Saarland 66123
- Skype Me