Anugunj Naman

I am currently pursuing a graduate degree at Purdue University, specializing in Machine Learning. Previously, I worked as a Software Engineer at Elevance Health, where I collaborated with the Responsible AI and Explainable AI (XAI) team to enhance fairness, interpretability, and mitigate data drift in machine learning models.

I hold a Bachelor of Technology in Computer Science and Engineering from the Indian Institute of Information Technology, Guwahati, which I completed in 2022.

My research experience includes an internship at NVIDIA, where I focused on speech recognition, meta-learning, and few-shot learning. I also interned at the AI-ML-NLP Lab at IIT Patna, where I worked on multimodal learning techniques.

In addition, I have made contributions to the open-source community, particularly to the HuggingFace Transformers library. My contributions include integrating two vision transformer models: Microsoft's CvT: Convolutions to Vision Transformers and Meta's LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference.

Contact: anugunjjha at gmail dot com

Recent Publications

2024

NeurIPS

Random Propagations in GNNs

Thu Bui, Anugunj Naman, Carola-Bibiane Schönlieb, Bruno Ribeiro, Beatrice Bevilacqua, Moshe Eliasof

2nd NeurIPS Workshop on Unifying Representations in Neural Models (UniReps), 2024

Abs PDF

Graph learning benefits many fields. However, Graph Neural Networks (GNNs) often struggle with scalability, especially on large graphs. At the same time, many tasks seem to be simple in terms of learning, e.g., simple diffusion yields favorable performance. In this paper, we present Random Propagation GNN (RAP-GNN), a framework that addresses two main research questions: (i) can random propagations in GNNs be as effective as end-to-end optimized GNNs? and (ii) can they reduce the computational burden required by traditional GNNs? Our empirical findings indicate that RAP-GNN reduces training time by a factor of two, while maintaining strong accuracy for node and graph classification tasks.

2021

ICNLSP

Indic Languages Automatic Speech Recognition using Meta-Learning Approach

Anugunj Naman, Kumari Deepshikha

4th International Conference on Natural Language and Speech Processing (ICNLSP), 2021

Abs PDF

Recently Conformer-based models have shown promising leads to Automatic Speech Recognition (ASR), outperforming transformer-based networks while meta-learning has been extremely useful in modeling deep learning networks with a scarcity of abundant data. In this work, we use Conformers to model both global and local dependencies of an audio sequence in a very parameter-efficient way and meta-learn the initialization parameters from several languages during training to attain fast adaptation on the unseen target languages, using model-agnostic meta-learning algorithm (MAML). We analyse and evaluate the proposed approach for seven different Indic languages. Preliminary results showed that the proposed method, MAML-ASR, comes significantly closer to state-of-the-art monolingual Automatic Speech Recognition for all seven different Indic languages in terms of character error rate.
TCSS

A Multimodal Author Profiling System for Tweets

Chanchal Suman, Anugunj Naman, Sriparna Saha, Pushpak Bhattacharyya

IEEE Transactions on Computational Social Systems, 2021

Abs PDF

The rising usage of social media has motivated to invent different methodologies of anonymous writing, which leads to an increase in malicious and suspicious activities. This anonymity has created difficulty in finding the suspect. Author profiling deals with the characterization of an author through some key attributes such as gender, age, language, dialect region variety, personality, and so on. Identifying the gender of the author of a suspect document is a salient task of author-profiling. The linguistic profile of a user can help in determining his/her demographics. Different social media platforms, such as Twitter, Facebook, and Instagram, are used regularly by users for sharing their daily life activities. Moreover, users often post images along with text on different social media platforms; thus, the usage of multimodal information is very common nowadays. In this article, the task of automatic gender prediction from multimodal Twitter data is posed as a classification problem and an efficient multimodal neural framework is proposed for solving this. The popularly used BERT_base is utilized for learning the encoded representation for the text part of the tweet, and recently introduced EfficientNet is used for extracting the features from images. Finally, a direct product-based fusion strategy is applied for fusing the text and image representations, followed by a fully connected layer for predicting the gender of a Twitter user. Plagiarism detection authorship analysis near end duplicate detection (PAN)-2018 author profiling data are used for evaluating the performance of our proposed approach. Our proposed model achieved accuracies of 82.05%, 86.22%, and 89.53% for pure-image, pure-text, and multimodal setting, respectively; outperforming the previous state-of-the-art works in all the cases. Moreover, a deep analysis is carried out to interpret the produced results; different words that serve as clues for gender classification are identified characterizing different gender classes.