Research Internships

Injecting Common Sense Knowledge into Unsupervised Multilingual Pretraining Models for MultiChoice Classification

Guide: Prof. Dr. Goran Glavaš, Data and Web Science (DWS) Group, Universität Mannheim, Germany
Scholarship: DAAD WISE Scholarship 2020

  • Trained adapter based multilingual BERT and XLM-R on ConceptNet relations and thus injected knowledge in transformers from various languages.
  • Used the transformer injected with common sense knowledge to train on SIQA multi-choice dataset which is in English and evaluated on the target language for which knowledge was injected from ConceptNet from XCOPA multi-choice dataset.
  • Showed that intermediate ConceptNet training improves the zero-shot transfer from English to the target language.

End to End Binarized Neural Networks for Text Classification

Researcher at NeuralSpace, London, United Kingdom

Accepted at SustaiNLP 2020 Workshop on Simple and Efficient Natural Language Processing @ EMNLP 2020

  • Trained end to end binarized networks on 4 intent classification datasets - AskUbuntu, Chatbot, WebApplication and 20NewsGroups.
  • Made use of binarized hyper-dimensional vectors for embeddings and showed that they were memory effcient and also achieved state of art results when compared to traditional embeddings (GloVe and Word2vec).
  • Showed that binarized networks were time efficient and reduced memory footprint compared to non binarized counterparts.

Incremental Training for Image Classification of Unseen Objects

Guide: Prof. Soumitra Nandy, Indian Institute of Science (IISc), Bangalore, India
Scholarship: Indian Academy of Sciences Fellowship 2019

  • Implemented YOLO object detection algorithm in Tensorflow and made modifications in training dataset to address fundamental questions related to YOLO architecture.
  • Worked on image classification of CIFAR-10 dataset and experimented whether incremental training is possible.
  • Trained the VGG-16 Net for image classification of 8 out of 10 classes and then incrementally trained the 2 classes on the last few convolutional and fully connected layers achieving decent accuracy and then replicated the same to YOLO.
  • The objective was to verify incremental training (detection of unseen objects is still yet to be verified) so that it could later be implemented on a video surveillance camera.