PROJECTS

Global NIPS Paper Implementation Challenge:

  • Implemented “Selective Classification For Deep Neural Networks” paper as the part of challenge. The authors proposed a new method to build a classifier on top of a model(CNN) where the classifier selects or rejects a particular instance. If the classifier predicts that the model can predict the instance with high confidence then it would accept the instance and passes it to model(CNN) otherwise it would reject the instance and tells that a human intervention is needed. The code for the implementation can be found here.

Number Plate Detection:

  • Detection of number plates on the Indian car vehicles is a difficult task due to low resolution CCTV cameras used and position of cameras fixed at different angles, Used YOLO (version2) to detect number plates in this scenario and used 1633 images to re-train YOLO (version2) for this task. The project videos can be found here.

Emotion Detection In User Dialogue Conversation - SemEval-2019:

  • Emotion detection in Dialogues is one of SemEval-2019 task. The goal of the project is to determine the emotion in the conversation between two users ( three dialogue turns ).
  • Proposed hierarchical attention network as the emotion in dialogue depends on each turn of the conversation. Achieved F1 score of 0.68.

Answer Selection In Question Answering System - Microsoft India AI Challenge:

  • This goal of the project is to select most probable answer among 10 candidate answers for a given query. Experimented with Siamese - BiLSTM network, Attentive pooling, CNN based siamese network, BERT and ELMO embeddings. Achieved an MRR score of 0.58.

Sales Prediction:

  • Prediction of Daily sales and cumulative monthly sales of a pharmaceutical client, Used Xgboost algorithm (boosting technique) with extracted features. Built R shiny application for visualization of daily and monthly sales.

Email Classification:

  • Automatically redirecting the customer complaint emails of a leading bank to the respective departments at three hierarchical levels. Developed a hierarchical algorithm, used text mining techniques and Naive Bayes algorithm for classifying the emails into three hierarchical levels.