Sourya Dipta Das

Sourya Dipta Das

Undergraduate Student of Dept. of Electronics & Telecommunication Engineering at Jadavpur University

My Research works

Code-Mixed Data Generation for Low Resource Language Pairs

Aug,19 – Feb,20

This research project is in collaboration with JU-NLP lab for generating code-mixed data from normal monolingual text for low resource language pairs( here, Bengali-English Pair). Generated code-mixed data can be used in chatbots and also as augmented data for various NLP tasks in code-mixed data domain.

Sentiment Analysis of Hindi-English Code-Mixed Data with Noisy Label

Oct,19 – Jan,20

This project is part of the feature extraction module from Dialogue Act Classification from Telephone Calls Project and also my submitted solution for Sentimix Challenge of SemEval 2020. Here, cascaded LSTM-GRU with Multi-Head Attention layers is used with Sub-word Embedding to make it more robust to the noisy data. Problem of Noisy label is also handled by the noise adaptation layer added at the end of the model.Paper of this project is currently under review.

Fine Grained Sentiment Analysis of Meme Image & Texts using Multi-Task Learning

Oct,19 – Feb,20

This project is submitted solution for Memotion Analysis Challenge of SemEval 2020. Here, ResNet, LSTM-GRU model with Attention are used to extract features from image and corresponding text respectively. Sentiment value and class of the memes are learned jointly by a single network. Paper of this project is currently under review.

Demoireing High Resolution Images Using Conditional GANs

Jul,19 – Sept,19

This project is my submitted solution for AIM 2019 Demoireing Challenge - Track 2: Perceptual of ICCV 2019. Here, i have used a Global Generator instead of local Generator to emphasize on global features more to increase fidelity & perceptual quality of the image. As Moire effect heavily dependent on scale of the image, i have made an image translation model in high resolution by using a Multiscale Discriminator instead of doing up scaling which is computationally heavy task.

Semantic Segmentation of Steel Surface Defects Using SE-ResNeXt-50 Unet

Aug,19 – Oct,19

Main goal of this project is fast detection and segmentation surface defects based on their defect type/class from steel sheets. SE-ResNeXt-50 Unet is used for semantic segmentation with Combined loss of Binary cross-entropy loss and Dice Loss.

Multiclass Noisy Text Classification using BERT

Jun,19 – Jul,19

This project is a part of’CIQ Challenge’organized by FIRE 2019 to identify Noisy Insincere Questions from Question Answer Community.Pipeline is consists of several text pre-processing techniques to reduce noise and Pre-trained BERT model for transfer learning.

Geometric data extraction from 3D point cloud of the LIDAR for a generative model

Oct,17 – Dec,18

This project is associated with a DRDO funded project for developing a Quadruped Robot for Military Applications in the Neptune Lab. We are developing a unsupervised algorithm for extracting geometric data from point clouds of a LIDAR,So that quadruped can do more optimized path planning. I had done this project under joint guidance of by Dr.Ananda Shankar Chowdhury of ETCE Dept. of Jadavpur University & Sanjoy Kumar Saha of CSE Dept. of Jadavpur University.

Indian Language Transliteration System using Seq2Seq Model

May,17 – Sept,17

Main goal of this system is to generate roman transliterrated represenatation of word from another Indian langauge(like Hindi,Bengali) based on user phonetic morhological trend. We have used LSTM based seq2seq model and validated our system using turing test.

Phonetic based Language Identification of Mixed Coded Data Using LSTM

March,17 – June,17

This RNN based model learns the phonetic relations from various words for individual languages. I designed a Phonetic based feature extractor for the words from the mixed coded data, which will be used for language classification. we have used our created dataset of Mixed Coded text (Bengali-English) twitter data. I have done this project under Dr. Dipankar Das of CSE Dept. of Jadavpur University.

Path planning of autonomous robots in crowded dynamic unknown environment using Artificial Immune System Optimization

Oct,16 – Sept,17

Most of obstacle avoidance algorithm like gravitation based search or charge based object avoidance algorithm failed in crowded dynamic unknown environment due to large number of static and moving objects as they have to consider so many objects and compute a lot of functions. In this scenario, our proposed algorithm gives a low computational approach with faster responses to the robot for obstacle avoidance. Our algorithm also worked on fast movement of the robot due to less secondary response time. Here, main object of each robot is to reach a goal position from an initial position. So, we design an online path planning algorithm which is capable of giving a fast response to a possible threat or moving object. This project is guided by Dr. Pratyusha Rakhshit of ETCE Dept. of Jadavpur University.

Energy Optimized Robot arm Path Planning using Differential Evolution Algorithm

Oct,16 – Jan,17

This project deals with a evolutionary approach to design a algorithm for energy efficient path planning of an industrial robot arm in a workspace with multiple obstacles using differential evolution (DE) algorithm. The path-planning problem is formulated as an optimization problem with an aim to determine the shortest and energy optimal path of the robot arm from its given initial position to the pre-defined goal location, without hitting obstacles. Application of such evolutionary algorithms in trajectory planning is advantageous because the exact solution to the path-planning problem is not always available beforehand and must be determined dynamically. This project is guided by Dr. Pratyusha Rakhshit of ETCE Dept. of Jadavpur University.

Collaborative Research work:-

Single view to Multiple view via Dimensionality Reduction: a Generalized framework

Sept,17-Dec,17

This project work is about Single view to multiple views via Dimensionality Reduction. Our main aim in this project is to develop a generalize algorithm to visualize data points from multiview datasets using single-view dimensionality reduction techniques.I had done this project with Sumit Mukherjee , PhD student of Seelig Lab of Dept. of Electrical Engineering at University of Washington. Here, my role was to make contribution to main algorithm and write the code in MATLAB/Python for Simulation.