Parag Dutta

Causal Feature Alignment: Learning to ignore suprious background features

Authors: Rahul V., Parag Dutta, Vikram M. and Ambedkar Dukkipati

Deep Representation Learning for Predicting Temporal Event sets in the continuous time domain

Authors: Parag Dutta, Kawin M., Pratyaksha S., and Ambedkar Dukkipati

arXiv Link -->

CRUSH: Contextually Regularized and User anchored Self-supervised Hate speech Detection

Authors: Souvic Chakraborty, Parag Dutta, Sumegh Roychowdhury and Animesh Mukherjee

ACL Anthology Link · arXiv Link · Presentation Video Link

Active² Learning: Actively reducing redundancies in Active Learning methods for Sequence Tagging and Machine Translation

Authors: Rishi Hazra, Parag Dutta, Shubham Gupta, Md. Abdul Qaathir and Ambedkar Dukkipati

ACL Anthology Link · arXiv Link · Presentation Video Link

Machine Reading Comprehension for ranking search engine query results by relevance - A review

Authors: Parag Dutta and Debayan Ganguly

Contextual detection of code-mixed textual hatespeech in Social Networks

We are working with graph neural networks and pretrained transformers for detecting code-mixed hate speech in social network posts. We are using pretrained transformers to exploit their language understanding aspects, we are using graphs to leverage the context from the tree structure among the posts/comments/replies, and training in an end-to-end manner.

CoviHawkes District wise daily Covid-19 case prediction

I was involved in temporal modelling of daily covid-19 cases for Indian states and districts. We used inter-district and intra-district mobility features in graphs to predict case counts with low MAPE values. We validated our model's effectiveness with extensive validation by comparing with various baselines like SIR, 2 layered MLP, LSTMs, GCN, GAT, etc.
I also helped with the website design and poster designs. This project was featured in the IISc website homepage and IISc EECS division website homepage.

Few-shot Transfer Learning in Medical Imaging from Sparse Annotations

I was involved in implementing two methods to predict medical image segmentation in particular spine segmentation, in the low data regime. Our aim was to train a model on data for which annotations are easily and readily available (lumbar spine images) such that it generalises well to similar but previously unseen regions (thoracic/cervical spine). First, we implemented a contour propagation model that uses priors obtained in a human-in-the-loop setting to generalise to regions of the spine not seen during training. We show that this reduces the effort involved in annotating new data. Second, we implement a transfer learning approach and pretrain a model on a large amount of partially annotated lumbar spine data (source domain) and fine-tune the network on a limited amount of data with complete whole spine annotations (target domain). We show that this method gives us a boost in performance with minimal complete volume supervision at the time of training and without any supervision required at inference. We achieved dice similiraty coefficients of upto 0.86 by using as low as just 20 volumes of CT scans.

Active Learning for Neural Machine Translation

I was involved in generating a deep learning model which can be used to identify and eliminate redundancies in existing Active Learning methods on a variety of tasks. While my co-authors focused on tagging based tasks, I focused on machine translation. I used Siamese networks pretrained with similarity loss to get embeddings of the sentences, followed by clustering and selecting only the most relevant sentences for training a NMT model. We showed that our model beats existing active learning methods by performing better after consuming same amount of data on all NLP tasks we experimented on.

Explainable classification based on mitochondria treatment from microscope images

I had to develop an end-to-end object classification model from the ground-up. Since this was a project in the biological sciences domain, the main focus was to develop a model that was not only accurate, but also had a very high explainablility. Additionally there was also a need for the model to be lightweight, becasue the target user circle are not supposed to have specialized hardware to run the model. I developed a novel convolutional archtitecture with just 4 million parameters, which not only achieved accuracies of 93% but also gave extremely explainable results with saliency maps. A non-explainable model based on Fast-RCNN based architecture achieved more than 96% accuracy.

Deepfake Detection in High Compression Videos

Deepfakes are images or videos which are forged by inserting the facial features of a person over an some other persons' face using some deep neural network like GANs or VAEs. It is hard to detect deepfakes as it is, even more so when the video is compressed. Most of the videos streaming online are compressed videos. Hence, I developed a simple CNN based architecture inspired by the discriminator from DCGANs to detect facial forgeries in highly compressed videos which detected 92% of the forged videos by only sampling 6 to 10 frames per video and less than 200ms per frame; thus outperforming all existing methods by over 9% in accuracy. Additional experiments pushed the accuracy to 95% by considering sequential relations in the video frames.

Polymorphic Malware Detection using Deep Learning

Using the Microsoft Malware Classification Dataset and files from my system, I created a CNN based binary malware/benign classifier after converting the files to their corresponding byte-wise grayscale images.
Then I also created a simple polymorphic malware in python, which was then used to train the classifier using transfer learning. It's just a proof of concept that polymorphic malwares can be detected using probabilistic deep learning model (because it can't be detected using deterministic approaches).

Prediction of colon cancer based on genome features

Comparative study on performance of various machine learning models on low data regime. A supervised dataset was provided which had around 2000 features but only 62 observations. The work was mostly driven towards dimensionality reduction. I expiremented with various dimention reduction algorithms like PCA, LDA, ICA, PLS, etc. On the reduced dataset I tried different machine learning algorithms like Random Forest Ensembles, Support Vector Machines, Bayesian Networks, K - Nearest Neighbors, etc. The best results was gained with PLS reduced dataset of 40 features along with Linear SVM.

Providing relevant answers to question asked in search engine queries using MRC

The problem was to provide to-the-point precise answers to questions like 'What is the speed of light?', to which the answer should be '299,792,458 m/s'. To do this, the query along with the top search results would be provided, and after sorting them according to our requirements, we needed to extract the relevant text and display it directly to the user if a certain confidence threshold was achieved. To do this I used a multi-hop attention networks for generating a ranking over relevant passages given a user query. It was a pipelined model, which used pretrained transformers, and significantly outperformed most of the models in the competition.

Social media sentiment anaylsis

A labelled dataset of 1.6 million tweets was provided for training a deep learning model which analyses sentiments in a given tweet on a scale of -1(negative) to 1(positive). I started with basic frequency based classification, followed by RNNs, LSTMs and GRUs. Later I also tried attention and gained significantly higher accuracy. It was basically a comparison based project to determine the capabilities of various recurrent neural network models.

Swarm Intelligence among drones for movement and reconnaissance

Proposed an intelligent system to crontrol the movement of a swarm of drones as if it were a single unit. The individual drones would learn how to collaboarate with others for information sharing and move stochastically without collision using particle swarm optimization techniques. Moreover the whole swarm would have a hive-mind like intelligence which could be used for group reinforement learing, and also share computational resources among themselves. The architecture was extremely scalable, and hence additional hardware could be used to implement features like FLIR, real time identity searches, signal jamming, and gait analysis, which are extremely useful for surveillance.

MERN Full-stack Web Development (Internship)

I was involved in the ground up development of a dynamic, responsive, and reactive progressive web application. The project was developed from scratch, using MERN stack. A RESTful web API service was developed in the backend, using MongoDB Cloud services for database, NodeJS as the server-side scripting language and a combination of ExpressJS & Restify as the framework for Authentication and REST API services respectively. As for the client-side frontend of the system, a Progressive Single Paged Application (P-SPA) was created using ReactJS to handle communication with the server.

Vehicle Numberplate Recognition at security checkpoint

I was involved in development of a system that could use live footage from the cameras positioned at security gates to automtically read the number plates of a vehicle that entered the premises and logged the same into a database. After detecting the region of interest using convloutional filters, I used basic SSD algorithm for OCR, and a modified version of busy wait locking protocol for synchroniztion.