Key Projects

• Performed the tasks of Forecasting of Coal Prices using Multivariate Time Series Data

• Performed extensive Exploratory Data Analysis for time-series and trained ARIMA and LSTM models to forecast coal prices with multivariate attributes

• R² score of 0.96 was achieved 

• Modelling human motion is a challenging task since it is highly non-linear, dynamic, and stochastic in nature

• Given sequences of prerecorded human motion data (a subset of Human3.6M Dataset), the task was to predict 24 frames in the future based on 120 previous frames

• We built upon the RNN Structured Prediction Layer (SPL) model which decomposes the pose into individual joints and can be augmented with different neural network architectures

• Introduced the SPL dropout layer and performed ablation study in terms of sevral parameters

• Using the per joint loss instead of the standard mean squared error as well as residual connection for modelling velocities help us stay on top of leaderboard in the Machine Perception course project at ETH Zurich

• Achieved the joint angle score of 2.07

• The goal of the project was to design a precise and scalable automated verifier for proving the robustness of fully connected and convolutional neural networks with rectified linear activations (ReLU) against adversarial attacks. 

• Considered adversarial attacks based on the L∞ -norm perturbations. A network is vulnerable to this type of attack if there exists a misclassified image inside an L∞-norm based ε-ball around the original image.

• Leveraged the DeepZ relaxation to build the verifier and improved the ReLU transformer in DeepZ. 

• Learned a value λ for each neuron in the network which maximizes the precision of the verifier. 

• Built a deep learning model to detect acute intracranial hemorrhage and its subtypes 

• Designed a model that utilizes multi-window (brain, blood, bone) 3D context from neighboring slices to improve predictions at each slice 

• Leveraged pretrained Inception-ResNet and BiLSTM sequence network to learn 3D context. Modeled the network as a two stage process where the embeddings of each slice are fed into the sequence model to model dependence

• Data Augmentation is used to improve the performance of the model: rotation, flipping, scaling, transposition 

• Our proposed architecture performs significantly better than standard single window based non-contextual models with 0.942 F1-score

• Trained the model for classification of Age, Gender and Ethnicity from Images

• Explored the augmentation of ResNet and CNN to further improve the model

• Extensive hyperparameter space was searched using TALOS

• Accuracy achieved was 88% 

• Built deep learning models to tackle online abuse in terms of toxic comments on social media by classifying comments into different labels

• Created a CNN-BiGRU model with GloVe embeddings and leveraged the Pre-trained BERT embeddings for this task

• We report the standard F1 scores and AUC scores for the models and produce relative increase of 4% accuracy to one of the best researches in toxic comment classification

• Accuracy as high as 98% was achieved with the BERT model and our models improve upon the previous results  

•Performed theoretical study and prototyping of blockchain platform for EHR sharing of patients to promote faster tele-medicine consultations and foster research and development in medical science

• Patients benefit from secure EHR sharing and better consultation while organizations benefit from ease-of-access and privacy issues 

• Patients get the option to share a part of the EHR along with time-limited access to practitioner or institute

• Created a tool that tries to make the process of consenting to cookies more efficient while ensuring that the legally required information is conveyed

• Designed a text summarizer using both extractive (text-rank) and abstractive methods (Seq2Seq) for a given cookie policy

• Analyzed the output of the summarizer in terms of GDPR and Big Data Law

• Automated cookie acceptance tool based on keyword classification using XGBoost and CNN with GloVe embeddings 

Familinet is a platform that allows sharing of multimedia content and text posts within a network of family and/or friends. The app allows users to post content using the in-app post feature, which displays in a scrollable feed for the user. Additionally, users can share a post from other social networks (Facebook, Instagram etc.) These networks are private and invite-only. Familinet address the problems of people above the age of 55 having troubled navigation because of improper icon signifiers, cluttered feed, sending messages to wrong group of people etc. We specifically uncovered the issues related to navigation of an iPad with a lot on the screen at one time, the tendency to choose the easy solution instead of the best solution when carrying out a task, and the problem of sometimes sending content to the wrong group.

• Developed a deep learning based air quality forecasting method to leverage sequential nature, trend and seasonality in time series of AQI 

• Discovered attribution of pollutants concentration such as PM2.5, PM10, CO and NO2 to weather, traffic conditions, festivals and season • Initial benchmarking using random forest regression by artificially injecting hand-crafted time features and compare with LSTM later

• We also leveraged Residual Convolution Network and LSTM to incorporate spatio-temporal behavior of the various pollutants 

• The RMSE achieved was as low as 1.1 for the next 24 hours forecasting beating other state-of-the-art methods 

Created a low-cost product which automates the process of recording a lecture in a classroom. The Raspberry Pi 3 was used to run the Real Time Object Detection Algorithm implemented using OpenCV. Background Modelling using Mixture of Gaussians and Camshift along with histogram back-projection Algorithm was implemented for subject detection and tracking. 

Designed and implemented a noninvasive continuous blood pulse wave monitoring system that consists of transducer, data acquisition, and post-processing. The signal obtained from a test subject was passed through a signal conditioning circuit, digitized and transferred to a computer. Signal post-processing was performed to extract clinical information about the test subject using appropriate algorithms in software. 

Designed a Neural Network controlled self-driving car, which runs on a white track with black obstacles. The system captures video frames every couple of seconds and passes them to a series of neural networks, which have been trained by watching a human drive in similar conditions. The trained neural network can then be passed live video frames and will predict how to steer to stay on the road ahead.