*Click on the title for details.

ContentNCF: Content Based Neural Collaborative Filtering – University of Waterloo

In this work, we extend Neural Collaborative Filtering (NCF), to content-based recommendation scenarios and present a CNN based collaborative filtering approach tailored to image recommendation. We build upon the Pinterest ICCV dataset used in so as to include image features, and use it to make content-based image recommendations. This content-based approach, Content-NCF, proves successful in predicting user item interactions on our new Pinterest Image 2019 dataset.

Emo-CNN for Perceiving Stress from Audio Signals: A Brain Chemistry Approach – CentraleSupelec, France

The emotion plays a key role in many applications like in healthcare domain to gather emotional behaviors. It is very crucial to determine “how it was said” other than “what it was said”. Therefore, in human- machine interaction applications, it is important that emotional states in human speech are fully perceived by a Machine Learning model. Our ultimate goal at FAST lab was to model the stress from emotions. The abstract of this work is accepted in the International Conference on Speech Emotion Recognition and Affective Computing (ICSERAC) 2019.

Automated Essay Scoring with Cross Feature Vector Generation – IIIT-B

One of the difficulties of grading essays lies in the varying subjectivity of the grading process. This also leads to the contradicting variation in grades awarded by different human assessors, which is clearly unfair to the students. This problem could be solved by the adoption of automated assessment tools for essays (text answers). We designed the model, Intelligent Text Rater (ITR) to automate the task of essay grading. We also implemented ITR with the proposed novel approach of feature vector generation of text essays and significantly reduced the required training data size.

Spatio-Temporal Features of Crowd Models – Murdoch University, Australia

Convolutional Neural Networks (CNN’s) have shown great performance in learning appearance representations from images. But learning of temporal features (Between multiple frames) and how they can be effectively used together with appearance features for video analysis, is a challenging problem. In the past few years, number of works on learning temporal features of crowd videos using CNN’s have been reported and during my internship we surveyed the state of the art techniques on this topic. For a brief time, I also worked under the assistance of Yasir Jan who extended the conditional GAN to crowd dataset like SHOCK and WIDER FACE to give the bounding boxes around the human entities in the general crowd images.

Detecting User Activities Using Cell Phone Accelerometers – IIIT-B

This project was part of the hackathon held at IIIT-B. The purpose of this project was to process data and classify behavior into four wide activity classes like running, climbing, cycling, sedentary and other using the predictive model. We used the FFT spectrograms and K-Means algorithm to calculate step counts and cluster different movements made respectively. We were able to produce the clear and well separated clusters for activities using K-means Algorithm in Matlab. No libraries and packages were used in the implementation.