EXPERIENCE

Worked on leveraging GAN-induced transformations for self-supervised training in GANs (Generative Adversarial Networks) for improved image generation quality and diversity and better latent space semantics.

The project involved working on state-of-the-art GAN architectures like BigGAN, StyleGAN, and StyleGAN2, on popular image datasets like CIFAR-10, CelebA-HQ, FFHQ and ImageNet-2012.

Succesfully improved latent space semantics of BigGAN (as measured using this framework) and StyleGAN (as measured using this framework), while also improving image generation quality and diversity (as measured using FID metric).

Work accepted at WACV 2021 (Winter Conference on Applications of Computer Vision); Also filed a patent for the same.

Technologies Used: PyTorch, TensorFlow and OpenCV.

Interned as a part of BITS Pilani's Practice School II programme.

Project Title: Multi-modal Bixby Interaction (Digital Appliance).

Worked on the following two tasks for a proof-of-concept:

• Food Image Classification:

  • Collected an image dataset of food items - consisting of around 50,000 training and 8,000 validation images belonging to a total of 120 high-level categories (like cake, etc.) and 694 low-level categories (like milk cake, coconut cake, etc.). It was prepared by scraping the allrecipes.com website.

  • Used Keras to fine-tune a pre-trained MobileNet-V2 network to be used as a backbone network for a two-level local classifier CNN to predict the food-item given the image.

• Question-Answer System based on Recipe Text:

  • The collected dataset also contained a textutal description of each recipe/food-item - list of ingredients, cooking steps, nutrition facts, etc.

  • Developed a minimal heuristic-based QA system which given a recipe/food-item name and a question (regarding the recipe), outputs the answer to that question based on the recipe's textual information present in the dataset.

• Technologies Used: BeautifulSoup, Selenium, Aiohttp, Keras, and NLTK.

• Sample Use Case: Smart refrigerator integrating the above two functionalities and enabling QA on image of food (kept in your refrigerator) with support from Bixby, Samsung's virtual assistant.

Worked on increasing the efficiency of DOER (Decentralized Distributable Disk of Offline Open Educational Resources) - a product of the Gnowledge Lab.

• DOER is packaged as a cluster of servers that can be installed on any PC by copying the DOER distribution image of a 1 TB hard disk. When we boot from this specially crafted hard disk, the PC boots with several servers like NROER, Khan Academy Lite, etc.

• DOER targets rural and underprivileged areas (without internet connectivity) and provides them with easy access to free educational resources. It can be accessed from the PC when it is connected in a LAN of a school or a college, without requiring internet connection.

Implemented a smart caching service in the backend, reducing the response time by more than 50% and thus ensuring better user experience on low end devices (as are common in target areas).

Technologies Used: Python, Django, Selenium and Beautiful Soup.

Interned as a part of BITS Pilani's Practice School I programme.

Led a group of nine teaching assistants to conduct two workshops (to teach Python, NumPy and PyTorch) and two graded assignments (related to implementation of fully-connected neural networks, genetic algorithm, and PID controller from scratch in NumPy) as a part of the course.

Also, personally mentored five groups (consisting of three students each) for the course project - implementing a research paper of their choice (related to Computer Vision and published in top-tier conferences) in TensorFlow, PyTorch or Keras.

Course Instructors: Dr. Surekha Bhanot, Dr. Bijoy Krishna Mukherjee

Software-Defined Networking (SDN) is an emerging network architecture approach that enables the network to be intelligently and centrally controlled, or ‘programmed,’ using software applications, thus helping operators manage the entire network consistently and holistically, regardless of the underlying network technology.

Hybrid SDNs contain both legacy and programmable network switches and allow operators to reap the benefits of SDN without upgrading the entire network.

Waypoint enforcement refers to techniques used in hybrid SDNs to ensure that all data packets flow through at least one SDN-compatible switch (which in turn is controlled by the SDN controller).

Currently, there are many waypoint techniques available, with five of them being Virtual-IP based approach, Multi-VLAN based approach (both proposed by SDN Lab, BITS Pilani), Telekinesis, Magneto, and Panopticon.

This project aims to simulate these five techniques on a Cisco three-tier topology using LBNL traffic and compare them based on certain parameters like average path length, percentage of traffic waypoint enforced, average load on links, link utilisation and critical load capacity.

This work was accepted at WAINA 2020 (The Workshops of the 34^th International Conference on Advanced Information Networking and Applications) [Link].

Mentors: Dr. K Haribabu, Ms. Sandhya Rathee

The Student Faculty Council is a body of the CSIS Department whose members are faculty members and a few selected students. It is chaired by the Head of CSIS Department.

Two meetings are organized per semester, in which we discuss issues regarding quality of class-room instruction, laboratory facilities and equipment, examinations, classroom attendance, possible electives to be offered in subsequent semesters, etc.