Meta Learning without memorization with exact information estimation
Last Updated: 3rd Sept
MThis project is improvision over Meta-Learning without Memorization by Mingzhang Yin, George Tucker, Mingyuan Zhou, Sergey Levine, Chelsea Finn. In this paper, the authors approximate the mutual information for tractability of loss function. Alternatively, this projects directly maximizes the mutual information without approximation, using the MI estimator MINE inturn improving results.
Multi-Agent Generative Network (MAGNET)
Last Updated: 3rd Sept
MAGNET is a multi agent cooperative learning based architecture for generating new images. MAGNET generates images similar to images in CelebA dataset. The algorithm of multi-agent cooperative learning is inspired from Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments. This is a dual agent cooperative game than the conventional competitive MAP learning like in GANs. The results are better than most of the proposed GAN variants.
Environment Adaptive Self multi task learning Navigation Robot
Last Updated: 5th May
Modified version of the paper ’Composable Action-Conditioned Predictors: Flexible Off-Policy Learning for Robot Navigation’. Simulated RC and a real-world RC car that can gather data and train fully autonomously without any human provided labels beyond those needed to train the detectors, and then at test-time be able to accomplish a variety of multiple tasks(In process of building real-world RC).
Non-Convex approaches to RPCA and applications to Computer Vision problems
Last Updated: 26th April
The method we developed transcend the problem of low-rank and sparse matrix decomposition and applicable to solve inverse problems often encountered in several areas. We developed an efficient alternative representation corresponding to the mini max-concave penalty (MCP): re-weighed MCP (RMCP). Also developed re-weighted 𝛾 –penalty (RGP), which is the RMCP defined over the singular values of a matrix, as a non convex relaxation to the matrix rank. This work out performs current SOTA which employs a heuristic weight update, and is submitted to IEEE Transactions on PAMI and IEEE CVPR. Paper Supplementary material
ChildrEN SafEty Retrieval (CENSER) System for Retrieval of Kidnapped Children from Brothels in Varanasi, India
Last Updated: 26th April
CENSER System was developed by me in collaboration with Skylark labs and is being used by Guria non-profit organization in India to retrieve the kidnapped children from brothels in India. The children buying makeup are recorded on the hidden cameras which are used by the CENSER system to establish a match with the database of missing children. Since the facial features of children have changed significantly since being kidnapped, the proposed system can perform age-invariant face recognition. A majority of the time, the image of the child is not available at the age when he/she was kidnapped. In that case, the CENSER system uses kinship analysis as well as matches with the sketch of the child. This work is submitted to IEEE Transactions on Bio metrics, Behaviour, Identity Science. This system uses Memory Augmented ScatterNet ResNet Hybrid Network inspired from the complementary learning system theory. CENSER has been used in 14 raid operations by Guria (http://www.guriaindia.org/) NGO in Varanasi, India and has so far saved 1 girl child.
Mask Detection and Offender Recognition(MADOR) System
Last Updated: 26th April
(MADOR) System was developed by me in collaboration with Skylark labs. Masks have been made mandatory by governments in several countries amid the COVID-19 pandemic. Our MADOR system can detect individuals without mask, using deep learning, and alerts (system or phone, messages) the authorities. The faces of the offenders can also be added to the database and recognised later by matching against a database.
An Intelligent Communication System based on 𝛽-VAE
Last Updated: 26th April
Designed a new communication system based on Variational Auto Encoders. The designed system can operate under Additive Wide Gaussian Noise (AWGN) and Rayleigh fading channels. The CNN based VAE architecture performs the encoding and modulation at the transmitter, whereas decoding and demodulation at the receiver.This work is submitted to Transactions on communications and wireless networks
I am a 2021 B.Tech graduate student at the National Institute of Technology,
Warangal, from Electronics and Communication Engineering. I started research in machine
learning in my freshman year of UG. My propensity to research led me to take up various
internships and research projects during my undergraduate study. I have collaborations at
the Indian Institute of Science, Bangalore; the Indian Institute of Technology, Dharwad, and
I also worked as a part-time Machine Learning Researcher at Skylark Labs LLC.
During my projects, I explored various domains of ML including non-convex
optimization, Deep Learning, Computer Vision, Deep Reinforcement Learning, and Meta learning. I'm
currently seeking positions to pursue MSc or a Ph.D.