Exploring Limits to Face Detection Alignment and Recogntion
As part of computer vision and biometrics team at RBS,
I had a chance to experiment with the state of art biometrics and face analytics.
Insipired from same I have made cetrain efforts to open source solution to a problem
that I inititally encountered while starting the innovation project at RBS of having a
onestop shop solution for all the face recogntiion tasks such as detection,
recogntion, alignment, landmark detection and verficiation. The current project aim is to provide multi
viewpoint invaraint detection and landmark detection that can help in better recognition on devices in real time.
The algorithm is lightweight and is designed to be able to produce near real time FPS on CPU and therefore,
is edge compute compatible i.e. can run on hardware with computing power constraints.
Optimal Reservoir Operations using Long Short-Term Memory Network
This work proposes a LSTM model for daily inflow determination and a naive anomaly detection algorithm
base-line using LSTM. In other words a strong baseline to predict floods and droughts for any prediction model.
Experiments to prove the efficacy of LSTM for inflow prediction are done on corpus of daily inflow data of
past 20 years for Bhakra Dam. Although, experiments are run on data from Bhakra Dam Reservoir in India,
LSTM model and anomaly detection algorithm are general purpose and can be applied to any basin with minimal changes.
Language Identification for Hindi Language transliterated text to Roman Script using GANs
Deepak Kumar Shamra, Anurag Singh, Abhishek Saroha
Book Chapter Link Acceptance
Towards Extensible and Adaptable Methods in Computing pp 267-279 Springer Nature Singapore, 2018
Use of DL for transliterated text launguage identfication.
This work aimed to identify text based content written in roman script which conveys meaning in Hindi
language. The project proposes a methodology to identify language
based on semantic meaning of the text. The model was tested and propsed with novel features for Hindi,
for a corpus of ~10 k lines of transliterated Hindi text scraped from web
AI Solver for 2048
An artificial intelligence for the 2048 game based on expectiminmax algorithm
Commerical DNA Storage
There are multiple aspects and areas of problems to solve to make DNA storage commercial.
Why are excited about storing information in DNA? How many of us can read the floppy drives now?
Its very hard to read them becuase the equipments available to read them are hard to find now.
We shift our storage medium iteratively copying the information. But no method of storage is currently long term.
The current storage systems shall render themselves redundant soon. So what is the solution?
DNA at its worst has a half life of over 500 years and DNA data from over 50k years of unattended fossils have been successfully
sequenced. With more than 90% of data being generated in past no matter how much horizontally we scale our servers data will
increase exponentially. Therefore the solution to offer to such massive data storage crisis is to find
a media that can store huge amount of data in lot less space.
Again we find DNA to be the answer.