Projects
Exploiting Correlation in Online Sequential Decision Making
About: The project deals with modelling and exploiting correlation between the successes and failures of multiple communication channels. Performance is determined by a metric relevant to time-sensitive applications called Age-of-Information (AoI). A generalization of the Multi-Armed Bandit (MAB) model is used as the decision framework.
Key Contributions:
Designed an MAB algorithm for regret minimization under a correlated MAB framework and demonstrated its superior performance on certain bandit instances.
paper / poster / slides / code
Proposed a variant of the successive-elimination algorithm for best-arm identification under the correlated MAB framework. Analyzed the sample-complexity for our (0, δ)-PAC algorithm.
report / slides
Proved instance independent lower bounds and matching upper bounds on AoI-regret for correlation- aware variants of the UCB and Thompson Sampling policies.
Achieved significantly improved performance on AoI-regret by exploiting underlying correlation.
paper / presentation / slides / code
Video Stabilization for a Spherical Mobile Robot
About: Camera stabilization equipment prevents physical disturbances from distorting video footage. However, the correction is not perfect since low frequency disturbances from actions like rolling on an irregular surface creep into a mechanically stabilized recording and degrade the viewing experience. The problem can be alleviated using software-based video stabilization.
formulation / code
- Implemented a linear-programming based offline video-stabilization technique that estimates and smoothens the underlying camera trajectory and performs a fresh rendering of the video.
- Devised a novel objective and additional constraints to use the technique in a windowed fashion allowing for parallelism and a fixed delay between live input and output video.
Sample Efficient Exploration in Reinforcement Learning
About: For reinforcement learning to be effective and tractable, knowledge about the environment must be gained within a reasonable number of time steps. The project looks at reward function synthesis and assistance from a near-optimal expert-policy as two possible approaches.
Part 1
Ran experiments to compare the number of samples required by an agent to learn an optimal policy via Q-learning when it uses potential-based reward-shaping and optimal reward-search on an environment having a sparse intrinsic-reward.
report / code
Part 2
Designed an algorithm combining EXP3 and policy roll-outs to weight and combine knowledge from the learned environment model and an external expert policy. Tested it on the RiverSwim MDP.
preliminary-slides / code
Digital Immunoassay Based on Particle Motion Analysis
About: Biochemical immunoassay technology detects proteins and other macromolecules in low concentrations. Our team worked on a novel biochemical immunoassay that estimates molecular concentrations by tethering down proteins and analyzing their motion through statistical methods.
slides
- Applied Gaussian-Mixture-Models on experimental data to derive an accurate calibration curve.
- Sped up the data processing pipeline through the use of efficient Python data structures. Brought down time taken for parameter extraction from 30min to 2min for a typical experiment.
Image Segmentation in Baby-Monitor Footage
About: Each year thousands of infants succumb to suffocation arising from unsafe sleep positions. The project aimed to alleviate this problem using computer vision. Cradlewise is an early stage startup whose smart crib was recently featured in TIME Magazine’s 100 best inventions of 2020.
USB Powered High Voltage Nixie Tube Clock Display
report / blog-posts / video / code
- Designed an efficient DC-DC isolated voltage convertor capable of safely stepping up 5V to 200V.
- Printed a custom circuit board for the clock display front-end. The board featured signal isolation.
- Wrote embedded software for the on-board controller to display digits in a time-multiplexed manner.
Vehicle Routing Problems
A supervised learning project where I learnt about Integer programming as a discrete optimization technique and applied this knowledge to Vehicle Routing Problems.
Offside Detection in Football
A self initiated Summer project. Our team created a program to identify offside situations in football. The project utilized the functionality of the openCV library in Python.
Course Projects
Hopfield Networks as Dynamical Systems
- Formalised the Hopfield recurrent neural network as a dynamical system using a state-space model.
- Proved convergence of the network to designated fixed-points using the Lyapunov stability criteria.
Deep Learning for Metric Learning and GANs
- Implemented the Siamese and Triplet network architectures for Metric Learning using PyTorch.
- Created a GAN and used it to generate hand-written digit images by training it on MNIST digits.
Kalman Filters for Non-Linear Estimation
- Applied Non-Linear Kalman filtering to fuse position, velocity, and aerodynamic information to track the trajectory of a spacecraft during simulated atmospheric re-entry.
- Compared the performance of the Extended and Unscented Kalman filters on the re-entry problem.
Automatic Key-Word Recognition
- Trained a Hidden-Markov-Model word recognition program to recognize a vocabulary of 10 words.
- Achieved a classification accuracy of 70% on the Google speech commands dataset.
- Implemented speech synthesis using the source-filter model of speech.
16-Bit Pipelined RISC Processor
GitHub Repo for pipelined variant / GitHub Repo for multi-cycle variant
- Designed a 6-stage pipelined processor in VHDL and demonstrated it on an Intel DE0-Nano FPGA.
- Implemented the IITB-RISC, 14 instruction, simplified Instruction Set Architecture (ISA) with cycle saving optimizations including data-forwarding and branch-prediction.
Face Recognition with Fischer and Eigen Faces
- Compared the performance of principal component analysis (PCA) based Eigen-Faces and linear discriminant analysis based Fisher-Faces on the Yale face-recognition dataset.
- Recognized the strengths and weaknesses of the methods in relation to the brightness of faces
Electronic Stethoscope Project
This project designed an electronic analog front-end that would allow for stethoscope signals to be visualized on an oscilloscope.