Astuary

Spry

Code for "Thinking Forward: Memory-Efficient Federated Finetuning of Language Models" (NeurIPS 2024). Spry is a federated learning algorithm that enables finetuning LLMs using Forward-mode Auto Differentiation; to achieve low memory footprint, high accuracy, and fast convergence.

Flow

Code for "Flow: Per-Instance Personalized Federated Learning" (NeurIPS 2023). Flow addresses the challenge of statistical heterogeneity in federated learning through creating a dynamic personalized model for each input instance through a routing mechanism.

MultiagentChatbot

Code for "CommunityBots: ... A Multi-Agent Chatbot Platform..." (CSCW 2023). A multi-agent chatbot system (A single chat window having multiple chatbots with different personalities interacting with the user, giving a feel of group discussion) built on ReactJS.

MultimodalFusion

A novel multimodal fusion model to classify 6 classes of emotions from visual, verbal, and vocal features of a person’s speech in CMU-MOSEI dataset.

BoseSoundTouchAPI

Created at Hack(h)er413 2020. Won "Best Use of the Bose API" by Bose and "Best Hack for Making your Home Accessible" by Wayfair

SkinCancer

Final Year Project (BEngg at MSU, Vadodara) on Skin Cancer Detection using Matlab Image Processing Toolbox

ColorImageDemosaicing

In digital cameras the red, blue, and green sensors are interlaced in the Bayer pattern. The missing values are interpolated to obtain a full color image. In this project, I implemented several interpolation algorithms.

LSTMVariantsForTimeSeriesForecasting

I have used the ExtraSensory data set. This is a human activity recognition data set containing data from 60 individuals. The task that I will focus on, is probabilistic activity forecasting. Given a sub-dataframe consisting of between one and 30 consecutive observations for a single individual and a timestamp value t, the objective is to predict the (log) probability that each of the five labels (label:LYING_DOWN, label:SITTING, label:FIX_walking, label:TALKING, label:OR_standing) is active (e.g., takes the value 1) at the future specified time, t.

ModernComputerArchitectureResearchReviews

My reviews of the research published for computer architecture which includes: Methodology, Experience, Wearout, Memory, Cache, Multithreading, Prediction, Pipelining, Instruction-level Parallelism, Branch Prediction, Virtualization, Secondary Storage, Security, Parallelism, Data Parallelism, Data Centers

The_Cost_of_Avoiding_Backpropagation

Astuary/Flow

Code for "Flow: Per-Instance Personalized Federated Learning" (NeurIPS 2023). Flow addresses the challenge of statistical heterogeneity in federated learning through creating a dynamic personalized model for each input instance through a routing mechanism.

Astuary/Spry

Code for "Thinking Forward: Memory-Efficient Federated Finetuning of Language Models" (NeurIPS 2024). Spry is a federated learning algorithm that enables finetuning LLMs using Forward-mode Auto Differentiation; to achieve low memory footprint, high accuracy, and fast convergence.

Astuary/MultimodalFusion

A novel multimodal fusion model to classify 6 classes of emotions from visual, verbal, and vocal features of a person’s speech in CMU-MOSEI dataset.

Astuary/ColorImageDemosaicing

In digital cameras the red, blue, and green sensors are interlaced in the Bayer pattern. The missing values are interpolated to obtain a full color image. In this project, I implemented several interpolation algorithms.

Astuary/LSTMVariantsForTimeSeriesForecasting

I have used the ExtraSensory data set. This is a human activity recognition data set containing data from 60 individuals. The task that I will focus on, is probabilistic activity forecasting. Given a sub-dataframe consisting of between one and 30 consecutive observations for a single individual and a timestamp value t, the objective is to predict the (log) probability that each of the five labels (label:LYING_DOWN, label:SITTING, label:FIX_walking, label:TALKING, label:OR_standing) is active (e.g., takes the value 1) at the future specified time, t.

Astuary/ModernComputerArchitectureResearchReviews

My reviews of the research published for computer architecture which includes: Methodology, Experience, Wearout, Memory, Cache, Multithreading, Prediction, Pipelining, Instruction-level Parallelism, Branch Prediction, Virtualization, Secondary Storage, Security, Parallelism, Data Parallelism, Data Centers

Astuary/AdvancedAlgorithms

This repository contains detailed description of solutions of the chosen problems from the book "Introduction to Algorithms" by Cormen, Leiserson, Rivest, and Stein.

Astuary/AspectBasedSentimentAnalysis

Data augmentation for sentiment analysis through synonym swapping, and smart masking.

Astuary/BlobDetection

Built a Laplacian scale space, starting with some initial scale and going for n iterations. Performed/ non-maximum suppression in scale space. Displayed resulting circles at their characteristic scales for points above a threshold.

Astuary/BoseSoundTouchAPI

Created at Hack(h)er413 2020. Won "Best Use of the Bose API" by Bose and "Best Hack for Making your Home Accessible" by Wayfair

Astuary/ChemistryCanvas

Astuary/ComputerVisionProblems

Answers and Explanations to some Computer Vision Problems including: Radiometry, Light, Color, Image Formation, Optical Flow, Image modeling, Linear Filtering, Image Filtering, Matching and Alignment, Recognition using K-Means Clustering

Astuary/DeepImagePrior

Denoising images with 1. AutoEncoder with Batch Normalization and Bilinear Upsampling and with 2. Deeper Architectures, those with Skip Connections, as well as Averaging the Results across iterations.

Astuary/DepthFromDisparity

I used a pair of images to compute a depth image of the scene. I did this by computing a disparity map. A disparity map is an image that stores the displacement that leads every pixel in an image X to its corresponding pixel in the image. The depth image is inversely proportional to the disparity.

Astuary/HandcraftedImageRepresentation

Implemented three different image representations for MNIST dataset: Raw Pixels, Histogram of Gradients and Local Binary Patterns

Astuary/ImageDenoising

Evaluated two simple denoising algorithms, namely Gaussian filtering and median filtering. In addition, also implemented a more advanced algorithm based on nonlocal means.

Astuary/Kunjal

Template for my portfolio website, built using GatsbyJS.

Astuary/Kunjal-MS

My portfolio website - During MS

Astuary/Kunjal-PhD-2

Portfolio Website - PhD

Astuary/MultiagentChatbot

Code for "CommunityBots: ... A Multi-Agent Chatbot Platform..." (CSCW 2023). A multi-agent chatbot system (A single chat window having multiple chatbots with different personalities interacting with the user, giving a feel of group discussion) built on ReactJS.

Astuary/OptimizedModels

Regularized and Augmented Linear Regression, Logistic Regression with Huber Loss, Geometric Regression, Support Vector Classifiers, Neural Network with Adam Optimizer and Stochastic Gradient Descent & CNN with Elastic Deformation; all these models are implemented from scratch using NumPy, SciPy, Scikit-learn and PyTorch.

Astuary/PanoramicStitching

Implemented the RANSAC algorithm to stitch two images. The input to the algorithm are two images which are related by an unknown transformation. I used the blobs detector implemented to extract keypoints and extract feature descriptors on them. Then estimated an affine transformation using feature matching and RANSAC to produce a combined image.

Astuary/PingMyFeeling

EMPath project.

Astuary/ResearchPapers

A common place to put all my academic papers.

Astuary/ScaledAndJitteredMNISTClassifier

Classifying MNIST normal, scaled and jittered dataset with Multiclass Logistic Regression and MatConvNet CNNs

Astuary/Surestart_Spring_2023_VAIL

Astuary/TextureSynthesis

A simple approach for generating a texture is to randomly tile the target image using patches from a source image. It results in an image with artifacts around the edge of the tiles. This is because the tiles do not align well. The approach of Efros and Leung avoids this by generating the ouput, one pixel at a time, by matching the local neighborhood of the pixel in the generated image and source image. The random tiling approach works reasonably well for some textures and is much faster than the pixel-by-pixel synthesis approach. However, it may have visible artifacts at the border. The “Image Quilting” approach from Efros and Freeman, proposes a way to minimize these by picking tiles that align well along the boundary (like solving a jigsaw puzzle).

Astuary/TourItYourself

HackUMass project on Geofencing for tourist safety

Astuary/VideoToSeamlessGIFs

From the original Microsoft paper on "Video Textures", here the goal was to improve distance metric by switching from a crude sum of squared distance to most sophisticated wavelet-based distance; add intensity normalization, cross-fading and morphing to the suggested basic algorithm. We also experiment on the trade-off between variety and smoothness.

Astuary/WhatsappDataAnalysis

Programs to analyze whatsapp exported chat data in various ways. Currently: Chat Frequency between Two people with respect to days