This is part of a commercial project that finds a face in an image and crops as symmetrically as possible to fit on ID cards .
Programmatically created region of interest (ROI) and associated Sentinel-2 Multispectral data with the measured Soil Organic Carbon (SOC) ground truths.
Using 3D convolutional neural network, developed a model that detects class of a hyperspectral image patch.
From a multi-labelled face dataset, predict age and gender.
Recognises emotion based on SAVEE dataset using CNN.
Currently, working on using transformer network to capture temporal relation for better accuracy.
Saliency map highlights the pixels (in computer vision), which are responsible in making a decision. This projects generates saliency map on an MNIST digit classifier.
Checks if the person in the image is wearing glasses. Used Resnet-18 as backbone.
Predicting soil moisture from time-series data from different sensors including hyperspectral camera.
Predicting Amazon stock price using LSTM.
Detects CKD with 98% accuracy.
Using partial derivative method, depicts impact of each feature in decision making process of the machine.
Simple NN regression model to estimate N2O emissions
Simple Autoencoder on MNIST dataset
Simple VAE on MNIST dataset. Used tensorboard to visualise continuous progress.
Simple FC network to recognise short speech commands (for example, forward, up, yes, no etc.).
An experimental work in search of explainable and deterministic neural networks.
First started with creating 28X28 filters for each classes (digits in this case) by accumulating the average distribution of the pixels.
Without any backpropagation, it works with 77% accuracy. Improving the filters with backpropagation (not randomly initialized) produces 88%+ accuracy.
Future work in this would be applying DSC and addressing the issue of lack of translational invariance.
CNN built from scratch to recognise speaker from mel spectrogram data.
A simple neural network from scratch for demonstration purpose
Aligning a polynomial function to sine curve for demonstration purpose
Classic snake game. A foundation work to learn reinforcement learning.
Simulate Conway's Game of Life. Auto zoom-out when generations exceed borders.
Implementing Quad (Samuel Beckett) by using dynamic mesh generation. Here is a quick demo.
An attempt to learn fundamental of genome sequencing to be able to apply machine learning.
In this game AI agent has to escape through the hole avoiding enemies.
Done: Basic gameplay
Remaining: AI Agent
Spotting tumor cells on brain using the BraTS dataset
Done so far: Processing nii files. Prerequisite knowledge - U-NET, 3D Convolution, DSC similarity etc.
Using CelebA dataset, generating new faces.
Done so far: Initial codes and tensorboard setup for debug and diagnosis.