a program which converts Indian sign language to text n real time using openCV and CNN. three filters were applied to the RoI (region of interest), a greyscale filter, a gaussian blur filter and a threshold filter to get the black and white image which was recognised by the neural network running over 5 epochs. the dataset we created had around 22,000 images in the training set and 11,000 images in the test set.
see on github
For my dissertation titled "Behavioural Differences in Intelligent Virtual Agents With and Without Active Systems", I trained a completely clueless virtual hummingbird present on a virtual island to locate and drink nectar from virtual flowers using only it's eyes (RGB camera, active vision) and then compared the behaviour to a hummingbird with access to location and orientation of the flowers and itself.
Read the thesis here.
An implementation of a Genetic Algorithm that I designed to maximise the revenue for a supermarket by deciding the prices of 20 different articles, each ranging anything between £0.01 to £10. There are two solution approaches attempted one of which is exploitative (almost elitist) and the other one is quite exploratory. Both the solution approaches use different crossover, mutation and selection approaches.
May the fittest solution survive.
See on Github
With a dataset of ~600,000 rows and a class imbalance of 99% to 1% in the data, I evaluate different classification algorithms and empirically decide on the best solution among the candidate solutions which predicts fraud transactions with an accuracy of ~75% without oversampling the data and ~99% with oversampling the data.
Click here to read the whole analysis.
or
See on Github
Drag and Drop Website Builder