Used ensemble methods to classify single nucleotide polymorphism (SNP) genotype data(synthetic) of 4000 cases and 4000 controls. Data contained 29623 features among which only 15 were causal. Used PCA to reduce dimensions and achieved Top-1 accuracy of 66.2%.

Developed a single layer neural network in python and implemented adaptive learning rate in gradient descent. Applied this method for both Hinge Loss and Least Squares Loss and studied its effects. From this project, I understood the profound effects a learning rate can have on the model.

Developed multiple classifiers using Convolution Neural Networks for a wide range of datasets from Kaggle for image classification problem. Models were developed using Keras.

Developed a 5 layer CNN network to perform classification on Mini-Imagenet dataset, to satisfy a requrment of minimmum accuracy of 80%. I developed a deep network with a 50% dropout for each layer, which sounds strange, but performed well. I learnt the effects of dropout during training and the effects of high dropout value.

Implemented feature space transformation for multi-modality data. Built a transformer that would perform feature space transformation of data from multiple unknown distributions into a known distribution. This helps reduce development time and engineering effort, by having a common modal which works on data from different modalities

A project to help users visually understand links between any aribitary datasets and also understand what information can be gained by such links (e.g., linking data sets A, D, and C might provide less information than linking datasets G, D, and F).