Navpreet Singh

GPT2 classifier using few shot learning and language model classification

Implemented zero and few shot learning using GPT2 transformer from huggingface to classify AI and non AI relevant papers with just 10 examples in the training. Also experimented with different classification layers and convolution on top of GPT2 language model output training it with 20 samples, getting better score on precision-recall metrics compared with the few shot learning approach

Fake image generator conditioned on facial attributes

Implemented and trained Conditional DCGAN as a part of NYU Computer Vision Course project, using the celebrity dataset with facial attributes and base deep convolutional GAN architecture with Minimax loss function. Conditioned the generator on facial attributes like hair in order to produce fake images containing such attributes.

Intelligent trading bots using Machine Learning.

Designed and developed trading bots running Percent of Volume (POV) algorithm to execute trades on Eurodollars securities leveraging Machine Learning to predict the market volumes and classify them as Bull/Bear and executing trades accordingly.

Track vehicles on the road and estimate their speed.

Implemented using Pretrained YOLOv4 on COCO dataset for object detection and DeepSort Model for object tracking. The application is able to detect vehicles in a video footage and estimate their speed in pixels/sec.

Simple chatbot trained on Movie Dialogues Dataset

I trained a simple seq2seq model Chatbot using Pytorch based on multi-layered Gated Recurrent Unit (GRUs) and Attention mechanism using movie scripts dataset from the Cornell Movie-Dialogs Corpus.

Detection of tumors using Lung Nodule Analysis in CT Scans

I trained an end-to-end complex deep neural network using Pytorch based on Lung Nodule Analysis dataset of chest CT scans, finding the nodules using simple CNN residual network architecture and then identifying the tumours using state of the art U-Net Architecture for Object Detection and Image segmentation and then using transfer learning to use the same resnet model to train and classify them as benign/malignant.

Classification of DNA Sequence based on promoter region (where protein contact with DNA helix)

I cleaned and prepared UCI DNA sequences data. Used exploratory data analysis to do feature engineering and trained the dataset on KNN, Decision Trees, SVM, Random Forest and Adaboost classifiers using pandas and sklearn to select the best performing model which was SVM.

Predicting house prices using Liner Regression trained on Ames Housing Dataset

I cleaned, prepared, and visualised Ames Housing dataset and performed feature selection and feature engineering using pandas, seaborn and scipy. Then trained the data on Linear Regression techniques lasso and gradient boosting using sklearn and used ensemble method stacking and averaging the models to produce the best fit for the data.