Projects

Neural Architecture Search on Quantized Neural Nets | Python, PyTorch

Implemented QNNs from scratch using Pytorch. Wrote Binary layers and Convolutional layers and was able to replicate the paper accuracy 96% on MNIST
Integrated the above layers with the Neural Network Intelligence (NNI) library’s ENAS module to perform experiments and judge the accuracy and feasibility of QNNs. Reproduced similar accuracy of 53% as achieved by a normal Neural network layer
Sarcasm Detection in Amazon Review Dataset | Python, PyTorch | Github Link

Advisor - Surekha Bhanot
Built a Sarcasm Detection model from scratch using Pytorch. Achieved an accuracy of 83.45% for the Sentiment model and accuracy in the range of 55 - 62% on the OCEAN dataset using a CNN model on the word embeddings
Extracted the features from the above models, concatenated them and used them to train a CNN-SVM classifier on the Amazon Review Dataset. Achieved an accuracy of 88%

Implemented IBM model 1 and model 2 from scratch to translate between Dutch and English
Optimized the usage of dataset by using random sampling to select around 10,000 samples on which training was performed
Achieved a cosine score of 0.54 on a randomly sampled test data set

Built a Registration Software (RegSoft) to handle registration, accommodation and participation of incoming participants for one of Asia’s largest fest - Oasis.
Handled the issue of large number of requests using Signals in Django. Reduced the load on the webapp by 20% from which other services were requesting access