Deep Transformer Soft Actor-Critic Network for Reinforcement Learning

  • Utilize Transformer as memory module for both Actor and Policy networks
  • Hyperparameter tuning for SAC performance

Sentiment Analysis on MyAnimeList User Ratings

  • MyAnimeList is a popular anime rating website.
  • Predict user rating based on review using Recurrent Neural Network (RNN)
  • Setup a data-mining pipeline utilizing self-hosted REST API with a Redis server for caching inside dockerized container
  • Used different models (RNN with LSTM, CNN, CNN with Word2Vec embedding layers) for training and stacking model for ensemble.
  • Achieved 94% validation accuracy with ensemble model

Classification of Extended MNIST using Persistent Homology

  • EMNIST dataset is MNIST (handwritten digit) dataset with handwritten characters
  • Applied Persistent Homology and Principal Component Analysis to reduce the dimensionality of dataset. Reduced feature size from 784 (28x28) to 35 while retaining 99% variance
  • Utilized libraries giotto-ai along-side standard deep learning libraries sklearn, NumPy, Tensorflow/Keras
  • Achieved 97%-91% training-testing accuracy with a shallow neural network with only 3 hidden layers

Utilization of CNN in speech recognition

  • Classify Google Speech Command using Convolutional Neural Network (CNN) on audio data
  • Created pipelines to process audio data to image features
  • Audio data augmentation with respect to image features
  • Used multiple CNN architectures (LeNet, MiniGoogleNet, AlexNet) for training and stacking model for ensemble.
  • Achieved 91% validation accuracy with ensemble model.