Implemented NLP model to translate pseudocode lines to C++ code by finetuning pretrained Transformer models. Achieved state of the art performance, with 87% of code translations passing test cases. Experimented with different input types and context lengths by leveraging self-attention.
Implemented a multiclass, multilabel classifier for the Yelp Restaurant Photo Classification task on Kaggle, achieving Mean F1 Score within 2% of Kaggle competition winner. Used convolutional neural networks with Keras by implementing methods to tackle data imbalance and weakly labeled data.
Implemented a Deep Q-Learner for the Tetris-like game of 1010!, using deep learning with Keras and employing decaying epsilon-greedy policy, target networks and experience replay. Achieved slightly better than average human performance by experimenting with different neural network architectures (convolutional, fully connected) and hyperparameters.