Project Report, CIC 4105, Carnegie Mellon University, Pittsburgh, PA
Presented a novel rein-forcement learning-based scheduler that performs dynamic partitioning of computation workload across multiple heterogeneous GPUs for large neural network inference tasks. Formated our results and contributions as a technical report and provided a slide presentation.
Poster Presentation, GHC 4 Commons, Carnegie Mellon University, Pittsburgh, PA
Provided a poster presentation on a Multimodal Dense Video Captioning pipleline. Presented our intuition of using knowledge distillation to expand pretrained LSTM vocabulary embedding space to a pretrained GPT-2 model with few-shot adaptation, raising performance on the baseline model.
Poster Presentation, GHC 6101 Collaborative Commons, Carnegie Mellon University, Pittsburgh, PA
Provided a poster presentation on a attention-based model for coreference resolution in interactive shopping, providing the results of our exploration such as tracking objects over turns, predicting object counts, and contrasting conversations to tackle prediction errors by existing models and optimizing performances Our model exceeded State of Art Performance (Max F1 score 69.5) on Situated Interactive Multimodal Conversation (SIMMC) 2.0 dataset