Sree Harsha Nelaturu

Deep Learning Researcher


Projects


Psych.ai Combining art+deep learning to enable media enhancement of entertainment content. First place winner at Microsoft GAINS AI Hackathon (2017)

Endangered Tales Automatic Comic generation for spreading awareness about climate change/sustainability. Best use of OpenAI API at Stanford Treehacks (2021)

StoryBoxVR Party game using VR and language modelling.

Fair + EfficiencyFederated Learning Gradient based learning leveraging sparsity is desirable in and has found success in centralized paradigms as well as in federated learning. A key challenge beyond the communication bottleneck in FL is device fairness, in algorithms like FedAvg [esp in non-IID settings] there is significant variance in performance. Our projects aims to leverage sparsity to address this.

Lips sometimes lie: Towards undestanding lip reading networks Class project for class on High Level Computer Vision, exploring the limitations of AV-HuBERT for lip-reading tasks.

Study of effect of Optimizers on Accelerated Backpropgation Techniques Investigation of optimizers on networks trained with approximate and sparse operations for a class on Optimization for Machine Learning.

Deep Greetings Using LSTMs and Neural Style Transfer for generating greeting cards.


Talks