GTC 2020: Cooperative Neural Networks

GTC 2020 S21871
Presenters: Harsh Shrivastava,Georgia Tech
We’ll walk you through a novel approach to come up with domain-specific deep-learning architectures called cooperative neural networks (CoNN). CoNN incorporates the known prior information of the domain in its architecture by exploiting the structure of the underlying probabilistic graphical model describing the domain. We’ll demonstrate our approach for the document classification task, where we transfer the independence structure of the popular Latent Dirichlet Allocation (LDA) model to a cooperative neural network, CoNN-sLDA. We’ll show that the CoNN-sLDA model outperforms existing state-of-the-art techniques. CoNN-sLDA model has considerably fewer parameters and gives a significant runtime improvement compared to existing deep-learning models.

Pre-requisites: Basic familiarity with Deep Learning and Probabilistic Graphical Models.
Nvidia Hardware details: Experiments were ran on the Tesla V100 GPUs using Nvidia DGX Workstation.

Watch this session
Join in the conversation below.