The Thesis Review

[13] Adji Bousso Dieng - Deep Probabilistic Graphical Modeling

Nov 26, 2020
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

Core Idea of Deep Probabilistic Models

  • Deep Probabilistic Graphical Modeling blends probabilistic graphical models with deep learning to gain interpretability and uncertainty quantification.
  • It uses neural networks within a generative process framework to add flexibility in model specification and inference.
ANECDOTE

Story of Topic RNN Creation

  • The idea for Topic RNN came from an internship where recurrent neural networks and topic models were combined.
  • It captures both local sequential dependencies with RNNs and global thematic context with latent topics.
INSIGHT

Value of Inductive Biases

  • Inductive biases in models allow faster training and control by encoding known structure explicitly.
  • Controllability in probabilistic models arises from encoding known dependencies, like patient-specific global latent variables in medical data.
Get the Snipd Podcast app to discover more snips from this episode
Get the app