SLIDE 1
Hybrid Models with Deep and Invertible Features
Eric Nalisnick*, Akihiro Matsukawa*, Yee Whye Teh, Dilan Gorur, Balaji Lakshminarayanan
*equal contribution
Hybrid Models with Deep and Invertible Features
Hybrid Models with Deep and Invertible Features Eric Nalisnick *, - - PowerPoint PPT Presentation
Hybrid Models with Deep and Invertible Features Eric Nalisnick *, - - PowerPoint PPT Presentation
Hybrid Models with Deep and Invertible Features Eric Nalisnick *, Akihiro Matsukawa*, Yee Whye Teh, Dilan Gorur, Balaji Lakshminarayanan *equal contribution Predictive Models Hybrid Models with Deep and Invertible Features Predictive Models
SLIDE 2
SLIDE 3
Hybrid Models with Deep and Invertible Features
Generative Models Predictive Models
SLIDE 4
Hybrid Models with Deep and Invertible Features
Predictive Models Generative Models
Can we efficiently combine them to model p(y, x)?
SLIDE 5
Hybrid Models with Deep and Invertible Features
Neural Hybrid Model
We define a computationally efficient hybrid model by combining normalizing flows with generalized linear models (GLMs).
SLIDE 6
Hybrid Models with Deep and Invertible Features
Neural Hybrid Model
We define a computationally efficient hybrid model by combining normalizing flows with generalized linear models (GLMs). Predictive Component Generative Component
SLIDE 7
Hybrid Models with Deep and Invertible Features
Neural Hybrid Model
We define a computationally efficient hybrid model by combining normalizing flows with generalized linear models (GLMs). Normalizing Flow Linear Model
SLIDE 8
Hybrid Models with Deep and Invertible Features
Neural Hybrid Model
We define a computationally efficient hybrid model by combining normalizing flows with generalized linear models (GLMs). Input features.
SLIDE 9
Hybrid Models with Deep and Invertible Features
Neural Hybrid Model
We define a computationally efficient hybrid model by combining normalizing flows with generalized linear models (GLMs). Normalizing flow acts as a deep neural feature extractor.
SLIDE 10
Hybrid Models with Deep and Invertible Features
Neural Hybrid Model
We define a computationally efficient hybrid model by combining normalizing flows with generalized linear models (GLMs). Flow’s output and params. are used to compute p(x) via change-of-variables.
SLIDE 11
Hybrid Models with Deep and Invertible Features
Neural Hybrid Model
We define a computationally efficient hybrid model by combining normalizing flows with generalized linear models (GLMs). Flow’s output is used as the feature vector in a (generalized) linear model, which computes p(y|x).
SLIDE 12
Hybrid Models with Deep and Invertible Features
Neural Hybrid Model
We define a computationally efficient hybrid model by combining normalizing flows with generalized linear models (GLMs). Optimization objective:
Weight to trade-off predictive and generative performance.
SLIDE 13
Hybrid Models with Deep and Invertible Features
Simulation: Heteroscedastic Regression
Gaussian process fitted to simulated data.
SLIDE 14
Hybrid Models with Deep and Invertible Features
Simulation: Heteroscedastic Regression
Gaussian process fitted to simulated data. Our model’s predictive component.
SLIDE 15
Hybrid Models with Deep and Invertible Features
Simulation: Heteroscedastic Regression
Gaussian process fitted to simulated data. Our model’s predictive component. Our model’s generative component.
SLIDE 16