SEARCH SESSIONS

Search All
 
Refine Results:
 
Year(s)

SOCIAL MEDIA

EMAIL SUBSCRIPTION

 
 

GTC ON-DEMAND

Presentation
Media
Abstract:
We'll explore how to discover properties of deep networks by looking at their learned parameters or measuring the patterns of the networks' input space. Emerging properties from individual samples can be measured by examining the common changes they undergo during training. We'll explain how this allows a hierarchical analysis that goes beyond explainability of individual decisions why a particular image was misclassified, for example and extends to entire classes or even the training dataset itself. We show how understanding these patterns can provide the foundation for more principled, stable, and robust definitions of future network architectures and more consistent learning procedures.
We'll explore how to discover properties of deep networks by looking at their learned parameters or measuring the patterns of the networks' input space. Emerging properties from individual samples can be measured by examining the common changes they undergo during training. We'll explain how this allows a hierarchical analysis that goes beyond explainability of individual decisions why a particular image was misclassified, for example and extends to entire classes or even the training dataset itself. We show how understanding these patterns can provide the foundation for more principled, stable, and robust definitions of future network architectures and more consistent learning procedures.  Back
 
Topics:
Advanced AI Learning Techniques, Computer Vision
Type:
Talk
Event:
GTC Silicon Valley
Year:
2019
Session ID:
S9287
Streaming:
Download:
Share:
 
Abstract:

Interpretability of deep neural networks has focused on the analysis of individual samples. This can distract our attention from patterns originating at the distribution of the dataset itself. We broaden the scope of interpretability analysis, from individual images to entire datasets, and found that some high-performing classifiers use less than half the information contained in any given sample. While the learned features are more intuitive to visualize for image-centric neural networks, in time-series it is much more complicated as there is no direct interpretation of the filters and inputs as compared to image modality. In this talk we are presenting two approaches to analyze the behavior of networks with respect to used input signal, which pave the way to go beyond simple layer stacking and towards a more principled design of neural networks.

Interpretability of deep neural networks has focused on the analysis of individual samples. This can distract our attention from patterns originating at the distribution of the dataset itself. We broaden the scope of interpretability analysis, from individual images to entire datasets, and found that some high-performing classifiers use less than half the information contained in any given sample. While the learned features are more intuitive to visualize for image-centric neural networks, in time-series it is much more complicated as there is no direct interpretation of the filters and inputs as compared to image modality. In this talk we are presenting two approaches to analyze the behavior of networks with respect to used input signal, which pave the way to go beyond simple layer stacking and towards a more principled design of neural networks.

  Back
 
Topics:
Artificial Intelligence and Deep Learning
Type:
Talk
Event:
GTC Europe
Year:
2018
Session ID:
E8450
Streaming:
Download:
Share: