Available Technologies

    Innovation Sphere

Projects Sharing Researchers

Communicate

Details

Project TitleInterpretable and High-Performance Predictive Models via Deep Neural Networks
Track Code2017-249
Short Description
Abstract
 
TagsDeep Learning, Interpretable predictive model, machine learning, neural networks, statistical interactions
 
Posted DateSep 25, 2017 7:16 PM

Market Opportunity

The interpretability of machine learning models is important due to implications for communicating these models to a variety of practitioners and stakeholders. Interpretability is especially important in domains where decisions can have major consequences, such as in healthcare. Deep neural networks have been recognized as some of the best performing machine learning methods. However, they are notorious for their uninterpretable black-box nature. Risks associated with using uninterpretable models often prevent practitioners from using such models.

USC Solution

USC researchers have developed a novel system for constructing an interpretable predictive model for deep learning applications in diverse domains, including healthcare, financial services, marketing, computer vision, speech recognition, etc. Empirical evaluation on both synthetic and real-world data showed the effectiveness of this system, which detects statistical interactions of input data more accurately and efficiently than does the state-of-the-art.

Value Proposition

  • Highly accurate and interpretable predictive model

  • Orders of magnitude faster than state-of-the-art models

  • Provides insights into both real-world data and neural network behavior

  • Provides understanding of how the model makes predictions

  • Allows a domain expert to reason about which input data affect a response variable in data

Applications

  • Predictive Modeling

  • Healthcare

  • Computer Vision

  • Speech Recognition

  • Finance

  • Marketing

Stage of Development

  • Experimentally validated

  • Available for exclusive and non-exclusive license

Intellectual Property

Status:

Provisional application to be filed


Key Publications:

Detecting Statistical Interactions from Neural Network Weights, 2017

Contact Information

Nikolaus Traitler, Licensing Associate

USC Stevens Center for Innovation

(213) 821-3550

traitler@usc.edu

Files

File Name Description
NCD 2017-249 - Interpretable and High-Performance Predictive Models via Deep Neural Networks.pdf None Download