Interpretable Machine Learning
Event details
Date | 10.05.2017 |
Hour | 10:00 › 11:30 |
Speaker |
Professor Fernando Perez-Cruz Associate Professor in Computer Science Stevens Institute of Technology |
Location | |
Category | Conferences - Seminars |
Extended Abstract
Machine learning relies on two general approaches to learning from data: discriminative and generative modeling. Discriminative techniques focus on solving a clear well-defined supervised learning task. These models use labeled datasets to predict an output given any potential input. On the contrary, generative approaches build a probability density model for the data, and do not have a clear task or metric at hand, so potentially they could solve any.
Another way of thinking about these two approaches is to look at discriminative learning as a one-way conversation, and at generative modeling as a two-way one. The goal of discriminative models is to build a machine as accurate as possible, given some labeled data and a specific metric. Generative modeling, on the other hand, requires the interaction between data owner (expert in the field) and a machine-learning practitioner (building the model). Generative models should include all prior intangible information known to the data owner and should also be understandable to the data owner because the purpose of these models is not reducing some error measure but learning and understanding the data. In this sense, generative models need to be either human interpretable and actionable, and/or point to causal interactions.
In today’s talk, we will first discuss how Bayesian nonparametric (BNP) models can be used to find hidden patterns in data. BNPs return latent variable models that following De Finetti’s Theorem, find latent variables for which the observed data are conditional independent. From the analysis of this latent variable model, we can gain knowledge about the system that generated the data and draw actionable conclusions and propose tests to verify the found connections.
Then, towards the end of the presentation we will focus on the application of BNPs to modeling psychiatric disorders, geo-localization with LTE networks and trading relationships. These latent variable models yield not only intuitive, but perhaps unsurprising findings, but also relevant interactions not evident to the naked eye.
Machine learning relies on two general approaches to learning from data: discriminative and generative modeling. Discriminative techniques focus on solving a clear well-defined supervised learning task. These models use labeled datasets to predict an output given any potential input. On the contrary, generative approaches build a probability density model for the data, and do not have a clear task or metric at hand, so potentially they could solve any.
Another way of thinking about these two approaches is to look at discriminative learning as a one-way conversation, and at generative modeling as a two-way one. The goal of discriminative models is to build a machine as accurate as possible, given some labeled data and a specific metric. Generative modeling, on the other hand, requires the interaction between data owner (expert in the field) and a machine-learning practitioner (building the model). Generative models should include all prior intangible information known to the data owner and should also be understandable to the data owner because the purpose of these models is not reducing some error measure but learning and understanding the data. In this sense, generative models need to be either human interpretable and actionable, and/or point to causal interactions.
In today’s talk, we will first discuss how Bayesian nonparametric (BNP) models can be used to find hidden patterns in data. BNPs return latent variable models that following De Finetti’s Theorem, find latent variables for which the observed data are conditional independent. From the analysis of this latent variable model, we can gain knowledge about the system that generated the data and draw actionable conclusions and propose tests to verify the found connections.
Then, towards the end of the presentation we will focus on the application of BNPs to modeling psychiatric disorders, geo-localization with LTE networks and trading relationships. These latent variable models yield not only intuitive, but perhaps unsurprising findings, but also relevant interactions not evident to the naked eye.
Practical information
- General public
- Free
- This event is internal
Organizer
- Dr. Olivier Verscheure
Contact
- Dr. Olivier Verscheure