GTC ON-DEMAND

 
SEARCH SESSIONS
SEARCH SESSIONS

Search All
 
Refine Results:
 
Year(s)

SOCIAL MEDIA

EMAIL SUBSCRIPTION

 
 

GTC ON-DEMAND

Presentation
Media
Abstract:
We'll talk about how we're incorporating physics into deep learning algorithms. Standard deep learning algorithms are based on a function-fitting approach that does not exploit any domain knowledge or constraints. This makes them unsuitable for applications like robotics that require safety or stability guarantees. These algorithms also require large amounts of labeled data, which is not readily available. We'll discuss how we're overcoming these limitations by infusing physics into deep learning algorithms, and how we're applying this to stable landing of quadrotor drones. We've developed a robust deep learning-based nonlinear controller called Neural-Lander, which learns ground-effect aerodynamic forces that are hard to model. We'll also touch on how Neural-Lander can land significantly faster while maintaining stability.
We'll talk about how we're incorporating physics into deep learning algorithms. Standard deep learning algorithms are based on a function-fitting approach that does not exploit any domain knowledge or constraints. This makes them unsuitable for applications like robotics that require safety or stability guarantees. These algorithms also require large amounts of labeled data, which is not readily available. We'll discuss how we're overcoming these limitations by infusing physics into deep learning algorithms, and how we're applying this to stable landing of quadrotor drones. We've developed a robust deep learning-based nonlinear controller called Neural-Lander, which learns ground-effect aerodynamic forces that are hard to model. We'll also touch on how Neural-Lander can land significantly faster while maintaining stability.  Back
 
Topics:
AI & Deep Learning Research
Type:
Talk
Event:
GTC Silicon Valley
Year:
2019
Session ID:
S9732
Streaming:
Download:
Share:
 
Abstract:
Learn about tensors, higher-order extensions of matrices that can incorporate multiple modalities and encode higher-order relationships in data. After an introduction to tensor methods, we will discuss which tensor methods can be used in deep learning and in probabilistic modeling. We'll show how tensor contractions, which are extensions of matrix products, provide high rates of compression in a variety of neural network models. We'll also demonstrate the use of tensors for document categorization at scale through probabilistic topic models. These are available in a python library called Tensorly that provides a high-level API for tensor methods and deep tensorized architectures.
Learn about tensors, higher-order extensions of matrices that can incorporate multiple modalities and encode higher-order relationships in data. After an introduction to tensor methods, we will discuss which tensor methods can be used in deep learning and in probabilistic modeling. We'll show how tensor contractions, which are extensions of matrix products, provide high rates of compression in a variety of neural network models. We'll also demonstrate the use of tensors for document categorization at scale through probabilistic topic models. These are available in a python library called Tensorly that provides a high-level API for tensor methods and deep tensorized architectures.  Back
 
Topics:
AI & Deep Learning Research
Type:
Tutorial
Event:
GTC Silicon Valley
Year:
2019
Session ID:
S9733
Streaming:
Download:
Share:
 
Abstract:
标准深度学习算法基于不使用任何领域知识或约束的功能拟合方法。 这有几个缺点:高样本复杂性,并且缺乏鲁棒性和泛化性,尤其是在领域或任务转移下。 我将展示几种注入结构和领域知识以克服这些限制的方法,即张量,图形,符号规则,物理定律和模拟。 标准的深度学习算法基于不使用任何领域知识或约束的功能拟合方法。 这使其不适用于数据量有限或需要安全性或稳定性保证的应用程序,例如机器人技术。 通过将结构和物理注入到深度学习算法中,我们可以克服这些限制。 有几种方法可以做到这一点。 例如,我们使用张量神经网络对多维数据和高阶相关性进行编码。 我们将符号表达式与数值数据相结合,以学习功能域并获得强大的概括性。 我们将基线控制器与学习到的残余动力学相结合,以改善四旋翼无人机的着陆。 这些实例表明,将结构构建到 ML 算法中可以带来可观的收益。
标准深度学习算法基于不使用任何领域知识或约束的功能拟合方法。 这有几个缺点:高样本复杂性,并且缺乏鲁棒性和泛化性,尤其是在领域或任务转移下。 我将展示几种注入结构和领域知识以克服这些限制的方法,即张量,图形,符号规则,物理定律和模拟。 标准的深度学习算法基于不使用任何领域知识或约束的功能拟合方法。 这使其不适用于数据量有限或需要安全性或稳定性保证的应用程序,例如机器人技术。 通过将结构和物理注入到深度学习算法中,我们可以克服这些限制。 有几种方法可以做到这一点。 例如,我们使用张量神经网络对多维数据和高阶相关性进行编码。 我们将符号表达式与数值数据相结合,以学习功能域并获得强大的概括性。 我们将基线控制器与学习到的残余动力学相结合,以改善四旋翼无人机的着陆。 这些实例表明,将结构构建到 ML 算法中可以带来可观的收益。  Back
 
Topics:
Artificial Intelligence and Deep Learning
Type:
Talk
Event:
GTC China
Year:
2019
Session ID:
CN9402
Download:
Share:
 
 
Previous
  • Amazon Web Services
  • IBM
  • Cisco
  • Dell EMC
  • Hewlett Packard Enterprise
  • Inspur
  • Lenovo
  • SenseTime
  • Supermicro Computers
  • Synnex
  • Autodesk
  • HP
  • Linear Technology
  • MSI Computer Corp.
  • OPTIS
  • PNY
  • SK Hynix
  • vmware
  • Abaco Systems
  • Acceleware Ltd.
  • ASUSTeK COMPUTER INC
  • Cray Inc.
  • Exxact Corporation
  • Flanders - Belgium
  • Google Cloud
  • HTC VIVE
  • Liqid
  • MapD
  • Penguin Computing
  • SAP
  • Sugon
  • Twitter
Next