GTC ON-DEMAND

 
SEARCH SESSIONS
SEARCH SESSIONS

Search All
 
Refine Results:
 
Year(s)

SOCIAL MEDIA

EMAIL SUBSCRIPTION

 
 

GTC ON-DEMAND

Presentation
Media
Abstract:
The neural network architecture (e.g. number of layers, types of layers, connections between layers, etc.) plays a critical role in determining what, if anything, the neural network is able to learn from the training data. The trend for neural network architectures, especially those trained on ImageNet, has been to grow ever deeper and more complex. The result has been ever increasing accuracy on benchmark datasets with the cost of increased computational demands. In this talk we demonstrate that neural network architectures can be automatically generated, tailored for a specific application, with dual objectives: accuracy of prediction and speed of prediction. Using MENNDL--an HPC-enabled software stack for neural architecture search--we generate a neural network with comparable accuracy to state-of-the-art networks on a cancer pathology dataset that is also 16x faster at inference. The speedup in inference is necessary because of the volume and velocity of cancer pathology data; specifically, the previous state-of-the-art networks are too slow for individual researchers without access to HPC systems to keep pace with the rate of data generation. Our new model enables researchers with modest computational resources to analyze newly generated data faster than it is collected.
The neural network architecture (e.g. number of layers, types of layers, connections between layers, etc.) plays a critical role in determining what, if anything, the neural network is able to learn from the training data. The trend for neural network architectures, especially those trained on ImageNet, has been to grow ever deeper and more complex. The result has been ever increasing accuracy on benchmark datasets with the cost of increased computational demands. In this talk we demonstrate that neural network architectures can be automatically generated, tailored for a specific application, with dual objectives: accuracy of prediction and speed of prediction. Using MENNDL--an HPC-enabled software stack for neural architecture search--we generate a neural network with comparable accuracy to state-of-the-art networks on a cancer pathology dataset that is also 16x faster at inference. The speedup in inference is necessary because of the volume and velocity of cancer pathology data; specifically, the previous state-of-the-art networks are too slow for individual researchers without access to HPC systems to keep pace with the rate of data generation. Our new model enables researchers with modest computational resources to analyze newly generated data faster than it is collected.  Back
 
Topics:
HPC and Supercomputing
Type:
Talk
Event:
Supercomputing
Year:
2019
Session ID:
SC1934
Streaming:
Share:
 
Abstract:
There has been a surge of success in using deep learning as it has provided a new state of the art for a variety of domains. While these models learn their parameters through data-driven methods, model selection through hyper-parameter choices remains a tedious and highly intuition-driven task. We've developed two approaches to address this problem. Multi-node evolutionary neural networks for deep learning (MENNDL) is an evolutionary approach to performing this search. MENNDL is capable of evolving not only the numeric hyper-parameters, but is also capable of evolving the arrangement of layers within the network. The second approach is implemented using Apache Spark at scale on Titan. The technique we present is an improvement over hyper-parameter sweeps because we don't require assumptions about independence of parameters and is more computationally feasible than grid-search.
There has been a surge of success in using deep learning as it has provided a new state of the art for a variety of domains. While these models learn their parameters through data-driven methods, model selection through hyper-parameter choices remains a tedious and highly intuition-driven task. We've developed two approaches to address this problem. Multi-node evolutionary neural networks for deep learning (MENNDL) is an evolutionary approach to performing this search. MENNDL is capable of evolving not only the numeric hyper-parameters, but is also capable of evolving the arrangement of layers within the network. The second approach is implemented using Apache Spark at scale on Titan. The technique we present is an improvement over hyper-parameter sweeps because we don't require assumptions about independence of parameters and is more computationally feasible than grid-search.  Back
 
Topics:
HPC and Supercomputing, Artificial Intelligence and Deep Learning
Type:
Talk
Event:
GTC Washington D.C.
Year:
2017
Session ID:
DC7200
Download:
Share:
 
Abstract:
There has been a surge of success in using deep learning as it has provided a new state of the art for a variety of domains. While these models learn their parameters through data-driven methods, model selection through hyper-parameter choices remains a tedious and highly intuition-driven task. We''ve developed two approaches to address this problem. Multi-node evolutionary neural networks for deep learning (MENNDL) is an evolutionary approach to performing this search. MENNDL is capable of evolving not only the numeric hyper-parameters, but is also capable of evolving the arrangement of layers within the network. The second approach is implemented using Apache Spark at scale on Titan. The technique we present is an improvement over hyper-parameter sweeps because we don''t require assumptions about independence of parameters and is more computationally feasible than grid-search.
There has been a surge of success in using deep learning as it has provided a new state of the art for a variety of domains. While these models learn their parameters through data-driven methods, model selection through hyper-parameter choices remains a tedious and highly intuition-driven task. We''ve developed two approaches to address this problem. Multi-node evolutionary neural networks for deep learning (MENNDL) is an evolutionary approach to performing this search. MENNDL is capable of evolving not only the numeric hyper-parameters, but is also capable of evolving the arrangement of layers within the network. The second approach is implemented using Apache Spark at scale on Titan. The technique we present is an improvement over hyper-parameter sweeps because we don''t require assumptions about independence of parameters and is more computationally feasible than grid-search.  Back
 
Topics:
HPC and Supercomputing
Type:
Talk
Event:
SIGGRAPH
Year:
2017
Session ID:
SC1704
Download:
Share:
 
Abstract:
We'll present graphs as powerful tools when analyzing complex relationships between entities. We'll share how many structures commonly found in computer science, like social networks, computer networks, and the world wide web, can be modeled as graphs. Since many of the real graphs are very large and complex, the associated analysis algorithms must be very efficient and highly parallel. We present two implementations of a key graph-based analysis such as the triangle enumeration for two different parallel paradigms: GPU programming and Apache Spark. We'll reveal the performance of the two different implementations for the different paradigms as the characteristics of the graph change.
We'll present graphs as powerful tools when analyzing complex relationships between entities. We'll share how many structures commonly found in computer science, like social networks, computer networks, and the world wide web, can be modeled as graphs. Since many of the real graphs are very large and complex, the associated analysis algorithms must be very efficient and highly parallel. We present two implementations of a key graph-based analysis such as the triangle enumeration for two different parallel paradigms: GPU programming and Apache Spark. We'll reveal the performance of the two different implementations for the different paradigms as the characteristics of the graph change.  Back
 
Topics:
Algorithms & Numerical Techniques, Tools & Libraries, Big Data Analytics
Type:
Talk
Event:
GTC Silicon Valley
Year:
2016
Session ID:
S6424
Streaming:
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