Scribd
Upload
140 views16 pages

Brumby GTC2010 Slides

The document describes research at Los Alamos National Laboratory to develop large-scale, real-time functional models of mammalian visual cortex using high-performance computing systems like the Roadrunner supercomputer. The goals are to explore new neuroscience-inspired artificial neural networks for robust feature and object extraction from video at the scale of the human visual cortex, processing 1080p HD video in real-time. The technical approach involves building a full-scale model of the visual cortex using models of neurons, synapses, color processing, feedback and learning rules, and implementing it on petaflop and GPU-accelerated supercomputers.

Uploaded by

Mossad News
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
140 views16 pages

Brumby GTC2010 Slides

The document describes research at Los Alamos National Laboratory to develop large-scale, real-time functional models of mammalian visual cortex using high-performance computing systems like the Roadrunner supercomputer. The goals are to explore new neuroscience-inspired artificial neural networks for robust feature and object extraction from video at the scale of the human visual cortex, processing 1080p HD video in real-time. The technical approach involves building a full-scale model of the visual cortex using models of neurons, synapses, color processing, feedback and learning rules, and implementing it on petaflop and GPU-accelerated supercomputers.

Uploaded by

Mossad News
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 16

Visual Cortex on a Chip Large-scale, real-time functional models of mammalian visual cortex on a GPGPU

Steven P. Brumby (brumby@lanl.gov), " Amy E. Galbraith, Michael Ham, Garrett Kenyon, and John S. George ! Los Alamos National Laboratory, Los Alamos, New Mexico, USA ! GPU Technology Conference 2010, San Jose CA!

http://synthetic-cognition.lanl.gov"
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

Large-scale, real-time functional models of " mammalian visual cortex on a GPGPU!


Project Goals Concept
Science Goal: Explore new neuroscience-inspired artificial neural networks using unique petascale computing resources at LANL (Roadrunner) for robust, human-like feature and object extraction from video and imagery datasets Performance Goal: Full-scale, real-time model of human visual cortex ~ 1080p HD video at >10 fps.

Technical Approach
1. Full-scale model of visual cortex using: synaptic plasticity / machine learning, models of color processing, feed-back/lateral connectivity, and spiking dynamical neurons. 2. Unique petaflop (Roadrunner/Cell) and multi-teraflop single-rack GPGPU-accelerated high performance computing systems. 3. A $6M R&D investment by DOE/LANL LDRD-DR, NSF, and DARPA. http://synthetic-cognition.lanl.gov
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

Full-scale simulation of human vision is a petascale computing challenge, and is now feasible with Roadrunner.

UNCLASSIFIED LA-UR-10-05430!

LANL Roadrunner IBM Cell processor


NVIDIA Fermi GPGPU Xbox 360 IBM Xenon

Hans Moravec, When will computer hardware match the human brain?, J. Evolution &Technology, 1998.

Computing hardware for large-scale simulations of human cortex exists now: With petascale supercomputers, such as LANLs Roadrunner, large-scale simulations can be run at real-time on full-scale visual datasets (~months of HD video). GPGPU enables experiments at small-mammal scale using machines costing a few k$.

Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

!!Mouse cortex: !!Cat cortex:

4,000,000 neurons 300,000,000 neurons

!!Monkey cortex: 6,000,000,000 neurons !!Human cortex: 11,000,000,000 neurons (~100G neurons whole brain) !!~10,000 synapses (connections) per neuron !!130 types of neurons in human brain

Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

!"#$%&'()*+,"'()*+-./"01)
9B&1#7A)5/6+(1) C"/11@+"ADE'"F+7-+")!GH) :/AA&/)5IH)4/6+(1) <+E,7)J++F)7+-./"01) #K&L)

(+>+()/?)@&/(/A&%'()6+-'&()!"

Spiking Dynamics
2&/%3+4&%'()5/6+(1) 2(,+)2"'&7)89:;<=)

Feedback Loops

Synaptic Plasticity
#K&L) #K&L)

Next generation Artificial Neural Networks: Goal is functional models of visual cortex based on recent work in neuroscience.
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

WHERE PATHWAY

V1 V2, V4 IT WHAT PATHWAY


Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

E.M. Izhikevich & G.M. Edelman, (2008) Proc. National Acad. Sci. 105:3593-3598 Large-Scale Model of Mammalian Thalamocortical Systems,

Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

Processing in Retina Local contrast equalization, for each patch set mean to zero and local Euclidean norm to 1. Processing in S-cells Radial Basis Functions with standard oriented Gabor neuron weight vectors [3-5].
2& # 1 sRBF ( r , wG ) = exp % ! " ( r ! wG ) ( $ 2 '

' 1 2 ' 2- x! * 2 * wG = w (! , " , # , $, % ) = exp ) & x! + " y! cos ) + %, , ( 2# + ( $ +

Processing in C-cells MAX Function over S-cell receptive fields.


cMAX ( s ) = max ({ si })
i !I j

Standard model of visual cortex


LANLs PANN code is a high performance implementation in C++/C and Python of a Neocognitron model [Fukushima; Poggio et al.].
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

Blu-Ray 1080p @ 24 fps Eye ~2 Mpixels @ 10 fps SDTV 480p @ 30 fps

WebCam 240p @ 15 fps

LANL supercomputing assets have allowed us to reach full-scale processing at real-time video rates. Codes can also run on GPU or conventional multi-core machines.
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

LANL codes exploit conventional multi-core/ multi-cpu clusters, vector processing on cpu cores (SSE), and cluster on a chip hardware (CELL, GPGPU).
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

Brumby et al., IEEE Proc. AIPR 2009.


Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

HebbOha learning rule: !w = " y ( r , w ) ( r # w ) Visual cortex models are self-organizing using neuroscience-inspired learning rules.
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430 ! UNCLAS SIFIED

RF 5x5

RF 7x7

Left top: Sorted histogram of number of activations of learned 5x5 prototypes in response to different numbers of retinal patches (cf., grey dashed line = random imprinting): blue=1,800, aqua=18,000, green=180,000, yellow=1,800,000, red = 18,000,000. Left middle: Sorted histogram of number of activations of learned 7x7 prototypes in response retinal patches (cf., grey dashed line = random imprinting): blue=1,800, aqua=18,000, green=180,000, red = 1,800,000. Left bottom: KullbackLeibler divergence of activation distributions for 5x5 (red points) and 7x7 (blue points), relative to the final (red curves) in upper and middle panels.

Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

PANN!

Receiver-Operator Curves Training Solid Line Testing Dashed Line

Visual cortex models can be applied to non-traditional computer vision tasks, such as vehicle identification in Brumby et al., IEEE Proc. AIPR 2009. satellite/aerial imagery and video.
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

UNCLASSIFIED LA-UR-10-05430!

Large-scale, real-time functional models of " mammalian visual cortex on a GPGPU!


Project Goals Concept
Science Goal: Explore new neuroscience-inspired artificial neural networks using unique petascale computing resources at LANL (Roadrunner) for robust, human-like feature and object extraction from video and imagery datasets Performance Goal: Full-scale, real-time model of human visual cortex ~ 1080p HD video at >10 fps.

Technical Approach
1. Full-scale model of visual cortex using: synaptic plasticity / machine learning, models of color processing, feed-back/lateral connectivity, and spiking dynamical neurons. 2. Unique petaflop (Roadrunner/Cell) and multi-teraflop single-rack GPGPU-accelerated high performance computing systems. 3. A $6M R&D investment by DOE/LANL LDRD-DR, NSF, and DARPA. http://synthetic-cognition.lanl.gov
Operated by Los Alamos National Security, LLC for the U.S. Department of Energys NNSA

Full-scale simulation of human vision is a petascale computing challenge, and is now feasible with Roadrunner.

UNCLASSIFIED LA-UR-10-05430!

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy