Synthetic Cognition at Los Alamos

Last updated: Aug 14, 2012

The synthetic cognition group at Los Alamos National Laboratory is building supercomputer simulations of the human visual system. They are using IBM Roadrunner - a petaflop supercomputer which was the world's fastest when built in 2008 - as well as various other GPU-accelerated multi-teraflop machines. The simulations are not highly biologically detailed, but nor are they as trivial as standard artificial neural networks. Instead they strike a balance at a level of complexity thought to be sufficient to reproduce the function and behaviour of natural nervous systems.

The simulations are built using custom software written in C, C++, and Python. Running at a peak of 1.14 PFLOPS they have simulated the primate visual system including the retina, LGN, and visual cortex areas V1, V2, and V4. The models are trained to visually recognise objects such as enemy vehicles on a battlefield - a possible application being their use in military UAVs. The research aims to prepare for the time when supercomputers become sufficiently powerful to simulate an entire human brain.

Contents    (hide)

Research details

The goal of this resarch is to understand the computational principles of high-level sensory processing and cognition in the human brain. This is to be achieved by using a petascale supercomputer to create a synthetic cognition system that emulates the functional neural architecture of the primate visual cortex.

Complete physiological realism is not the goal. It is attempted only to the extent that is needed for understanding neural computation and for solving complex information processing tasks. The cortical models fall between the traditional (and over-simplified) artificial neural networks and biologically-inspired cellular and molecular descriptions.

A petascale supercomputer is considered sufficiently powerful based on the following calculation:

  • There are ~10 billion neurons in the human visual cortex
  • Each neuron recieves ~10,000 synaptic inputs
  • Each synapse requires ~10 FLOPS based on a 1 Hz firing rate
  • Thus: ~10 G neurons x 10 K synapses x 10 FLOPS = 1 PFLOP

The IBM Roadrunner supercomputer at LANL fullfills this requirement with its 1.7 PFLOP peak capability. The neural network model achieved computation speeds of 1.14 PFLOPS in 2008 - a world speed record for scientific computation at the time, and the first to break the petaflop barrier. In the future, when exascale supercomputers become available, it should be possible to build simulations that match the scale of the complete human brain.

The synthetic cognition team have developed a completely new methodology for neural computation and a novel software architecture for emulating cortical columns. It is not based on Michael Hines' NEURON, as is the case with many other biologically-inspired simulations. The model enables petascale simulation of the mammalian visual system, including the retina, LGN, and the visual cortex ventral pathway V1 -> V2 -> V4 -> IT. It uses a hierarchical feed-forward architecture in the family of Neocognitron and HMAX models.

The model implements the following features:
  • Spiking dynamics - Conductance-based leaky integrate and fire neurons. A neuron spikes when the membrane potential exceeds a threshold. Excitatory and inhibitory synaptic input modifies the associated conductance.

  • Lateral connectivity and feedback loops - 80% of cortical synapses are lateral connections and feedback connections from higher visual cortex areas.

  • Spike-timing-dependent plasticity (STDP) - Causal pairing of a pre- and post-synaptic spikes (i.e. input precedes output) strengthens synapses. Acausal pairing decreases it.

News / current status

Most recent research paper, published September 2011: Petascale visual cortex models and object detection

IBM Roadrunner

Roadrunner is a supercomputer built by IBM at the Los Alamos National Laboratory in New Mexico, USA. It was the world's fastest supercomputer when it first went live in mid-2008. As of November 2011 it is ranked the world's 10th fastest. It has a peak performance of 1.7 PFLOPS.

The supercomputer is one of a kind, custom built from off-the-shelf parts at a cost of $133 million. It occupies ~560 m2 of floor space. The hybrid design uses two different types of processors, as follows:

  • 12,960 x IBM PowerXCell 8i CPUs (116,640 cores)
  • 6,912 x AMD Opteron dual-core processors (13,824 cores)
  • Total: 130,464 cores (both computing and operation nodes)

Roadrunner is used by the DOE to analyse the safety of the USA's aging arsenal of nuclear weapons. It is also by the science, financial, automotive, and aerospace industries.

Roadrunner supercomputer

Los Alamos

Los Alamos National Laboratory (LANL) is a US Department of Energy (DOE) national laboratory, located in Los Alamos, New Mexico. It is one of the largest science and technology institutions in the world. The lab conducts multidisciplinary research in fields such as national security, space exploration, renewable energy, medicine, nanotechnology, and supercomputing. It receives $2.2 billion in annual funding and employs ~9,000 staff and ~700 students.

Funding

Funding for this research comes from a $6 million R&D investment made in ~2009 by DOE/LANL, NSF, and DARPA.

The National Science Foundation (NSF) awarded $1,168,982 (award no. NSF-OCI-0749348) in May 2008 with an estimated expiry date of April 2012.

Details of the DOE/LANL and DARPA awards are unknown.

People involved

Craig Rasmussen   Linkedin profile
Technical staff member 
  Cristina Rinauldo  
Student research assistant 
  Cyrus Omar   Google profile Linkedin profile Homepage
PhD student 
Garrett Kenyon   Google profile Homepage
Co-principal investigator 
  Ilya Nemenman  
Collaborator 
  John George   Linkedin profile
Co-principal investigator 
Karissa Sanbonmatsu   Linkedin profile Homepage
 
  Luís Bettencourt   Linkedin profile Homepage
Principle investigator 
  Marian Anghel   Linkedin profile
Technical staff member 
Melanie Mitchell   Linkedin profile Wikipedia page Homepage
Collaborator 
  Michael Ham   Google profile Linkedin profile Homepage
Postdoc research assistant 
  Mick Thomure   Google profile Linkedin profile Homepage
PhD student 
Peter Loxley   Linkedin profile Homepage
Postdoc research assistant 
  Shawn Barr   Linkedin profile
Student research assistant 
  Steven Brumby   Google profile Linkedin profile Homepage
 
Tsvi Achler   Google profile Linkedin profile Homepage
Postdoc research assistant 
  Vadas Gintautas   Google profile Linkedin profile Homepage
Collaborator 
  Will Landecker   Homepage
PhD student 
Zhengping Ji   Google profile Linkedin profile Homepage
Postdoc research assistant 
  

Research papers