Home

Large Scale Computing Models

Spiking Models

People

Publications

Research in the News

Job Opportunities

PetaVision

Petavision is a functional model of visual cortex that enables petascale simulation of mammalian vision. Petavision is based on a biologically-inspired hierarchical feed-forward architecture originally proposed by Kunihiko Fukushima ("Neocognitron") [1], as later developed by Tomaso Poggio, et al., [2], which we have modified in several key ways:


Spiking Dynamics: Conductance-based leaky integrate and fire neurons. The basic equation for the membrane potential is given by

where V is the membrane potential, g_exc and g_inb are the excitatory and inhibitory conductances, and V_exc and V_inh are the corresponding reversal potentials. A neuron spikes when the membrane potential exceeds a threshold V_thresh which is adaptive. Excitatory and inhibitory synaptic input modifies the associated conductance.


Feedback: Neurons in visual cortex receive 1000's of synaptic connections, the majority (~80%) of which are lateral connections to neighbors and feedback connections from higher visual cortex areas. These lateral and feedback connections are thought to convey contextual information that helps to resolve ambiguity in local features and may contribute to the formation and/or enhancement of functional groups.


Spike Timing Dependent Plasticity (STDP) (figure from [3]): Connections between neurons in the visual cortex are thought to be modified according to an STDP rule in which a causal pairing of a pre- and post-synaptic spike (i.e. input precedes output) increases the connection strength whereas an acausal pairing (i.e. output precedes input) decreases the connection strength. Petavision seeks to simulate the developmental process via which the visual cortex wires itself in response to visual experience.

Initial Results

Simulations of a leaky integrate-and-fire neural model with cocircular excitation running on the Roadrunner supercomputer achieved computation speeds of 1.14 Petaflops.

References

[1] K. Fukushima, Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position, Biological Cybernetics, 36(4), pp. 193-202 (April 1980).

[2] T. Serre, A. Oliva and T. Poggio. A feedforward architecture accounts for rapid categorization. Proceedings of the National Academy of Science, 104(15), pp. 6424-6429, April 2007.

[3] Y Dan and MM Poo, Spike Timing-Dependent Plasticity: From Synapse to Perception, Physiol. Rev. 86: 1033-1048, 2006; doi:10.1152/physrev.00030.2005