Kwabena Boahen 2006 human brain at rest dissipates about 20 watts 10^12 neurons x 10^3 synapes x 10 spikes/s = 10^16 events/s = 10^-15 J/op Blue Gene: 10^15 ops/s @ 10^6 watts = 10^-9 J/op must make up six orders of magnitude! Carver Mead: "Analog VLSI" -- use transistors as analog devices 1pA per ion channel means avg. neuron has 100 channels open at a time terrible noise levels (1/sqrt(n) = 10%), compare to 60dB S/N ~ 1M:1 Neurogrid: 1M neurons with 6B connections @ < 1 watt total by 2008 Dharmendra Modha 2007 human brain: about 10^15 synapses assume 1 Hz. average firing to get 2 * 10^15 synaptic events per second if each event represents 350 OPS, then the humans compute at 350 PFLOPS mice at 45TFLOPS -- within reach of Blue Gene mouse neurons via Braitenberg & Schuz 1998: 16M neurons, 80% of which are excitatory 8K synapses/neuron axonal delays of 1-20ms 0.09 local probability of connection uses 4-parameter spiking neuron model by Izhikevich 2004 uses STDP synapse model by Song, Miller, Abbott 2000 cat visual cortex connectivity via Binzeggar, Douglas, Martin 2004: 77% excitatory -> excitatory 11% inhibitory -> excitatory 10% excitatory -> inhibitory 2% inhibitory -> inhibitory macaque white matter connectivity via cocomac.org (meta-analysis) currently performance is non-realtime: 1 sec of model in 10 real seconds Henry Markram 2006 1M cortical colums in human brain, 100K neurons per column about 20 anatomical types of cortical neuron (9 inhibitory 10 excitatory) each anatomical type can produce different electrical behaviors electrical behavior depends on mixture of ion channels expressed by cell several million ion channels can be expressed per neuron overall column "recipe" is remarkably similar from column to column any neuron can connect to any other by just growing a synapse each neuron connects to approximately 10% of other neurons in the column connection morphology is characteristic of pre-post neuron types first simulation is one column with 10K neurons uses Blue Gene, one processor per neuron speculation: not enough information in spikes to explain qualia 1 petaflop general-purpose computer about enough for mouse brain Almut Schuz 2007 human cortex: ~ 10^10 neurons, ~ 10^8 input neurons most of the cortex connects to itself 85% of the neurons are pyramidal, connecting both locally and distantly probability of connection falls off with distance connection probability per distance curve looks the same in all layers a pair of cells prefers to share no more than one synapse about 90% of cortical synapses are excitatory results suggest Hebbian cell assemblies connectivity in other brain areas isn't as supportive of cell assemblies Ed Callaway 2007 cortical circuits display "fine-scale and cell-type specificity" ~ 12 types of inhibitory cells in cortex, accounting for 20% of all cells inhibitory types fall into two groups: fast-spiking and non-fast-spiking method: record from brain slices bathed in caged glutamate UV activates glutamate -- more accurate than electrical stimulation about 24 neurons respond to each stimulus results: neurons connect preferentially based on cell type only 1 in 5 neighboring pairs of neurons are directly connected relatively independent subnetworks exist within cortical columns on the horizon: molecular and genetic methods in vivo allatostatin & GFP transgenic mice -- turn specific cell types on and off modified rabies -- light up all neurons one hop upstream of target cell