From Biohack

Henry Markram assembled a team of programmers and scientists to put together 400 software packages in computational neuroscience to simulate the rat cortical column. There was an article here a while back, over at Blue Brain project. What were these software packages? Can we get a list of all of these packages?

Markram also did research on autism re: intense world syndrome.

Notes on Markram's talk

The talk can be found here.

• 200 ion channels (in the neocortex)
  • mostly due to combinatorial effects
  • database of different types of ion channels
    • which ones do we need to use to model the different types of behaviors?
    • ion channels are inserted in different locations, and this adds new functions to the neurons depending on location
      • exactly where do we put the ion channels? Several million ion channels per neuron.
      • freeze fracture scanning microscopy. Locate single channels, single molecule to constrain the model and to say certain distributions and so on.
• how many types of neuron?
• patch clamp robot to extract the cytoplasm
  • figure out which genes are on in that cell
    • this can be used to discover ion channels

• The recipe of the neocortex - of cell types used to build up the neural circuit
• Microcolumn is conserved over evolutionary history, though subtle recipe might change, the core stays the same to give you a column. This is the template.
• It is not possible - the cells are not doing specific targetting, but rather emergence and so on.
  • biases on neuronal branching / targetting - Jerry predicted no bias
    • every axon attaches every cell, it's an all-to-all circuit that is ready and in place for a specific functional circuit. It's just structurally in position. It's only going to transmit information to about 10% of them, and those 10%, all that it has to do is grow a synapse.

Blue Builder is a neuronal circuit compiler; it compiles it from a database of known ion channels and stereotyped structures of the microcircuits.

• Axodendritic collision detection
  • they touch 100 million times, and you have 8000 processes running over Blue Gene - algorithms to do this faster? We can do the first iteration, but we need to do 10k iterations while we spin and jitter the neurons, so that we can check whether or not the locations are matching the lock-and-key principles that we've devised before.
• positioning data from statistical data from experiments (Jerry/Jerri) -- statistics of connectivity

• synaptic physiology map of pre- and post- neuron connections, we need to model the right amount of current at the right time (this has been derived)
  • now go to the circuit builder - who are going to have synapses with it? We can now assign all of the synaptic properties which can be put into an algorithm that will act as a synapse. Through this model, we can capture all of the different types of dynamics. 30 Hz simulation of the source cell. Whole spectrum of different types of dynamics. synaptic injection

• spike-timing dependent plasticity algorithm (there are many others)
• 20 million nonlinear learning synapses -- you need algorithms to align them; experiments to derive these algorithms

• p-NEURON
• MPI messaging as axons ... one neuron per software. Two software packages that allow these simulations. The problem is that it produces 1 TB of data in 1 sec of simulation. So we needed another supercomputer because we want to decide which terabyte to keep for analysis - no need to spend half a year to analyze nonsense. So, visualization to immediately assess whether something is interesting or not. SGI and shared memory. 90% discount on a supercomputer with 300 GB shared memory. Media center. 3D representations. You can sit inside the neocortical column. You want to build the puzzle and sit inside and actually see it works - not sure if it's sci/pleasure, but it is a lot of fun.

• version 10 by the end of the year - 10k neurons that are more or less functional.
• steps of biological refinement -- this is a huge chart. GET THIS CHART. (38:25)
• Action Potential Paradigm v. Synaptic Paradigm
• soma-centric to dendro-centric
  • the data is represented in analog form on dendrites, and the perceptions are animated by the spatiotemporal spikes. The thoughts occur on the dendrites themselves. We can get dendrite maps (MEG, fMRI), they go across dendrites.

• painting 3D pictures on to the dendrites ---> This is a compressed way of saying it. Eh. Not really -- it's more about programming the neurons for thoughts, or something.
• analog scenes across dendritic space
• voxalization :-) 2.5D hurray.
• block of dendritic tissue
• a whole scene within a region of dendrites, an ecology within a microcircuit
• How does the circuit restructure itself?
  • We could particularly insert intelligence via specific stimulation to make it build specific structures.
  • distance between synapses - spillover from one input is almost impossible to avoid, to another one. Within 200 nm there's another synapse that is related to a different neuron. What the brain is trying to do is control the voltage within a certain voxel compartment - not so much a particular neuron. This would be an attempt to control pain/energy/action. Electromagnetic dendritic objects. We do not see the world. What we do is we use any clue that our senses can provide us is to build a virtual analog model in our brain; the world we see is the world we build. (51:05). "So it may just be, what the brain is trying to do is control the voltage within a certain voxel compartment. So if we summarize, let's look at the world as electromagnetic dendritic object."
• The World as an Electromagnetic Dendritic Object - see electromagnetic world object
  • distributed dendritic object
  • circuitry provides the rules for controlling the synapticity
  • cognition - to minimize/optimize adaptation of those circuits
  • different brains can learn to build the same models.
  • Similar rules to build the same rule, so that we can see the same world and interact. But animals with different brains, and more than likely, they build a different world.

The world as an Electromagnetic Dendritic Object - a virtual environment

We do not see the world, we use any clue that our senses can provide us to build a virtual analog model of the world on the dendrites

• Neurons learn to contribute to the circuit to build a distributed dendritic object
• The circuitry provides the rules to build and animate dendritic objects
  • animation is the spikes, aggregation
• Advanced cognition is the ability to run a simulation of the model into the future to optimize adaptation
• Different brains can learn to build a similar model
• Animals see the world differently because they build it differently, based on major differences in the brains capabilities - lots of genetics

Spikes as animators of perceptions

• Spikes emitted on top of dendritic perceptual waves
• Minimize the number of spikes used as we learn to transfer just the required information

We use spikes to transfer the minimum information required to change or transfer perceptions

• Perceptions before the spike
• Resting potentials contribute to perceptions

• spikes transmit the minimal information to change perceptions; small updates - spikes are the carriers of RSS? They could probably transmit complete perceptions. Spikes on top of dendritic-perceptual-waves, that spikes are perhaps the byproduct that are driving and animating our perception; minimize the number of spikes to transfer just the required information, the key thing in learning is that you actually start spiking less and less and less, and you only want to become more efficient to transmit more information. Perception may be there without a spike. And that you might even have perceptions / resting potentials there.
• Analog perception. interaction. There is not enough information in the spikes to recreate the world; it's a challenge.

• Igan Segev
• Philip Gooddman
• ONR in Israel
• lock and key rules of connectivity
  • statistics of branch order between neurons of specific type
  • branch order specifity
  • probabilities to decide the connectivity ... but the refining - molecular causes of the rewiring in plasticity, turning this into an algorithm. Start with an unconnected circuit; show it a stimulus set; and it will begin to rewire; we will not use probabilities of connectivity, there will be very specific wiring. If the dendritic object theory is something that one can explore further, one demand would be is that every synapse must have a very specific reason for being there - it will be contributing very specifically to the object being formed.
• multiplicity of circuits?
  • generalizing -- how do you generalize the different types of circuits? How do you get a specific instantiation of circuits? Powerful genetic rules.

• synaptic precision as the differences between species re: their intelligence?

• IBM may be designing specialized circuitry to compute the neurophysics without microprocessors. The microprocessors aren't going to be useful, really. But the problem is that if you do silicon fabrication, how are you going to update the circuit equations without manufacturing more? :-(

axil resistance - propagation too far? bad, diagnostics.

Javier Defilpe