Can somebody like davidad tell me if this is actually a significant result? I sorta assume it's not.
I've asked my mom to tell me more; she's a neuroscientist. She hadn't seen the paper or heard about it - she's now semi-retired so not necessarily staying on top of all the latest. I've sent her the paper (and the PR for good measure), she'll be taking a look and also asking around if her colleagues have had time to form an impression of how significant this is.
The linked ScienceDaily article is one of however many clones of a press release written by a freelance science journalist, about a paper that as far as I can tell is still embargoed, but will be out on PNAS within a week. It has the typical breathless style and vagueness of journalistic reporting on science.
As best I can make out, the PR is claiming that someone can "make a near perfect prediction" of something that it simultaneously claims results from synapses that "randomly bump into each other". Hm, okay.
I'm not "somebody like davidad" but I would wait until the actual article is out before getting either excited or disappointed; by default I'd assume no more than the most tenuous of connections between press releases and actual scientific content.
(But I'm tempted to ask you, "What do you know - and why?")
about a paper that as far as I can tell is still embargoed, but will be out on PNAS within a week.
Abstract:
It is well-established that synapse formation involves highly selective chemospecific mechanisms, but how neuron arbors are positioned before synapse formation remains unclear. Using 3D reconstructions of 298 neocortical cells of different types (including nest basket, small basket, large basket, bitufted, pyramidal, and Martinotti cells), we constructed a structural model of a cortical microcircuit, in which cells of different types were independently and randomly placed. We compared the positions of physical appositions resulting from the incidental overlap of axonal and dendritic arbors in the model (statistical structural connectivity) with the positions of putative functional synapses (functional synaptic connectivity) in 90 synaptic connections reconstructed from cortical slice preparations. Overall, we found that statistical connectivity predicted an average of 74 ± 2.7% (mean ± SEM) synapse location distributions for nine types of cortical connections. This finding suggests that chemospecific attractive and repulsive mechanisms generally do not result in pairwise-specific connectivity. In some cases, however, the predicted distributions do not match precisely, indicating that chemospecific steering and aligning of the arbors may occur for some types of connections. This finding suggests that random alignment of axonal and dendritic arbors provides a sufficient foundation for specific functional connectivity to emerge in local neural microcircuits.
Thanks!
I'd like to know where "accuracy ranging from 75 percent to 95 percent" in the press release came from.
Looking at the paper, I'm guessing it's the histogram intersection (HI) measurement, which actually ranged from 50% to 95% with an average of 75%, so that's one glaring error I could find in the first few minutes of looking at a paper in a field that I know next to nothing about. (A further guess is that the journalist thought 75% accuracy didn't sound all that impressive, and so went on a hunt for a grander number, hitting on the 95%.)
Maybe I'm all wet, but what I'm taking from the paper is that it's actually a null result? As in, some other researchers had theorized that the fine structure of synaptic connections was somehow "guided" by chemical signals, but the simulation in fact suggests that "randomly bumping into each other" is closer to the truth. This does seem to (maybe decisively) settle a debate which the authors say has gone on for decades; I'm wondering if the term "breakthrough" is warranted though.
As best I can make out, the PR is claiming that someone can "make a near perfect prediction" of something that it simultaneously claims results from synapses that "randomly bump into each other". Hm, okay.
It's like describing as "a near perfect prediction", the prediction that a tossed coin will come down heads with 50% probability. The 50% figure may be really precise, but that still tells you nothing about the next coin toss. The press release seems to confuse knowledge about a distribution with knowledge about individuals drawn from it.
As I understand it, what the paper, and Henry Markram as quoted in the press release, are actually saying is that their experiments show a distribution of connections that one would get by things randomly bumping into each other. This implies the predictions that there is no other organising mechanism in play yet to be discovered, and that to build an artificial network capable of the same functionality, one can build the connections in the same random manner. Or as they put it in the title, "Statistical connectivity provides a sufficient foundation for specific functional connectivity".
Note that the title is a prediction from what they have observed, not the observation itself. Functional comparisions of real neural tissue and artificial networks constructed on these lines have not yet been done.
I sorta assume it's not.
Based on the journalistic style alone? The important indicators point at it being very significant: PNAS published, Blue Brain Project (which is a serious effort I've semi-followed for a long time), the direct quote from the PI ("This is a major breakthrough ...")
That is the place we'd expect such breakthroughs to originate from, it's already been peer-reviewed, why so skeptical?
Personally, I would assume that it would be quite difficult for a random distribution of neurons to form the exact same network of synapses 85% of the time. Two random graphs with the same number of vertices have a very low probability of being isomorphic, but the spanning trees of those random graphs would have a trivial isomorphism and I assume neural networks formed by synapses randomly bumping into each other are more like the union of a few random spanning trees (lots of local connections, fewer long ones) than fully random graphs. Still, it's probably not 85% or 95% likely for there to be an isomorphism. However, if the shape of the neurons is what determines the properties of the synapses then it may be that how the synapses grow isn't as important as how the types of neurons are initially distributed. I haven't read the paper either, but if the summary is correct they distributed the different types of neurons "randomly", which of course doesn't say if there was a complex probability density over the 3D space or if it was, e.g., uniform probability for each neural shape.
From http://www.sciencedaily.com/releases/2012/09/120917152043.htm
Could this be a tiny step towards an AGI?
'Blue Brain' Project Accurately Predicts Connections Between Neurons
One of the greatest challenges in neuroscience is to identify the map of synaptic connections between neurons. Called the "connectome," it is the holy grail that will explain how information flows in the brain. In a landmark paper, published the week of 17th of September in the Proceedings of the National Academy of Sciences, the EPFL's Blue Brain Project (BBP) has identified key principles that determine synapse-scale connectivity by virtually reconstructing a cortical microcircuit and comparing it to a mammalian sample. These principles now make it possible to predict the locations of synapses in the neocortex.
"This is a major breakthrough, because it would otherwise take decades, if not centuries, to map the location of each synapse in the brain and it also makes it so much easier now to build accurate models," says Henry Markram, head of the BBP.
A longstanding neuroscientific mystery has been whether all the neurons grow independently and just take what they get as their branches bump into each other, or are the branches of each neuron specifically guided by chemical signals to find all its target. To solve the mystery, researchers looked in a virtual reconstruction of a cortical microcircuit to see where the branches bumped into each other. To their great surprise, they found that the locations on the model matched that of synapses found in the equivalent real-brain circuit with an accuracy ranging from 75 percent to 95 percent.
This means that neurons grow as independently of each other as physically possible and mostly form synapses at the locations where they randomly bump into each other. A few exceptions were also discovered pointing out special cases where signals are used by neurons to change the statistical connectivity. By taking these exceptions into account, the Blue Brain team can now make a near perfect prediction of the locations of all the synapses formed inside the circuit.
Virtual Reconstruction
The goal of the BBP is to integrate knowledge from all the specialized branches of neuroscience, to derive from it the fundamental principles that govern brain structure and function, and ultimately, to reconstruct the brains of different species -- including the human brain -- in silico. The current paper provides yet another proof-of-concept for the approach, by demonstrating for the first time that the distribution of synapses or neuronal connections in the mammalian cortex can, to a large extent, be predicted.
To achieve these results, a team from the Blue Brain Project set about virtually reconstructing a cortical microcircuit based on unparalleled data about the geometrical and electrical properties of neurons -- data from over nearly 20 years of painstaking experimentation on slices of living brain tissue. Each neuron in the circuit was reconstructed into a 3D model on a powerful Blue Gene supercomputer. About 10,000 of virtual neurons were packed into a 3D space in random positions according to the density and ratio of morphological types found in corresponding living tissue. The researchers then compared the model back to an equivalent brain circuit from a real mammalian brain.
A Major Step Towards Accurate Models of the Brain
This discovery also explains why the brain can withstand damage and indicates that the positions of synapses in all brains of the same species are more similar than different. "Positioning synapses in this way is very robust," says computational neuroscientist and first author Sean Hill, "We could vary density, position, orientation, and none of that changed the distribution of positions of the synapses."
They went on to discover that the synapses positions are only robust as long as the morphology of each neuron is slightly different from each other, explaining another mystery in the brain -- why neurons are not all identical in shape. "It's the diversity in the morphology of neurons that makes brain circuits of a particular species basically the same and highly robust," says Hill.
Overall this work represents a major acceleration in the ability to construct detailed models of the nervous system. The results provide important insights into the basic principles that govern the wiring of the nervous system, throwing light on how robust cortical circuits are constructed from highly diverse populations of neurons -- an essential step towards understanding how the brain functions. They also underscore the value of the BBP's constructivist approach. "Although systematically integrating data across a wide range of scales is slow and painstaking, it allows us to derive fundamental principles of brain structure and hence function," explains Hill.