Can we also map the synaptic strength with this method? If not, I think we only have half the picture; as for brains, configuration of connections are as important as layout of the circuitry.
Late to the party, but corresponding author on the paper here. Cool to see this on HN!
We have less than "half the picture" here. Not just weights; also missing electrical synapses, neurotransmitters, etc. We also don't know the spatial scale of neuronal arbor integration. Furthermore these are just the image data, not the complete connectome; people still have to trace circuits by hand in this dataset. Collaborators are starting to crack the segmentation problem, but it is still early days.
"URL to this view" lets you share URLs to whatever you're looking at.
----
edit:
in mammals there is pretty good circumstantial evidence that post-synaptic density size correlates with evoked postsynaptic potential, but this hasn't been clearly and directly calibrated yet, and could vary from cell type to cell type
What technology needs to be developed to get the data on synapses and neurotransmitters? A high-resolution Raman imaging microscope? And is a dead brain sufficient, or would you need a real-time noninvasive scan of a living one?
In the fruit fly, neuronal cell types are highly morphologically stereotyped and identifiable across animals. This means that for a given cell type, you can collect data on electrical synapses in animal A, on transmitters in animal B, and on electrophysiology in animal C, and in this fashion assemble a unified, multimodal view of the parts involved. Our whole brain EM volume lets you see how those parts are connected.
In the above examples dead brains are okay except for electrophysiology, where the brain needs to be alive.
I was talking about synaptic strength (strength of connections between neurons which can vary from +ve to -ve) [0]. There is no mention of synapse in that article.
> We will continue to improve connectomics reconstruction technology, with the aim of fully automating synapse-resolution connectomics and contributing to ongoing connectomics projects at the Max Planck Institute and elsewhere. In order to help support the larger research community in developing connectomics techniques, we have also open-sourced the TensorFlow code for the flood-filling network approach, along with WebGL visualization software for 3d datasets that we developed to help us understand and improve our reconstruction results.
People are rightfully pointing out that this is only part of the picture, however this is still a really important and useful result!
CNNs came by mimicking the layout of a visual cortex, it's very possible other similar breakthroughs could come from better understanding the layout of animal brains.
Since bio neurons are more complex than "neural network nodes" let's say we dedicate one CPU core per neuron. 100k cores is a pretty standard supercomputer.
Unfortunately connectivity is only one property for models. Biorealistic models are lacking sorely in modeling plasticity, which is prerequisite for understanding learning. We 're way behind on mapping that and it is a very complex problem.
We cannot even model C elegans behavior realistically and it has orders of magnitude fewer neurons. The issue right now is not (solely) computational power.
C elegans has only 302 neurons and rougly 7,000 connections between the neurons. Neuronal connectivity is similar between individuals. The full connectome has been mapped a long time ago. They are not even spiking neurons and their working is still not fully understood.
There exists opens source simulator for the whole worm (roughly 1000 cells total) http://openworm.org/
I think it's fair to say that this map of the fly brain will be extremely useful in validating models of neurons.
One slight challenge may be that we are certainly not close to being able to build a physical robot fly that resembles a real one, but perhaps a simulated one could do.
Simulated wings. Simulated muscles. Why bother hooking it up to anything physical. Even so it won't fly (apologies), it's way more complex than other challenges that are still unsolved in this field.
I don't quite understand what you mean... is it that the logical complexity of the fly brain is lower than the hardware complexity so it should be possible to replicate the "algorithm" to fly without simulating the hardware of the fly brain?
Neuron is the hardware. We have no idea how the software on it works. The way we model neural networks might be a very different from the operating system of the brain. It's like saying we mapped the x86 chip so we should be able to to simulate all it can do without reverse engineering the software.
Directly or indirectly connected isn't my point. I was referring to the large amount of electrical activity in the brain and we don't really understand what it actually encodes. Unless we reverse engineer it (which is insanely hard), it's hard to know what's "running" on the hardware.
Or if we can perfectly replicate the brain using our model of deep learning. I'm a skeptic but that would prove it too.
When you emulate NES games, you don't emulate all of the atoms in the NES processor, you just go with instruction set and write down some simple aproximation.
You would do the same in biology.
c elegans connectome has about 300 neurons. If you would simulate the neurons on a hardware level it would be enormously complex task, you would have to deal with ion channels and quirky stochastic effects, perhaps even some quantum physics simulation.
The brain of c elegans is insanely complex. The brain structure, on the other hand, is ridiculously simple.
You can just take connection and activation weights and write a simple neural network equivalent in an evening. And it would work about as good as a hardware simulation.
> You can just take connection and activation weights and write a simple neural network equivalent in an evening. And it would work about as good as a hardware simulation.
I'm under the impression that this is very far from a settled thing. In effect, we don't know whether the stochastic and quantum effects are important or not. Perhaps that is the secret sauce, and it is not possible to approximate it using "crude" mathematical models we use today.
Do you have any links to research that supports your idea?
You can just take connection and activation weights and write a simple neural network equivalent in an evening. And it would work about as good as a hardware simulation.
Isn't this currently an unsettled question? We don't know if simple approximations of easily observable brain features (like connections) works as well for imitating brains as emulating NES instructions works for imitating the NES. Among other obstacles, we haven't had complete connection maps of large, interesting brains to experiment with. It seems like this fruit fly work may be a stepping stone toward more sophisticated "emulation" experiments.
> You can just take connection and activation weights and write a simple neural network equivalent in an evening. And it would work about as good as a hardware simulation.
And yet, nobody has managed to do so, despite years of effort.
Neuron is the hardware. We have no idea how the software on it works. The way we model neural networks might be a very different from the operating system of the brain. It's like saying we mapped the x86 chip so we should be able to to simulate all it can do without reverse engineering the software.
That's a huge oversimplification. There isn't some hard drive from where software is loaded onto neurons. A closer model would be saying that the software is the neuron's interconnections and it's activation weights. The neurons are both the hardware and the software.
We don't know that. For instance, there is a "chemical clock" that teaches the hard to beat in feutuses (and it does some other stuff too). There is much behavior training in other parts as well, for instance to make you breathe there is a reflex. I think it's a pretty safe assumption to make that we only know the big honking in-your-face examples of genetic behavior. *
So while there is no hard drive with data it does "load data on the networks" in the sense that it puts them into train mode and trains a specific set of actions.
Secondly, from ~16 days after conception the brain (at that point a human embryo is something close to a fish) the brain is trying to learn. At that point, all it can perceive is the mother. So it learns ~8.5 months from the mother before drawing it's first breath. That's also "loading from the harddrive", just with more efficiency.
* one revealing experience I've had is owning an animal and then going to a dog show. The character of dogs of the same breed, even ones that have never seen eachother, ever, in uncannily similar. For horses and cats, I hear it's even worse. I've heard claims that some horse trainers know the exact family relations of a group of horses after observing them for 10 minutes just by seeing behavior differences. But the cats and dogs examples are pretty strong, since those are matching character traits in animals that don't share any close family bond, just the breed, and have never seen eachother. Their mothers and fathers have never seen eachother. And still, they react to strangers exactly the same way, down to whether their head is tilted left or right, how long they look, and many such small things.
How would you be so confident that neuron is both the software and the hardware? Do we actually understand how we have cognition ? There is tremendous amount of electrical activity in the brain (and the entire nervous system). One could argue it's a low dimensional signal like an activor or a much higher abstraction like the software we run on our chips. Just because we use neural network like structures to generate activation weights, how does it follow that's how the brain works?
From a former neuro person: The physical world is not an abstraction- it's material. If you talk to anyone with any reasonable level of knowledge about neuroscience, you aren't going to find anyone who finds the Cartesian dualism you're effectively proposing credible anymore. The only real philosophies that hold their own in the psych/neuro world are reductionism/materialism and biological naturalism.
That and lesion studies / patients like H.M. have pretty effectively demonstrated that neurons have both processing and storage functions that aren't necessarily mutually exclusive.
It's also worth pointing out that there are multiple types of neurons, so I suppose that you could argue that some of them are more "software-like" while others are more "hardware-like", as well as glial cells that essentially serve as dumb cabling (huge oversimplification). Purkinje cells in the cerebellum, for instance, are known to be more involved in muscle memory, although they still exhibit properties that are both hardware/software-like.
