Network of 75 Million Neurons of the Mouse Brain Mapped for the First Time | Hacker News Reader

Network of 75 Million Neurons of the Mouse Brain Mapped for the First Time

singularityhub.com

87 pointssethbannon34 comments

Like its artificial neural net cousins, I feel like black-box approaches are generally weak for studying intelligence. Programming a complex model is relatively simple, given decades of neurobiological research. But running this brain, absent a physical embodiment or model simplifications, offers few lessons about the nature, structure, and form of intelligence.

The purpose of a brain is to control homeostasis, motor, and perception functionality. Without a rich world simulation, the brain is just an isolated processor. By modelling at the neuron level, the only inputs can be at the neuron level. That means you have to physically model the entire mouse's body to connect all the inputs.

Even if you had a realistic body simulation, you also need to run the system for an extended period of time. Intelligence is an emergent behavior, and it needs time to work and learn. Since you have to model every single neuron firing at the same time, the speed at which you can run this model is unreasonably slow.

I feel like approaches which slowly build a system of models, each of which simplifies the function of a particular neural region to something computationally tractable is the more realistic approach. In doing so, you develop models which reveal the structure of intelligence.

That's a very cool way to think about it. Kind of like aliens finding my Core i7 CPU and announcing they had replicated its transistor network ... except for all they can do is run power to it and they have no idea how the RAM, GPU, or any I/O works.

You have to start somewhere though.

I totally agree with this!

I can't say I agree with this "Intelligence is an emergent behavior, and it needs time to work and learn. Since you have to model every single neuron firing at the same time, the speed at which you can run this model is unreasonably slow."

Part of the issue is semantics, what do you consider intelligent? And part is empirical. Insects pop out of a an egg and start doing their 'thing.' Some argue its "intelligence" and others argue is programming of an automaton. Having seen a number of mice go through the whole cycle, from birth to snake chow, my observation was that capabilities developed over time but intelligence appeared to be innate. (which is different than learning, it is in my opinion the ability to learn). I don't know how one my structure an experiment of keeping a mouse effectively in a coma until its body was fully grown, and then measuring its ability to learn, but I did observe that baby mice learned readily things like how to avoid losing track of mom, and how to keep warm. Programming or instinct or something else I don't know.

I do agree that without inputs the brain does not learn but I would not agree that it is not intelligent. So the question might be how fast can it learn. An interesting question might be is it possible to measure the speed of learning independent of a mouse acting on that learning.

> I don't know how one my structure

my -> might.

> build a system of models, each of which simplifies the function of a particular neural region to something computationally tractable

I think I've worked on too many large software systems, but my guess?

The human brain is the world's worst spaghetti code and this will be impossible.

Why not? What they achieved is a perfectly scientific way of resolving a mystery. Besides, they provide a huge amount of data to mine possibly with a very high information content. This is similar to using telescopes to observe distant galaxies or cosmic background radiation and using that data to prove facts about how our universe works.

> The purpose of a brain is to control homeostasis, motor, and perception functionality.

The fact you wrote this sentence, let me assume your prefrontal cortex has actually some more abilities :)

Excuse the plug, but I’m Tim Ragan the CEO of TissueVision. We’re the MIT startup who developed the microscope technology responsible for all the mouse brain images in the Allen Connectivity map that you see here.

We’re thrilled to have been a part of this groundbreaking project. If you’re excited about this sort of science too, drop me a line. We’re hiring talented developers who can push not just the microscope technology but also the super cool bioinformatics just waiting to be mined from these and other projects.

Tim

Wow just wanted to say thank you for developing this technology. From your white paper [1] I can see how serial two-photon tomography is a huge leap in imaging and a crucial platform tool for the kind of ambitious large-scale surveys we badly need.

I can see too how your main challenge is going to be registration/segmentation on such huge datasets but there's plenty of ML/computer vision work and talent next door in Cambridge. Maybe pair-up or present a poster during the incoming New England Machine Learning day at Microsoft Research this May 13th? (Poster deadline April 25) [2]. Good luck with your recruiting!

[1] http://www.tissuevision.com/nihms344616.pdf

[2] http://research.microsoft.com/en-us/events/neml2014/

It says the data's available at http://mouse.brain-map.org/

Is there a torrent available?

And what is the probability my ISP will cut me off for downloading 1.8 PB?

For that matter...I guess I'd need to buy ~600 3 TB hard drives to store it...that would cost $60,000. Plus networking and servers...eep...

Why is it so big?

Suppose you have 75M neurons where each neuron has on average 1000 weighted connections with other neurons. Then you would need 4 bytes per connection to store the address and 4 more bytes to store the weight, for a total of

    (4 + 4 bytes) * 1,000 * 75,000,000 ~ 600 GB

This back-of-the-envelope computation is off by over three orders of magnitude. What gives?

The data they're talking about isn't anything like that at all, it's actually micron-resolution 3D scans. Here's the abstract of the paper:

>Comprehensive knowledge of the brain’s wiring diagram is fundamental for understanding how the nervous system processes information at both local and global scales. However, with the singular exception of the C. elegans microscale connectome, there are no complete connectivity data sets in other species. Here we report a brain-wide, cellular-level, mesoscale connectome for the mouse. The Allen Mouse Brain Connectivity Atlas uses enhanced green fluorescent protein (EGFP)-expressing adeno-associated viral vectors to trace axonal projections from defined regions and cell types, and high-throughput serial two-photon tomography to image the EGFP-labelled axons throughout the brain. This systematic and standardized approach allows spatial registration of individual experiments into a common three dimensional (3D) reference space, resulting in a whole-brain connectivity matrix. A computational model yields insights into connectional strength distribution, symmetry and other network properties. Virtual tractography illustrates 3D topography among interconnected regions. Cortico-thalamic pathway analysis demonstrates segregation and integration of parallel pathways. The Allen Mouse Brain Connectivity Atlas is a freely available, foundational resource for structural and functional investigations into the neural circuits that support behavioural and cognitive processes in health and disease.

You describe a very abstract type of a synapse. Nature is more complex than this. I think it's just a definition problem. What do we mean with modeling a brain? Is it the same like modeling a neuronal network? On which level do we define a brain. E. g. We can also include all chemistry and even fluid dynamics in a cell. E. g. to understand a computer: do we have to have to simulate on the electron leven with doped compounds and barriers, or is it enough to model it with characteristic curves of transistors, or is it even enough to model it on the binary level as switches between 0 and 1?

They seem to be coming up with great work there. Unfortunately, thousands of neuroscientists who have been sharing huge funds the past decade couldn't come up with a big project like this one, it had to take private initiative to map a mouse. To me this shows how badly allocated funding sets the bar too low for most neuroscience teams.

And for those who say that this is not a big deal, it actually is a very big deal. Not only for the connection map of the brain, which may or may not lead to simulating a brain (we are far from that yet), but because we can analyze the distributions of connections and their weights to infer possible functional roles.

All of the data is available through a well-documented API: http://www.brain-map.org/api/index.html

Fully documented with plenty of examples and relevant white papers.

"It’s hard to comprehend how, in this age of advanced science, experts don’t know more about how the brain works. But it has a vastly complex structure."

Yes, we have a map of 75 million thingies and how they interact in some ways ... but we don't really know what even the interactions of just two neurons "really mean," which parts of their many interactions are the important ones, how they change over time and so-forth.

It does we humans a lot time to figure what an artifact from an alien human civilization does, even a fairly simple one - and that figuring generally involves something like figuring out the "blue print", the "purpose", the "programming" and so-forth of such an artifact. These are all analogies with our own intentional processes.

"Figuring out" the brain even as if it was a human-created artifact seems immensely hard. Yet like DNA, the brain isn't human-created or created at least-all. It just happened, in ways whose understanding will clearly take a lot of unraveling to really understand.

Not sure where I'm going with this aside from awe of the challenge.

Pretty much all human brains, that I know of, were created by humans.

I guess they meant to say design.

For those of you with institutional access: http://www.nature.com/nature/journal/v508/n7495/full/nature1...

What is your guess about the amount of memory needed to save a human brain. I mean all necessary information to save the personality. The article mentions 98,000 petabytes. But does this contain everything?

In Wolfram alpha and Wikipedia you can read, that the storage capacity of the human brain is around 2-2.5PB.

When I estimate it assuming 10^14 synapses and 24 bytes (weight, connection info, etc.) for each synapse, I come close to 2PB.

Well the article near the end says that the data for the experiment was 1.8 PB. I am not sure what the answer is, but given a mouse brain is significantly smaller than a human brain I would say Wolfram Alpha and Wikipedia are wrong.

I am not sure, but after looking at their sites they probably mean images. So, a synapse in a let's say high resolution jpeg will need much more memory than my abstract definition of a synapse.

i'm wondering if the massive amounts of Petabytes can be compressed in some efficient way. probably.

Well, of course. I would suggest paq8l ;)

I think you can store very efficiently. A neuron or synapse has a huge error rate. So, saving parameters of the synapse as 4 byte floating point number could be overkill. Connection information could be also reduced significantly by an intelligent sorting. E. g. you don't have to store the source index and neurons connect probably more to their neighbors then to neurons far away. Therefore I am wondering, what must be saved in 24 byte. Well, we will know, when we try it some day. At the moment it's just speculation.

The article lists 1.8P for a brain with 75 million neurons. Unless I've misplaced my zeroes, that's over ~20M per neuron, which seems much higher than I'd expect.

A neuron is a lot more complicated than you would expect.

I wish I'll be able to understand how the brain really work before dying. I have the feeling that a deep understand of the brain will open scary and amazing opportunities in AI.

Summarizing previous points, a big criticism of studies like these is that they don't have a huge potential to unearth real understanding of real neural networks. The suggestion is to figure out the 302 neurons in [this little worm](https://en.wikipedia.org/wiki/Caenorhabditis_elegans) before attacking problems orders of magitude more difficult, like mouse and human brains.

Is it just me or people actually think it'd be even cooler to have a visualization like the first image?

How do they actually know they have mapped everything right? How do they test that?

For once I find the use of the "Singularity Hub" domain for mirroring a news article entirely acceptable. Whole Brain Emulation for mice would be a big step towards hyperintelligent lobster-ems, and from there it's only a short step to killing all humans and spawning a Matrioshka Brain of capitalistic rational agents that have optimized their consciousness, subjectivity, and humanity away as market inefficiencies!

</Accelerando references>

two questions for someone more familiar with this area:

1. what are the immediate applications of having such a dataset? either for the mouse or the brain?

2. once we have the map, will we also have the technology to identify precisely what neurons are lighting up and how information is travelling during certain actions / responses? in short how far are we from a green / yellow / red Google Maps style traffic map for the brain?

1. There are some people who claim to make whole-brain simulations based on earlier anatomical data [1,2]. I don't expect that anyone can make meaningful predictions by simulating the new map, because the neuronal models used are either rudimentary, or do not capture neuronal variability (also non-neuron cells may be involved). However, the analysis of the connectivity between various regions as well as intra-region may lead to accurate predictions about what many circuits "do".

2. Functional imaging is advancing, but it's still very hard to capture large-scale neuronal activity. The current state of the art, calcium imaging can visualize activity from one focal plane, not too many neurons at a time, with slow time resolution, and only at the surface. I 'd say we are pretty far from that as it would require some amazing imaging technology.

1. http://www.izhikevich.org/publications/large-scale_model_of_...

2. http://bluebrain.epfl.ch/

If the human brain was simple enough to be understood it would be too simple to understand.

We will glean bits and pieces but a detailed functional understanding of an emergent system as large, containing as much feedback (within feedback), and with as enormous a connectivity as any large (or probably even small) mammalian brain is far beyond any reasonable expectation ever.

While this is a great first step, it is an amalgamation of images from over 1000 mice. Each mouse is going to have different connectome, though hopefully we can distinguish major features more clearly now.

Reply on news.ycombinator.com