5 ms·

600,000 neurons seems small enough to simulate - and microscopes are easily good enough to discern cells (neurons) and their connections. Though the weightings

600,000 neurons seems small enough to simulate - and microscopes are easily good enough to discern cells (neurons) and their connections. Though the weightings between them probably can't be detected. At least, that's my layman's view... what specific aspects of tools are actually deficient? It could be a useful starting point for eventually understanding ourselves. I expect we'll need to develop hierarchical models for how neurons operate, to cope with the complexity... perhaps even new pure mathematics.

atemerev 10y ago
Human Brain Project has the means to simulate around 30 thousand biological neurons — using a much larger supercomputer, of course. But it isn't a far cry from getting to 600,000.
biofox 10y ago
The problem isn't the number of neurons and their connections, but the complexity and diversity of them. There are hundreds of signalling molecules and receptors that haven't been characterised. Some molecules diffuse over long distances, others are taken up and recycled depending on the local environment. Surrounding non-neuronal cells alter the ionic and metabolic states of networks, and can physically block (or sometimes actively transport) neurotransmitters. Each neuron type has its own response characteristics (activation function) that can completely change depending on the context. Even different locations on the same neuron have different receptors and transmitters, and can alternate between inhibitory or excitatory functions. To give an example of the difficulty, we've had a complete map of the neural connectivity of C. elegans for 30 years, which consists of only 302 neurons, and we still don't know how many of the cells and circuits function. I think the only solution will be high-throughput phenotyping. Test cell types and mutants under different circumstances until we have a deep-enough understand of these systems and their interactions to model them. It takes humans far too long to do experiments. We need more lab automation.
- intrasight 10y ago
  Good point about the automation. Perhaps then in my lifetime.
- whamlastxmas 10y ago
  I wonder if lab automation would be a good topic for the YC Research efforts.
  - biofox 10y ago
    I think it would. As things are at the moment, biologists spend most of their time doing repetitive manual labour. One of the most promising things I've seen is the Emerald Cloud Lab, but it's still a long way off from what's needed: http://emeraldcloudlab.com/ http://emeraldcloudlab.com/
    - Kliment 10y ago
      Hey, I'd like to talk with you about automating some of that stuff, but there's no contact info in your profile. My email and IRC contact are in mine if you'd be willing to get in touch.
RobertKerans 10y ago
But knowing that still doesn't bring you that much closer to simulating it. There's a couple a of great papers where they got biologists to attempt to figure out how relatively simple, very well known things worked - one was a radio[1], one was an old Atari chip[2]. So they used biological techniques used to try to figure out how brains worked. For example lesions - they took transistors out of the arcade machine and tested if the game still started up. And if, say, they found that taking a and b out stopped Donkey Kong from working, and x and y stopped space invaders working, then maybe they could make the assumption that a and b were the Donkey Kong transistors, etc. Or if they destroyed tiny parts of the radio circuit board and it still appeared to work, then those parts were junk, unnecessary. But of course that's completely misleading. It doesn't matter how much data they had either. It's not to say the data that is currently held isn't important; it is, but it doesn't then translate to 'we can model this insanely complex system', because there's so, so much missing from the understanding of how it works [1] https://www.cmu.edu/biolphys/deserno/pdf/can_a_biologist_fix_a_radio.pdf https://www.cmu.edu/biolphys/deserno/pdf/can_a_biologist_fix... [2] http://biorxiv.org/content/biorxiv/early/2016/05/26/055624.full.pdf http://biorxiv.org/content/biorxiv/early/2016/05/26/055624.f...
- intrasight 10y ago
  It's just as revealing to have an engineer try to understand a living thing. There's a huge cultural and language gap between the disciplines.
  - regularfry 10y ago
    Apparently opticians can spot an engineer a mile off because they think of the eye as a simple lens system. It's not.
- merastius 10y ago
  This, and our failure to simulate something as well understood as the c. elegans' 302 neurons, makes me think that biologists need to push a new approach. When it comes to understanding a radio or a microchip, my first instinct would be to try to isolate components which are as small as possible, and actually try to map inputs to outputs... And if that's not working, either the component isn't small/basic enough, or I haven't really captured all the inputs. I'd hopefully start getting an idea about logic gates and voltage thresholds and eventually building up to larger and more complex processes and parts. 1) Would such an approach work in theory for a radio or chip? 2) I'm assuming I'm not the first person to think of this - is that approach basically what's already being done? (edit: a lot of stuff in this seem to be along those line: https://en.wikipedia.org/wiki/Electrophysiology https://en.wikipedia.org/wiki/Electrophysiology ) 3) Is it just much harder than I imagine to isolate a single neuron (or small set of neurons) and try some inputs to come up with the algorithm to properly simulate it? 4) Or are neurons too complex for this approach to be effective, and we need to dig down to more basic components or get a better understanding of inputs?

atemerev 10y ago

Human Brain Project has the means to simulate around 30 thousand biological neurons — using a much larger supercomputer, of course. But it isn't a far cry from getting to 600,000.

biofox 10y ago

The problem isn't the number of neurons and their connections, but the complexity and diversity of them. There are hundreds of signalling molecules and receptors that haven't been characterised. Some molecules diffuse over long distances, others are taken up and recycled depending on the local environment. Surrounding non-neuronal cells alter the ionic and metabolic states of networks, and can physically block (or sometimes actively transport) neurotransmitters. Each neuron type has its own response characteristics (activation function) that can completely change depending on the context. Even different locations on the same neuron have different receptors and transmitters, and can alternate between inhibitory or excitatory functions. To give an example of the difficulty, we've had a complete map of the neural connectivity of C. elegans for 30 years, which consists of only 302 neurons, and we still don't know how many of the cells and circuits function. I think the only solution will be high-throughput phenotyping. Test cell types and mutants under different circumstances until we have a deep-enough understand of these systems and their interactions to model them. It takes humans far too long to do experiments. We need more lab automation.

intrasight 10y ago
Good point about the automation. Perhaps then in my lifetime.
whamlastxmas 10y ago
I wonder if lab automation would be a good topic for the YC Research efforts.
- biofox 10y ago
  I think it would. As things are at the moment, biologists spend most of their time doing repetitive manual labour. One of the most promising things I've seen is the Emerald Cloud Lab, but it's still a long way off from what's needed: http://emeraldcloudlab.com/ http://emeraldcloudlab.com/
  - Kliment 10y ago
    Hey, I'd like to talk with you about automating some of that stuff, but there's no contact info in your profile. My email and IRC contact are in mine if you'd be willing to get in touch.

RobertKerans 10y ago

But knowing that still doesn't bring you that much closer to simulating it. There's a couple a of great papers where they got biologists to attempt to figure out how relatively simple, very well known things worked - one was a radio[1], one was an old Atari chip[2]. So they used biological techniques used to try to figure out how brains worked. For example lesions - they took transistors out of the arcade machine and tested if the game still started up. And if, say, they found that taking a and b out stopped Donkey Kong from working, and x and y stopped space invaders working, then maybe they could make the assumption that a and b were the Donkey Kong transistors, etc. Or if they destroyed tiny parts of the radio circuit board and it still appeared to work, then those parts were junk, unnecessary. But of course that's completely misleading. It doesn't matter how much data they had either. It's not to say the data that is currently held isn't important; it is, but it doesn't then translate to 'we can model this insanely complex system', because there's so, so much missing from the understanding of how it works [1] https://www.cmu.edu/biolphys/deserno/pdf/can_a_biologist_fix_a_radio.pdf https://www.cmu.edu/biolphys/deserno/pdf/can_a_biologist_fix... [2] http://biorxiv.org/content/biorxiv/early/2016/05/26/055624.full.pdf http://biorxiv.org/content/biorxiv/early/2016/05/26/055624.f...

intrasight 10y ago
It's just as revealing to have an engineer try to understand a living thing. There's a huge cultural and language gap between the disciplines.
- regularfry 10y ago
  Apparently opticians can spot an engineer a mile off because they think of the eye as a simple lens system. It's not.
merastius 10y ago
This, and our failure to simulate something as well understood as the c. elegans' 302 neurons, makes me think that biologists need to push a new approach. When it comes to understanding a radio or a microchip, my first instinct would be to try to isolate components which are as small as possible, and actually try to map inputs to outputs... And if that's not working, either the component isn't small/basic enough, or I haven't really captured all the inputs. I'd hopefully start getting an idea about logic gates and voltage thresholds and eventually building up to larger and more complex processes and parts. 1) Would such an approach work in theory for a radio or chip? 2) I'm assuming I'm not the first person to think of this - is that approach basically what's already being done? (edit: a lot of stuff in this seem to be along those line: https://en.wikipedia.org/wiki/Electrophysiology https://en.wikipedia.org/wiki/Electrophysiology ) 3) Is it just much harder than I imagine to isolate a single neuron (or small set of neurons) and try some inputs to come up with the algorithm to properly simulate it? 4) Or are neurons too complex for this approach to be effective, and we need to dig down to more basic components or get a better understanding of inputs?