Jump to content

Welcome to FutureTimeline.forum
Register now to gain access to all of our features. Once registered and logged in, you will be able to create topics, post replies to existing threads, give reputation to your fellow members, get your own private messenger, post status updates, manage your profile and so much more. If you already have an account, login here - otherwise create an account for free today!

These ads will disappear if you register on the forum

Photo

Fruit fly connectome mapped for first time


  • Please log in to reply
3 replies to this topic

#1
funkervogt

funkervogt

    Member

  • Members
  • PipPipPipPipPip
  • 414 posts
Two high-speed electron microscopes. 7,062 brain slices. 21 million images.
 
For a team of scientists at the Howard Hughes Medical Institute's Janelia Research Campus in Ashburn, Virginia, these numbers add up to a technical first: a high-resolution digital snapshot of the adult fruit fly brain.
 
Researchers can now trace the path of any one neuron to any other neuron throughout the whole brain, says neuroscientist Davi Bock, a group leader at Janelia who reported the work along with his colleagues on July 19, 2018, in the journal Cell.
 
"The entire fly brain has never been imaged before at this resolution that lets you see connections between neurons," he says. That detail is key for mapping out the brain's circuitry—the precise webs of neuronal connections that underpin specific fly behaviors.

 

 


  • wjfox, Casey, Yuli Ban and 1 other like this

#2
starspawn0

starspawn0

    Member

  • Members
  • PipPipPipPipPipPip
  • 505 posts
Neuroscientists seem to be exited by this -- e.g. David Markowitz (IARPA and DARPA):

https://mobile.twitt...103267584536576

Also see:

https://mobile.twitt...242969046618113

Connectomics has ENORMOUS potential to benefit the world. Not just by providing deeper insight into differences between healthy and diseased brains, which has literally $Trillion economic implications. But also by advancing engineering disciplines that support the IT industry.


I am not so sure. The worst-case scenario for connectomics is that there isn't much special about the fine-scale structure of the brain -- if you get the rough structure correct (that can already be deduced using crude scanning methods), and use a fairly simple learning algorithm, you get strong AI. All those extra cell types, strange connection patterns, and so forth, are blinding people to this very basic reality. The only reason we don't yet have strong AI has more to do with lack of the necessary computing power.

I suspect the worst-case scenario isn't true -- but is close to the truth, in that getting things roughly right is good enough to build decently-performing systems, that learn a little slower than humans; but that, through large numbers of training examples, can still match humans in most tasks.
  • Casey and Yuli Ban like this

#3
funkervogt

funkervogt

    Member

  • Members
  • PipPipPipPipPip
  • 414 posts

Marvin Minsky said a few years ago that the best route to AGI would be to first map a fruit fly brain at probably the same level of detail as was just done, and to develop a comprehensive algorithmic/schematic understanding of how it operates. Once we have "fruit fly level AGIs" we can apply the fundamental lessons learned to making the next most complex type of animal AGI, and so on until we're ready to make human-level AI. 

 

https://youtu.be/3PdxQbOvAlI?t=27m23s

 

It's hard for me to believe that we lack computers with enough power to emulate fly brains. 

 

 

 

I am not so sure. The worst-case scenario for connectomics is that there isn't much special about the fine-scale structure of the brain -- if you get the rough structure correct (that can already be deduced using crude scanning methods), and use a fairly simple learning algorithm, you get strong AI. All those extra cell types, strange connection patterns, and so forth, are blinding people to this very basic reality. The only reason we don't yet have strong AI has more to do with lack of the necessary computing power.

I suspect the worst-case scenario isn't true -- but is close to the truth, in that getting things roughly right is good enough to build decently-performing systems, that learn a little slower than humans; but that, through large numbers of training examples, can still match humans in most tasks. 

Let me indulge in wishful, pointless thinking for a moment about how I'd like to see the world run differently to say that I'd like to see a "Brain/AGI Manhattan Project that costs 20% as much as the atom bomb Manhattan Project," where 500+ talented scientists working in dozens of teams would explore different approaches to the problem. They'd be assured of indefinite funding and would have adequate lab equipment for the task. 

 

I know there'd be enormous whining about the ~$1 billion annual cost, but if it brought about the advent of AGI even ONE YEAR sooner than would have happened otherwise, it would pay for itself many times over and possibly save the lives of 55 million humans. 

 

To put that price tag into perspective, consider that $2.6 billion of Air Jordan sneakers were sold in 2014.

https://themarketmog...r-jordan-brand/


  • Yuli Ban likes this

#4
starspawn0

starspawn0

    Member

  • Members
  • PipPipPipPipPipPip
  • 505 posts
The problem is that just knowing the connectome and how individual neurons behave is not enough:

https://www.biorxiv....18/07/08/364489
 

First, by varying the network's nonlinear activation and rate regularization, we show that RNNs reproducing single neuron firing rate motifs may not adequately capture important population motifs.


But if you record population behaviors, you can train deep neural nets to emulate the dynamics. Somehow, Deep Learning "knows" what the network is trying to do, and can emulate it shockingly well.

This is going to pose a challenge for "mind uploading" proponents, who think if they just copy the connectome and behavior of individual neurons, they can simulate a brain. They are going to need a lot more information than that.

On the other hand, using much cruder information, in the form of population dynamics, they can at least emulated some important aspects of the dynamics. Not only that, but these trained models generalize nicely to new tasks:
 

Finally, we show that these dynamics are sufficient for the RNN to generalize to a target switch task it was not
trained on.


Minsky is a fine one to talk about brain emulation, given how he tried to kill Deep Learning with his book with Papert.

https://mobile.twitt...301851731243010

https://mobile.twitt...513328283082753
 

I disagree. At the time, there was already work on MLPs, and M & P undersold their value. Based on some things I’ve heard about their feelings re Rosenblatt, it sounds like their book was not aiming for balance, it was aiming for a knock down.


  • Yuli Ban likes this




0 user(s) are reading this topic

0 members, 0 guests, 0 anonymous users