f**d 发帖数: 768 | 1 作为开发下一代智能机器人平台的脑计算机仿真已受到人们的广泛关注。2008年2
月,美国国家工程院发表了用工程方法重建人脑的重大挑战书http://www.
engineeringchallenges.org/cms/8996/9109.aspx)。
IBM和DARPA(Defense Advanced Research Projects Agency) 目前都积极地投入到
这项科学研究中。欧盟也有“幼兽机器人”(RobotCub)大计划,以脑科学为基础
发展智能机器人。在2009 Kavli Futures Symposium有关“幻想的计算机”的讨论
会中,计算机学家、脑神经学家、物理学家们有一个共识:和计算神经科学的紧密
合作能帮助人们为未来的机器人设计出创新型的结构
(参见http://www.kavlifoundation.org/extreme-achine-2009-kavli-futures-symposium)。
美国国家工程院发表了用工程方法重建人脑的重大挑战书:
http://www.engineeringchallenges.org/cms/8996/9109.aspx
Reverse-engineer the brain
For decades, some of engineering’s best minds have focused their
thinking skills on how to create thinking machines — computers
capable of emulating human intelligence.
Why should you reverse-engineer the brain?
While some of thinking machines have mastered specific narrow skills —
playing chess, for instance — general-purpose artificial
intelligence (AI) has remained elusive.
Part of the problem, some experts now believe, is that artificial brains
have been designed without much attention to real ones. Pioneers of
artificial intelligence approached thinking the way that aeronautical
engineers approached flying without much learning from birds. It has
turned out, though, that the secrets about how living brains work may
offer the best guide to engineering the artificial variety.
Discovering those secrets by reverse-engineering the brain promises
enormous opportunities for reproducing intelligence the way assembly
lines spit out cars or computers.
Figuring out how the brain works will offer rewards beyond building
smarter computers. Advances gained from studying the brain may in return
pay dividends for the brain itself. Understanding its methods will
enable engineers to simulate its activities, leading to deeper
insights about how and why the brain works and fails. Such simulations
will offer more precise methods for testing potential biotechnology
solutions to brain disorders, such as drugs or neural implants.
Neurological disorders may someday be circumvented by technological
innovations that allow wiring of new materials into our bodies to do the
jobs of lost or damaged nerve cells. Implanted electronic devices could
help victims of dementia to remember, blind people to see, and crippled
people to walk.
Sophisticated computer simulations could also be used in many other
applications. Simulating the interactions of proteins in cells would
be a novel way of designing and testing drugs, for instance. And
simulation capacity will be helpful beyond biology, perhaps in
forecasting the impact of earthquakes in ways that would help guide
evacuation and recovery plans.
Much of this power to simulate reality effectively will come from
increased computing capability rooted in the reverse-engineering of
the brain. Learning from how the brain itself learns, researchers will
likely improve knowledge of how to design computing devices that process
multiple streams of information in parallel, rather than the
one-step-at-a-time approach of the basic PC. Another feature of real
brains is the vast connectivity of nerve cells, the biological
equivalent of computer signaling switches. While nerve cells typically
form tens of thousands of connections with their neighbors,
traditional computer switches typically possess only two or three. AI
systems attempting to replicate human abilities, such as vision, are now
being developed with more, and more complex, connections.
What are the applications for this information?
Already, some applications using artificial intelligence have
benefited from simulations based on brain reverse-engineering.
Examples include AI algorithms used in speech recognition and in machine
vision systems in automated factories. More advanced AI software should
in the future be able to guide devices that can enter the body to
perform medical diagnoses and treatments.
Of potentially even greater impact on human health and well-being is the
use of new AI insights for repairing broken brains. Damage from injury
or disease to the hippocampus, a brain structure important for learning
and memory, can disrupt the proper electrical signaling between nerve
cells that is needed for forming and recalling memories. With
knowledge of the proper signaling patterns in healthy brains,
engineers have begun to design computer chips that mimic the brain’s
own communication skills. Such chips could be useful in cases where
healthy brain tissue is starved for information because of the barrier
imposed by damaged tissue. In principle, signals from the healthy tissue
could be recorded by an implantable chip, which would then generate new
signals to bypass the damage. Such an electronic alternate signaling
route could help restore normal memory skills to an impaired brain
that otherwise could not form them.
“Neural prostheses” have already been put to use in the form of
cochlear implants to treat hearing loss and stimulating electrodes to
treat Parkinson’s disease. Progress has also been made in developing “
artificial retinas,” light-sensitive chips that could help restore
vision.
Even more ambitious programs are underway for systems to control
artificial limbs. Engineers envision computerized implants capable of
receiving the signals from thousands of the brain’s nerve cells and
then wirelessly transmitting that information to an interface device
that would decode the brain’s intentions. The interface could then send
signals to an artificial limb, or even directly to nerves and muscles,
giving directions for implementing the desired movements.
Other research has explored, with some success, implants that could
literally read the thoughts of immobilized patients and signal an
external computer, giving people unable to speak or even move a way to
communicate with the outside world.
What is needed to reverse-engineer the brain?
The progress so far is impressive. But to fully realize the brain’s
potential to teach us how to make machines learn and think, further
advances are needed in the technology for understanding the brain in the
first place. Modern noninvasive methods for simultaneously measuring
the activity of many brain cells have provided a major boost in that
direction, but details of the brain’s secret communication code
remain to be deciphered. Nerve cells communicate by firing electrical
pulses that release small molecules called neurotransmitters, chemical
messengers that hop from one nerve cell to a neighbor, inducing the
neighbor to fire a signal of its own (or, in some cases, inhibiting
the neighbor from sending signals). Because each nerve cell receives
messages from tens of thousands of others, and circuits of nerve cells
link up in complex networks, it is extremely difficult to completely
trace the signaling pathways.
Furthermore, the code itself is complex — nerve cells fire at different
rates, depending on the sum of incoming messages. Sometimes the
signaling is generated in rapid-fire bursts; sometimes it is more
leisurely. And much of mental function seems based on the firing of
multiple nerve cells around the brain in synchrony. Teasing out and
analyzing all the complexities of nerve cell signals, their dynamics,
pathways, and feedback loops, presents a major challenge.
Today’s computers have electronic logic gates that are either on or
off, but if engineers could replicate neurons’ ability to assume
various levels of excitation, they could create much more powerful
computing machines. Success toward fully understanding brain activity
will, in any case, open new avenues for deeper understanding of the
basis for intelligence and even consciousness, no doubt providing
engineers with insight into even grander accomplishments for enhancing
the joy of living.
References
Berger, T.W., et al. Restoring Lost Cognitive Function,” IEEE
Engineering in Medicine and Biology Magazine (September/October 2005),
pp. 30-44.
Griffith, A. 2007. Chipping In,” Scientific American (February 2007),
pp. 18-20.
Handelman, S. The Memory Hacker,” Popular Science (2005).
Hapgood, F. Reverse-Engineering the Brain,” Technology Review (July
11, 2006).
Lebedev, M.A. and Miguel A.L. Nicolelis. Brain-machine interfaces: Past,
present, and future,” Trends in Neurosciences 29 (September 2006), pp.
536-546. | n*****d 发帖数: 11 | | m******1 发帖数: 95 | 3 我们做得和他们很相近,但关注的问题不同。我们在硬件芯片上重建神经结构,譬如重
建spinal cord,然后在保证时间精度和神经网络规模的情况下让硬件以300倍的真实速
度运行。我们是通过这种高速仿真来重建儿童神经疾病的长期发展。不过组内包括我在
内的不少人都是EE出身,一转手就把这玩意当neural computer在玩了
【在 n*****d 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 补充一点有意思的信息: : Kwabena Boahen, Stanford, gave a talk on his Brain-on-a-Chip project : http://academicearth.org/lectures/googling-the-brain-kwabena-bo : http://www.stanford.edu/group/brainsinsilicon/index.html : http://www.stanford.edu/group/brainsinsilicon/goals.html : Jeff Hawkins' TEDTalk on how brain science will change computing : http://blog.ted.com/2007/05/23/jeff_hawkins_te_1/ : DARPA's brain-on-a-chip venture, called Systems of Neuromorphic Adaptive : Plastic Scalable Electronics ("SyNAPSE") : http://spectrum.ieee.org/tech-talk/semiconductors/devices/darpa
| f**d 发帖数: 768 | 4 你们这样做的目的是想在最近几年内构建人工SPINAL CORD来修复瘫痪病人嘛?
还是以此为设想,而现在所做的不过是骗些FUNDING来玩?
【在 m******1 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 我们做得和他们很相近,但关注的问题不同。我们在硬件芯片上重建神经结构,譬如重 : 建spinal cord,然后在保证时间精度和神经网络规模的情况下让硬件以300倍的真实速 : 度运行。我们是通过这种高速仿真来重建儿童神经疾病的长期发展。不过组内包括我在 : 内的不少人都是EE出身,一转手就把这玩意当neural computer在玩了
| m******1 发帖数: 95 | 5 长期疾病如child dystonia的发展期一般是15-20年,在这期间多种神经机制交互作用,周期短的如reflex以毫秒计,周期长的如神经元死亡以年计。如果关心的问题是“短期reflex怎么影响15-20年的疾病进程”这种,别说答案了,连猜想都没有。所以构建spinal cord这部分的第一个目的就是起码找到一些testable hypotheses
当然NIH给钱的另一个理由,是这套硬件工具可以用来做drug screening. 神经科在临床上有一个让医方和药方都很尴尬的现状,也就是用药非常盲目。以dystonia为例,神经科医生基本上是开了药试试看,不行再换一种,奏效了都不清楚为什么。我们做得东西相当于用电子技术弄个动物模型出来,然后拿来试药。
至于是不是在骗funding,应该不是吧,至少我觉得很serious. 毕竟老板也是既在NASA当过工程师又在医院做neurologist,想了很久积累出来的框架。加上刚才说了,这套东西实际上是个neural computer,所有工作加在一起足够干上二三十年
【在 f**d 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 你们这样做的目的是想在最近几年内构建人工SPINAL CORD来修复瘫痪病人嘛? : 还是以此为设想,而现在所做的不过是骗些FUNDING来玩?
| f**d 发帖数: 768 | 6 说的很好
有点不清楚
在动物体内用药,药物的新陈代谢,以及和神经元内部各种生物大分子产生生化作用。
。。进而影响神经元活动
对于你们这样做成的电子模型,没有新陈代谢过程,没有生物大分子的化学活动
你们怎么以此来研究药物的作用?
用,周期短的如reflex以毫秒计,周期长的如神经元死亡以年计。如果关心的问题是“
短期reflex怎么影响15-20年的疾病进程”这种,别说答案了,连猜想都没有。所以构
建spinal cord这部分的第一个目的:
临床上有一个让医方和药方都很尴尬的现状,也就是用药非常盲目。以dystonia为例,
神经科医生基本上是开了药试试看,不行再换一种,奏效了都不清楚为什么。我们做得
东西相当于用电子技术弄个动物模:
NASA当过工程师又在医院做neurologist,想了很久积累出来的框架。加上刚才说了,
这套东西实际上是个neural computer,所有工作加在一起足够干上二三十年
【在 m******1 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 长期疾病如child dystonia的发展期一般是15-20年,在这期间多种神经机制交互作用,周期短的如reflex以毫秒计,周期长的如神经元死亡以年计。如果关心的问题是“短期reflex怎么影响15-20年的疾病进程”这种,别说答案了,连猜想都没有。所以构建spinal cord这部分的第一个目的就是起码找到一些testable hypotheses : 当然NIH给钱的另一个理由,是这套硬件工具可以用来做drug screening. 神经科在临床上有一个让医方和药方都很尴尬的现状,也就是用药非常盲目。以dystonia为例,神经科医生基本上是开了药试试看,不行再换一种,奏效了都不清楚为什么。我们做得东西相当于用电子技术弄个动物模型出来,然后拿来试药。 : 至于是不是在骗funding,应该不是吧,至少我觉得很serious. 毕竟老板也是既在NASA当过工程师又在医院做neurologist,想了很久积累出来的框架。加上刚才说了,这套东西实际上是个neural computer,所有工作加在一起足够干上二三十年
| m******1 发帖数: 95 | 7 Glad to answer,两方面来说吧。第一,All models are wrong, but some are
useful,这也是名言了。动物模型wrong的部分在于忽略掉比人低等的,或忽略掉和人不同的部分,而useful的部分就像你说的,能照顾到大部分的分子级、原子级的机理。电子模型是另一种逻辑,它wrong的部分是忽略掉一部分机理,譬如我们现在就只能到细胞级和部分的分子级,再往下就是有模型也没有足够的计算能力来重建;电子模型useful的部分是快,快于实际速度,试想用老鼠猴子来测试15-20年的疾病发展或者药效,也得花上15-20年。
第二呢,我们关注的疾病是运动障碍movement disorder,所以在basic science上我们研究的是运动控制neural control of movement。运动控制这个领域和cognitive neuroscience紧密相关,属于从宏观到微观,自顶向下研究问题的那一支。和自底向上的同行们比,我们可能粗放了一点,但如果放到cognitive领域里面我们又算是关注的尺度很超前的了。
总之do our best吧
【在 f**d 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 说的很好 : 有点不清楚 : 在动物体内用药,药物的新陈代谢,以及和神经元内部各种生物大分子产生生化作用。 : 。。进而影响神经元活动 : 对于你们这样做成的电子模型,没有新陈代谢过程,没有生物大分子的化学活动 : 你们怎么以此来研究药物的作用? : : 用,周期短的如reflex以毫秒计,周期长的如神经元死亡以年计。如果关心的问题是“ : 短期reflex怎么影响15-20年的疾病进程”这种,别说答案了,连猜想都没有。所以构 : 建spinal cord这部分的第一个目的:
| f**d 发帖数: 768 | 8 给我的印象是,
你们这样做,一不小心就可能成了电子游戏,对实际疾病机理没可比拟之处
不过我承认,你们做的正是计算神经领域将来的重点发展方向之一
现在基于生物的神经元和网络模型做细致了,再发展下去的目标就是你们现在正做的
--可能你们现在做的所用模型太粗了点(?是不是?)
用。
是“
以构
例,
【在 m******1 的大作中提到】![](/moin_static193/solenoid/img/up.png) : Glad to answer,两方面来说吧。第一,All models are wrong, but some are : useful,这也是名言了。动物模型wrong的部分在于忽略掉比人低等的,或忽略掉和人不同的部分,而useful的部分就像你说的,能照顾到大部分的分子级、原子级的机理。电子模型是另一种逻辑,它wrong的部分是忽略掉一部分机理,譬如我们现在就只能到细胞级和部分的分子级,再往下就是有模型也没有足够的计算能力来重建;电子模型useful的部分是快,快于实际速度,试想用老鼠猴子来测试15-20年的疾病发展或者药效,也得花上15-20年。 : 第二呢,我们关注的疾病是运动障碍movement disorder,所以在basic science上我们研究的是运动控制neural control of movement。运动控制这个领域和cognitive neuroscience紧密相关,属于从宏观到微观,自顶向下研究问题的那一支。和自底向上的同行们比,我们可能粗放了一点,但如果放到cognitive领域里面我们又算是关注的尺度很超前的了。 : 总之do our best吧
| m******1 发帖数: 95 | 9 讲的很对,就算做出来的神经硬件再牛叉,如果医疗界不买账那就成了个游戏,高级版电子宠物嘛。I like the way you put it
所以谁来做还是很重要的,grant里面的PI和co-PI都是背景非常煊赫的MD+PhD双料,会治病的和会编程的都服他们。大家也都希望这种人给领域折腾出点新风向
模型粗不粗取决于疾病,如果是spinal cord injury这样的硬伤型疾病,走到细胞级都过头了。对于dystonia这种主要因细胞数量减少而造成的疾病来说,细胞级+部分分子级的建模并没有硬伤。如果拿我们这些东西来研究ion channel,那就粗得有点扯淡了
【在 f**d 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 给我的印象是, : 你们这样做,一不小心就可能成了电子游戏,对实际疾病机理没可比拟之处 : 不过我承认,你们做的正是计算神经领域将来的重点发展方向之一 : 现在基于生物的神经元和网络模型做细致了,再发展下去的目标就是你们现在正做的 : --可能你们现在做的所用模型太粗了点(?是不是?) : : 用。 : 是“ : 以构 : 例,
| f**d 发帖数: 768 | 10 good,this make sense,
would you tell me the PI name who has the MD+PhD and doing this
spinal cord in silicon simulation technique?
I want to read his papers
thanks
【在 m******1 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 讲的很对,就算做出来的神经硬件再牛叉,如果医疗界不买账那就成了个游戏,高级版电子宠物嘛。I like the way you put it : 所以谁来做还是很重要的,grant里面的PI和co-PI都是背景非常煊赫的MD+PhD双料,会治病的和会编程的都服他们。大家也都希望这种人给领域折腾出点新风向 : 模型粗不粗取决于疾病,如果是spinal cord injury这样的硬伤型疾病,走到细胞级都过头了。对于dystonia这种主要因细胞数量减少而造成的疾病来说,细胞级+部分分子级的建模并没有硬伤。如果拿我们这些东西来研究ion channel,那就粗得有点扯淡了
| | | m******1 发帖数: 95 | 11 我们做silicon建模仿真目前还真没有published work,相关工作多是Neural
computation方面的数学理论。不过年底SfN欢迎你来看我们的poster,PI的名字叫
Terence Sanger
Co-PI是Gerald Loeb,老爷子在Spinal cord建模上几十年下来搞得巨细靡遗,自己还
开公司造机器人。我们采用他的部分模型,跑在300倍实时速度上。最近的paper你可以
看这个:
http://www.jneurosci.org/cgi/reprint/30/28/9431
【在 f**d 的大作中提到】![](/moin_static193/solenoid/img/up.png) : good,this make sense, : would you tell me the PI name who has the MD+PhD and doing this : spinal cord in silicon simulation technique? : I want to read his papers : thanks
| f**d 发帖数: 768 | 12 两年前,偶尔遇到做spinal cord neuron regeneration的大牛stephen g. waxman,
他说,从分子机理做成neuron regeneration,路漫漫遥遥无期而不可及。
我当时在听他的讲座,提出的问题就是,可不可以用电子模型电路的方式来人工修复
spinal cord受损的neuron? 从而使瘫痪病人恢复运动能力。当时Stephen看了我好几眼,
没说话,呵呵--因为如果这样做是最终可行的路的话,stephen老先生相当于做了很
长时间的无用功。
--现在看来,Terence Sanger做的正是这个东西啊。
好的,希望来拜访观摩你们的最新进展。
谢谢
【在 m******1 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 我们做silicon建模仿真目前还真没有published work,相关工作多是Neural : computation方面的数学理论。不过年底SfN欢迎你来看我们的poster,PI的名字叫 : Terence Sanger : Co-PI是Gerald Loeb,老爷子在Spinal cord建模上几十年下来搞得巨细靡遗,自己还 : 开公司造机器人。我们采用他的部分模型,跑在300倍实时速度上。最近的paper你可以 : 看这个: : http://www.jneurosci.org/cgi/reprint/30/28/9431
| m******1 发帖数: 95 | 13 :)你看看John Donoghue老爷子组的东西吧,已经做到用电子技术让四肢瘫痪的人在
跑步机上走了。你提的问题是prosthetics修复方面的问题,关于这点,做脑机接口BMI
的人几乎都在想办法用电路绕开神经通路。我们这里不是做prosthetics,而是研究病
理、用药和computational neurosci. 这样的话首先是屁股坐的位置不一样,其次对“
电路能否/是否完全仿真神经通路”这个问题容忍程度不一样。
眼,
【在 f**d 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 两年前,偶尔遇到做spinal cord neuron regeneration的大牛stephen g. waxman, : 他说,从分子机理做成neuron regeneration,路漫漫遥遥无期而不可及。 : 我当时在听他的讲座,提出的问题就是,可不可以用电子模型电路的方式来人工修复 : spinal cord受损的neuron? 从而使瘫痪病人恢复运动能力。当时Stephen看了我好几眼, : 没说话,呵呵--因为如果这样做是最终可行的路的话,stephen老先生相当于做了很 : 长时间的无用功。 : --现在看来,Terence Sanger做的正是这个东西啊。 : : 好的,希望来拜访观摩你们的最新进展。 : 谢谢
| f**d 发帖数: 768 | 14 I SEE, 从外行看来,你们做的和prosthetics 交叠挺多的
真正做起来,就是不同的两个大方向。
John Donoghue做的令人叹为观止
做计算神经,如果不能了解和清晰这几个方向的技术、算法等,在将来就处于落后地位。
现在国内做这个的有没有?
BMI
了很
【在 m******1 的大作中提到】![](/moin_static193/solenoid/img/up.png) : :)你看看John Donoghue老爷子组的东西吧,已经做到用电子技术让四肢瘫痪的人在 : 跑步机上走了。你提的问题是prosthetics修复方面的问题,关于这点,做脑机接口BMI : 的人几乎都在想办法用电路绕开神经通路。我们这里不是做prosthetics,而是研究病 : 理、用药和computational neurosci. 这样的话首先是屁股坐的位置不一样,其次对“ : 电路能否/是否完全仿真神经通路”这个问题容忍程度不一样。 : : 眼,
| m******1 发帖数: 95 | 15 我听说过的没有。
位。
【在 f**d 的大作中提到】![](/moin_static193/solenoid/img/up.png) : I SEE, 从外行看来,你们做的和prosthetics 交叠挺多的 : 真正做起来,就是不同的两个大方向。 : John Donoghue做的令人叹为观止 : 做计算神经,如果不能了解和清晰这几个方向的技术、算法等,在将来就处于落后地位。 : 现在国内做这个的有没有? : : BMI : 了很
|
|