主题：【原创】围绕脑科学而发生的若干玄想 -- 鸿乾

this is a great one, thanks.

I watched Hawkins presentation again, very interesting, keeping me coming back.

1.

first of all, you following piece is hitting the nature of 智能 right on the money:

我的理解是这样的：类比和联想其实是神经系统自然的动作 [ 鸿乾 ]

I want to copy it here, hoping not to bother admin.

2.

I watched Hawkins presentation again, he said he does not want a robot, or AI in general, he wants to replicate human cortex's 智能 in a software, and that software will be able to do a "类比和联想" automatically, and potentially in a hardware as well.

he wants to develop a cortex based 智能 paradigm for the next generation of computing of next decade, like MSFT, GOOG kind of scale and impact.

Hawkins wants something big;

3.

原手's "Moravec的悖论" actually already touched that critical difference between animals, human being' s computing power and the computing power of average "technician/computer/robots"

obviously, humans (may be some animals) can do "类比和联想", but "technician /computer/robots" as we know today can't do it, although they can do many things much faster than human beings.

obviously, Hawkins' code wants to cross that border, making his code as powerful as animal or human being's cortex, in terms of being able to 类比和联想"

4.

I have posted a few on the same subject, "类比和联想的神经基础", & "量子力学的心脏” 晓兵

I also made some point on the following paper.

"Paper也出来了，可以在这里看" [ Fuhrer ]

http://www.alexwg.org/publications/PhysRevLett_110-168702.pdf

Dr. Alexander Wissner-Gross

http://www.alexwg.org/

5.

"类比和联想": if you cannot figure it out and do it, someone will make it, the "make or fail" gap here is basically information, and/or innovation.

In the humanity's forever game of innovation (提高劳动生产率), most people will fall behind, they may never have a chance to learn and understand the name and nature of that game in the first place, all their lives, for whatever reasons.

and consequently, money/capital will be taken away from those falling behind folks relentlessly, and transferred into the hands of innovators, that is basically

modern capitalist "温水煮青蛙" 金融 model (理论 "理想"model) , under "华盛顿共识"生产关系.

"自身的知识和勇气" 的稀有性 (value): "唯一能真正推动财富增长的，就是人类自身的知识和勇气"

"汉密尔顿ABC"講金融 (3): "决战千里之外" [ 晓兵 ]

6.

the barrier between innovators and followers is more of "朗道势垒" in US and outside of china, and in china we have both "朗道势垒", and/or "毛林势垒" for you, baby, which one do you try to jump and cross? what is your talent? 朗道 type or 毛林 type?

if you manage to cross that "朗道势垒" anywhere (including in china), you will be a capitalist innovator, and money (capital) will fall onto your head from sky;

in china, if you manage to cross "毛林势垒", TG's 政治局 got a seat for you, better or worse?

obviously, 林's model missed 势垒 between "毛and 林", big time, then he got eaten alive by Mao.

So, right modeling is important(:).

7.

I have posted quite bit on Erik Verlinde's theory

of gravity and entropy, which can also be viewed as an information theory or model.

basically, using an analogy: critical information is in the system or global heatbath, it is not evenly distributed (general relativity theory, gravity), but it is there, and with Verlinde's gravity entropy model, you can kind of do some modeling or computing to look for that "critical information", and Dr. Alexander Wissner-Gross (Harvard physics PHD) AI model is a try in that direction.

8.

Hawkins wants to do it in his way, kind of a "quick and dirty" of software based replication of human cortex, typical silicon valley way of doing things.

why not? he has made a lot of progresses, with many commercial apps already making money for his company.

----------

鸿乾:

我的理解是这样的：类比和联想其实是神经系统自然的动作

而不是外部刻意追求的。

具体怎么讲？首先从反面讲，即从现有的计算技术中的记忆存储和相应的功能讲。在现在的计算机中，你存一个图像就是一个图像，你取出来，还是那个图像，非常精确，如果有所误差，你就根本取不出来。而且这个图像的记忆和对这个图像的理解一点关系都没有，记忆是记忆，理解是理解，记忆是存储体中的，理解是存储外面的软件的运行的结果。因此，这个计算技术体系中，不可能产生类别和联系，即使有，也不是自然产生的，而是外部刻意追求而加进去的。扩大了讲，就是说，基本上没有可能产生智能。

但是在脑中，完全不同。记忆的东西，是神经系统分解进而理解（这个理解的含义，请注意，有所不同）的东西。因此，在这个基础上，如果有所误差（有意的，比如说放开了幻想，或无意的，比如说醉酒），那么就自然产生有些不同的记忆提取，但是其主体又是同样的。这样就产生了类比和联想的神经基础。现在大家都比较认同，这其实是智能的最基础的属性和特征。

当然了，具体究竟神经系统是如何分解信息，进而在不自觉的程度上做了理解（含义，再次提醒），又怎么做的存储，这些都是未知的。但是，我认为Hawkins对这些方面正在做的工作是非常好的，非常有意思的。遗憾的是，整个学术界对此做了几乎全部的冷待遇。他做的对或者错都不重要，但是重要的是，是一个很好的方向。

这个Pentti Kanerva的工作，也应该对Hawkins的影响很大。他是Hawkins的研究所的成员。

这位朋友说的：底层应用字典编码倾向的表层反映，其实就是说如何分解信息等（他称为字典编码）。现在的DNN也不是没有这个分解信息。他们的分解基本上是解很大的不定线性方程组，但是加上极值限制。这种分解，有很强的数学指导，因此容易为大家接受，分解也就是线性组合，理解起来容易。但是，可以明确讲，我的观点是，这样的数学，明显不是脑中的活动，差得很远，恐怕因此效果就差了。