From Molecule to Metaphor: A Neural Theory of Language

Thought is structured neural activity.

Language is inextricable from thought and experience.

Neural computation involves continuously finding a best match between the inputs and current brain state, including our goals.

2020-05 / 07 / 16:47:33 / embodied mind


Thought and language are neural systems. They work by neural computation, not formal symbol manipulation.

注意这里的 embody 指的是”与身体结合”

2020-05 / 12 / 19:01:21 / 信息立足于可感知的差异

Information in the technical sense, and therefore information processing, is based on detectable differences.

2020-05 / 13 / 14:07:28 / functionalism

Philosophical functionalism holds that everything important about language and thought can be understood completely using information processing models, without looking at the brain at all. An even stronger position claims that any information processing system of sufficient complexity will automatically have all of the mental powers of the mind, including consciousness. This stance is also called strong artificial intelligence—-there is nothing to the mind but abstract information processing.

2020-05 / 13 / 14:15:18 / 分子——磁铁和弹簧

My best physical metaphor for this is to imagine elaborate abstract sculptures made of strong springs and magnets. Fitting two of these contraptions together could cause a total reconfiguration of shape and might even cause some pieces to break off and take on a new shape. This is what happens in molecular binding, but the forces involved are much stronger relative to the size of the molecules than could actually be built with magnets and springs. The changing shape of molecules is the basis for all animal action. Think of the forces exerted by leaping whales or frogs to get a feeling for the power involved. A frog can jump twenty times its length—-equivalent to a 100-foot standing broad jump for a person.

2020-05 / 13 / 14:47:49 / 神经元投票和学习

If that were all there was to it, the information processing would be easy: if the positive votes outnumber the negative votes by a big enough margin (called a threshold), then output a signal. But not all input votes get equal weight in this election; some count more than others. And, as in the United States, only a fraction of voters usually participates. There is also a time limit, so the neuron needs to compare the weighted sum of positive votes against the weighted negative votes arriving in an interval. If the positive total outweighs the negative total by a big enough margin, a signal is sent out.

Compelling evidence exists that learning depends on the change of synaptic weights; this modification is realized through several kinds of chemical changes involving protein synthesis at both the transmitting and receiving sides of the synapse.

2020-05 / 13 / 18:49:08 / 视觉系统察觉变化和模式

For example, one way the visual system avoids being flooded by data is by only transmitting changes in a scene. We have all experienced this phenomenon in hearing—-a repetitive noise is dis- turbing for a while and then we stop noticing it.

The visual system has cells (as our simulation has people) that take signals from a collection of feature detectors and send an output if there is some pattern of interest. For example, a center-surround cell can signal when it sees a green center surrounded by red. A vertical-edge cell can signal when the detectors on its left report more brightness than those on its right. There are also intermediate cells that use memory to compare signals across time and report motion.

2020-05 / 18 / 19:28:52 / 学习改变脑部结构

When the strength of connections between neurons is modified, we have a funda- mental structural change. The neural network is now different and will respond differently to new experience. That is, learning does not add knowledge to an unchanging system—-it changes the system.

2020-05 / 18 / 19:30:16 / 基因图灵机

In protein synthesis it matters not just which genes are involved, but also which are expressed, that is, actually used to make proteins. Every cell has the same genetic informa- tion; other factors determine exactly which proteins are synthesized. The modern view of this process is more like a computer program than a stencil. Often the role of one protein is to facilitate or block the expression of genetic material in the synthesis of some other proteins.

2020-05 / 18 / 19:38:28 / 神经系统如何布线

Neurons are also cells and, in early development, behave somewhat like an amoeba in approaching and avoiding various chemicals. But rather than the whole cell moving, neural growth involves the cell's connecting pathway (axon) reaching out toward its downstream partner neurons. The basic layout of visual and other maps is established during devel- opment by billions of neurons each separately following a pattern of chem- ical markers to its predestined brain region and specific subareas within that region.

2020-05 / 18 / 19:55:45 / 化学梯度和行为强化

For the task of wiring up visual maps, gradient following greatly simpli- fies nature's information processing requirements. The axon of a develop- ing visual neuron will follow the gradient of a marker chemical related to its destination, analogous to the amoeba's moving in the direction of food. A neuron destined for the upper left section of the scene in a visual map will therefore follow two separate chemical gradients that mark areas cor- responding to leftness and to higher elevation in the target map.

In fact, the initial wiring is only approximate and leaves each neu- ronal axon connected to several places in the neighborhood of each of its eventual partner neurons. A second, activity-dependent mechanism is required to complete the development process.

2020-05 / 18 / 20:41:52 / 婴儿的神经元更多

The initial chemical wiring actually produces many more connections and somewhat more neurons than are present in adult brains. The detailed tuning of neural connections is done by eliminating the extra links as well as strengthening functional synapses. This has been known for decades.


2020-05 / 18 / 20:43:54 / 如何训练视网膜

However, there is no vision in the womb. Recent research shows that systematic moving patterns of activity are sponta- neously generated prenatally in the retina. A predictable test pattern, changing systematically over time, provides excellent training data for tuning the connections between visual maps (Stellwagen & Schatz 2002).

2020-05 / 18 / 20:58:58 / 记忆

We now know that this kind of short-term memory depends on ongoing electrical activity in the brain. You can keep something in mind by rehearsing it, but this interferes with your thinking about anything else.

These long-term memories are known to be based on structural changes in the synaptic con- nections between neurons. ... , and this can take several hours.

In addition to bridging the time gap, the brain needs mechanisms for converting the content of a memory from electrical to structural form. (写硬盘和写内存的区别)



There is general agreement and consid- erable evidence that dreaming is important in consolidating memory and involves simulating experiences.

2020-05 / 18 / 21:12:09 / 衰老的步伐

Participants who had been given the words related to old age walked significantly slower than participants who had done a version of the word jumble task that did not prime the elderly stereotype. This result sounds unlikely, but the finding has since been repli- cated by a number of labs. Thus, it appears that unconsciously activating a concept may influence motor processes (e.g., walking) related to that concept.

#read #neuroscience