Since AlphaGo beat Lee 乭 last year, deep learning has become hot. But no one seems to know how it works, just think of it as a black box. Some people say that deep learning is a nonlinear classifier [1]? Some people say that deep learning is a simulation of the human brain [2]… But I don’t think I even got through that.

After reading Jeff Hawkins’ On Intelligence [3], “This is it!” . I was surprised to find that the original book was published in 2004! I was afraid THAT I had read a fake book or one of the works of American civil science, so I went to Douban and Zhihu to check, and found that almost everyone who had read the book praised its theory. Strangely, no one seems to be standing up for it. The theory stops there, as if everyone is trying to hide the fact that they have read the book. It clearly explains the entire workings of human intelligence. Note Real Intelligence, not just Artificial Intelligence!!

Three insights

The author’s insight is much simpler and deeper than most papers on brain science:

  1. For a long time, people have equated “intelligence” with “acting intelligently” because of our inability to observe the mind from the inside. But when we read a book, it doesn’t look like anything has changed, and we know that we are generating countless associations, epiphanies, memories. Therefore, “understanding” cannot be measured by external behavior, it is an indicator of internal measurement.

  2. Does nature design a system of intelligence for each creature, from paramecium to man, or does nature follow a system of intelligence, or does nature come up with a new system of intelligence from a certain generation and use it today (what species first developed this system?)? ? Is what we mean by intelligence unique to humans, or is it a universal (but very diverse) feature of all living things? The author believes that intelligence cannot be designed by God for human beings, but must come from some idiomy of nature.

  3. The cerebral cortex, structurally and functionally, has the same structure/mechanism (not strictly an insight of the author, but discovered by Vernon Mountcastle in 1978).

These three insights lead naturally to the following questions:

  1. If intelligence is not defined by behavior, how is it defined?

  2. Looking ahead, how did intelligence evolve?

  3. Looking inward, how does the structure of the cerebral cortex capture the structure of the world?

In a nutshell, the authors conclude:

  1. Intelligence is not as mysterious as people think, it is simply “the ability to predict the future”.

  2. The essence of these predictions is nothing but a by-product of “biological stress” under “biological self-balancing mechanism” & “environmental pressure”.

  3. At its core, intelligence is something “stable and unchanging.” And this is thanks to the homogenous hierarchical structure of the cortex.

Let’s take a look at how the author extrapolated the nature of intelligence from those three simple insights.

Swing of life

From the small human body to the large economic system, there is a similar mechanism in complex systems to reduce shocks and return the system to steady state. Low blood sugar concentration, glucagon secretion will increase, pull hyperglycemia; And when blood sugar is high, insulin secretion will increase, pulling down blood sugar. By regulating these hormones, the system struggles to keep blood sugar within a certain range. This homeostasis mechanism appears in every corner of life, maintaining the homeostasis of life [4].

It’s like an “invisible hand,” always trying to push the squeeze away while grabbing the “deserter” back. This “invisible hand” weaves countless “right places” (” constant representations “in brain scientists’ slang) into our brains. When we deviate, we become alert and mobilize multiple systems to respond. Take the example in the book, when a ball comes to us, we don’t calculate its trajectory and landing point, but direct the body to adjust accordingly until we catch the ball. This adjustment algorithm is called prediction. In this case, what is the essential difference between the intelligence of the human being in catching the ball and the stress of the paramecium’s movement towards the food?

Why prediction is the foundation of intelligence?

Normally, people understand the “prediction” step is too big, like a serve from the precise calculation of its landing point, but the human brain “prediction” is more like “stress”, the ball moves, a little bit of fine tuning. Modern society moves so fast that we can’t see the historical face of concepts, and are therefore more easily confused by the mists of appearances. When we go back to the beginning of history, the fog clears naturally. What is the greatest benefit of intelligence to us? Not to create, but to survive. Human beings are constantly torn between survival and development. But few people see that development is nothing more than a response to the unknown challenges of survival.

How should we define intelligence? Perhaps evolutionary history can tell us more. Intelligence is the ability to help us survive: the ability to spear a fish swimming in a stream, the ability to tell a friend from a beast based on a blurred image… We should be looking at things like “how to keep balance”, not solving ballistic problems. That’s not nature’s evolutionary goal, and nature doesn’t have any brain mechanisms.

All survival problems boil down to a meta-problem: how to identify the constants in the problem. Such as: fish in the stream, the direction home… If there are any other components of intelligence, such as imagination, tool creation, problem solving, they can be reduced to some abstract means. In the final analysis, there is only one way for human beings to solve all problems — to use abstraction, to elevate and contradict in higher dimensions.

No escape from “invariable representations”.

Nature of abstraction

Just as after people recognized the concept of “negative number”, they could unify the operation of “addition” and “subtraction”, which are completely different from each other in appearance (one increase, one decrease), into “addition in the integer domain”. Reconciling contradictions from higher dimensions is exactly how the cerebral cortex is constructed and how it works. Constantly finding common ground in phenomena, extracting it, and giving it a name; These names in turn become the building blocks (or “words”) for the next level of abstraction. And so on, layer by layer, until you get to that holy grail of intelligence, constancy.

For example, how do we recognize edges [5]?

We first examine a small 3×3 retina labeled #1 to #9 (as shown below). When a vertical line appears (#1, #4, and #7 are all activated), the electrical signal is transferred to the second layer. Each neuron in the second layer responds separately to the activation of a group of cells in the retina. For example, the second layer, on the far left, responds to the activation of individual retinal cells. Another example: the second layer of neurons on the left responds to the activation of any two retinal cells. And so on…


Edge recognition: Retinal cells at the bottom; When a combination of retinal cells is activated, it activates corresponding neurons in the layer above it. And when a combination of neurons in the next layer is activated, that chain activates the neurons in the next layer

If we take time into account, and assume that the signal does not disappear immediately but decays over time, the input of (#1, #4, #7), (#2, #5, #8), and (#3, #6, #9) will activate one of the neurons in the third layer, representing “finding a vertical line”, as long as it is short enough.

See, each neuron is actually a “word” (or “concept”/” abstraction “/” feature “). But the “words” described by the lower neurons are less abstract. For example, the #(1, 4, 7) neuron in the second layer represents a vertical line on the far left side of the retina, while the upper layer has no restriction on the far left side of the retina.

The role of Memory

Neurons can gather, integrate and output information in just five milliseconds, equivalent to 200 operations per second. Humans can recognize images and make choices in half a second (equivalent to 100 steps)… 100 steps. A machine can’t do that. Of all the algorithms known to man, perhaps the only one that can do this is tabulating. So the entire cortex is a memory system, not a computer.

What is deep learning doing right?

A multi-layer network provides a layer – by – layer abstraction channel. Today, image recognition systems do exactly that: edge recognition at the bottom, then a specific shape, then a feature at the top…

Convolution provides a means of obtaining a constant representation.

What else don’t we know?

When we want to retrieve a memory, all we need is a few words. That is, memories seem to be stored in a holographic form. Any fragment contains the whole.

And we still don’t know how the brain makes decisions in 100 steps. We don’t know why there are so many feedback links, right? What are the functional differences between axon v.S. dendrites? …

Now let’s go back to the author’s three insights, to be repeated in slang:

  • Understanding is an internal measure of how the brain forms memories and uses them to make predictions.

  • Prediction is a byproduct of some self-regulating mechanism.

  • The cerebral cortex is remarkably homogenous in appearance and structure. In other words, the cortex uses the same computation to perform all of its functions. All the intelligence (sight, hearing, movement…) that human beings exhibit. It’s all based on a uniform set of algorithms.

Man is getting closer to the essence of the world, where matter is just a vehicle for information patterns. Anything other than the human brain is just a fleet of supplies for that purpose.

  1. For details, see Chapter 30 of the Beauty of Mathematics (second edition) written by Teacher Wu Jun. ↩

  2. Teacher Wang Chuan wrote about deep learning, this may be the most easy for you to read into the popular science paste ↩

  3. The original title was On Intelligence. The Chinese translation by Overseas Chinese Publishing House is the Age of Intelligence. Although the content is well translated, the title is a bit of a gimmick. ↩

  4. If you want to trace the mathematical basis of this idea of homeostasis, you have to go back to the early days of the discovery of cybernetics. For this history see Chapter 2 of Thomas Rid’s Rise of the Machine: a Lost Cybernetic History. ↩

  5. What needs to be specially explained here is: in reality, the human eye does not recognize the edge of the object according to the following principle, but a very clever, but also very simple mechanism. It is mentioned in Chapter 36 of Feynman’s Lectures on Physics, Volume 1. ↩