AI的两种实现思路。
This episode is a special one: we’re talking about artificial intelligence, or AI. AI became a hot topic in 2016 because of AlphaGo, and stayed hot for more than a year. By this year, it had actually cooled off quite a bit. But two things happened recently that pulled it back into our line of sight, so let’s use the opportunity to talk about AI.
What were those two things?
First, a few weeks ago Google demonstrated its latest phone-assistant app. In the demo video, this AI-based assistant could already make phone calls for its owner using voice—and the content, the intonation, practically everything it said sounded almost indistinguishable from a real person. That really startled people.
The second thing was, of course, Luo Yonghao’s Smartisan Technology releasing a workstation called TNT, supposedly capable of putting today’s computers out of business. TNT stands for Touch and Talk—“hold your finger down, then speak.” Luo Yonghao came up with this because of the current state of AI: voice recognition, especially AI-based voice recognition, is still pretty stupid. If you simply give AI commands by voice, it tends to get things wrong in spectacular fashion. So you have to hold down a target before speaking; locking onto the target greatly reduces AI’s computational difficulty.
Naturally, the messages sent by these two events were contradictory. One suggested that AI was still pretty dumb; the other showed us that AI’s intelligence is evolving rapidly.
So what do I want to say while this news is still making waves? I want to talk about two approaches to creating artificial intelligence.
The distinction between these two approaches isn’t a technical term. It’s my own summary after reading some material. I’m certainly no AI expert, so I can’t walk you through the technical details. What I want to discuss is some very basic stuff at the level of AI’s overall framework—and I suspect that’s more interesting to everyone anyway.
So what are the two approaches to implementing AI? I call them the “outcome-simulation approach” and the “mechanism-simulation approach.”
“Outcome-Simulation Approach”
Let’s start with the outcome-simulation approach. The idea is that as long as we can simulate the things human intelligence can do at the level of the results, that’s enough. How it works internally, and whether it works the way the human brain does, doesn’t really matter.
AlphaGo, which dazzled everyone a few years ago, is essentially a product of this approach. The foundation behind AlphaGo is one of the most eye-catching AI technologies of recent years—Deep Neural Network.
At the most basic level, this technology achieves intelligence by simulating the way neurons in the human brain are organized. But don’t hear that and assume it simulates the human brain; let me keep going, and I’ll explain that in a moment.
The predecessor of the deep neural network was a program called the Artificial Neural Network model.
Here is an illustration of an Artificial Neural Network model:
I’ll explain the key features of the Artificial Neural Network model as simply as I can. This section may be a little harder than what we usually discuss on the show, so you may need to prick up your ears and listen carefully.
An Artificial Neural Network model looks like this:
As you can see in the picture, some artificial neurons in the model encode incoming signals, some output computed results, and others are hidden between input and output, handling the transmission and transformation of signals.
Each neuron in the model performs only very simple operations on the information it receives. For example, under a rule like this: if the input signal is greater than 1, output 1; otherwise, output 0.
These artificial neurons are connected to one another with different strengths. In other words, every line in the diagram has a number representing the strength of its connection.
For example, if the strength is 0.5 and the signal output by the previous neuron is 1, then the number passed along this line to the next neuron becomes 1*0.5=0.5.
The most important point about the model is that every time it takes in data, the connection strengths between neurons—that is, the numbers on those lines—change according to a rule established beforehand. In this way, AI researchers can use enormous amounts of data to adjust those connection strengths.
At first, a model may be completely unable to turn input signals into the output signals you want. But after “training” on massive amounts of data, the connection strengths between its neurons keep adjusting, and the results increasingly match the requirements. In other words, it becomes more and more “intelligent.”
When you often hear in the news that experts train an AI with several million images, this is the process they’re talking about.
That is the basic concept of the Artificial Neural Network model.
The deep neural network in AlphaGo, understood in the simplest and crudest way, is basically a super-strengthened, upgraded version of the elementary Artificial Neural Network model shown in the picture—
Its biggest feature is that it has especially many intermediate layers. That’s why it’s called a “deep” network.
Although an Artificial Neural Network model may look at first glance like the network of neurons in the human brain, that’s where the resemblance ends.
In an Artificial Neural Network model, signal transmission and the rules for adjusting connection strengths are all based on functions and algorithms, and the network’s connection structure differs greatly from that of neurons in the human brain. So rather than a simulation of the brain, it is better understood as a purely mathematical model. The way it solves the same problems as the human brain may well be completely unrelated.
That is essentially AlphaGo’s situation: its playing strength far surpasses that of humanity’s strongest Go players, but its “way of thinking about Go” has nothing to do with human players. It has merely produced the seemingly “intelligent” result of being exceptionally good at Go; its underlying mechanism is completely unrelated to human intelligence.
That is why so many people have worried in recent years, “Oh no, if AI keeps developing, will it wake up to self-awareness and then rule humanity?” Yet experts like Dr. Jun Wu of Google aren’t worried about this at all; they keep pushing back against this AI-threat theory everywhere.
Because in these experts’ view, if the outcome-simulation approach represented by deep neural networks continues to develop, then even if AI becomes a hundred times more intelligent than it is now, it will still be nothing more than an optimal algorithm for some particular problem. No matter how far it develops, it is unlikely to “think” like humans—and self-awareness is an even more fantastical matter.
But (there’s always a but), besides the outcome-simulation approach, which is currently absolutely mainstream, there is another approach that seems marginal right now but cannot be ignored: the mechanism-simulation approach.
“Mechanism-Simulation Approach”
If outcome simulation is a roundabout way of saving the country, then mechanism simulation is a head-on collision.
This approach says that if we want machines to become intelligent, we first have to understand what real human intelligence actually is. We have to clarify human intelligence, and even the core of intelligence—what consciousness essentially is—and then use that as the basis for reproducing human intelligence in a computer.
If outcome simulation is an engineer’s approach aimed at “solving problems,” mechanism simulation is a scientist’s approach. Scientists often aren’t concerned with solving practical problems; they like to “探寻本质” or search for the essence.
At present, the most important representative of this “non-mainstream” approach is probably Douglas Hofstadter, the American cognitive scientist and AI scholar.
Hofstadter is not Chinese American. His name is actually Douglas Hofstadter, and he is a professor in the School of Arts and Sciences at Indiana University. He knows a little Chinese and chose this Chinese name for himself.
In Hofstadter’s view, mainstream AI research today may be powerful, but it has nothing to do with true artificial intelligence. True AI should be built on the foundation of unlocking the mysteries of intelligence itself.
Hofstadter’s thinking about the nature of intelligence is concentrated in his monumental book Gödel, Escher, Bach: An Eternal Golden Braid, also known as GEB.
This book is extremely well known among researchers in science and engineering. Many people have bought it, but I suspect few have read it from beginning to end, because parts of it are genuinely rather obscure. A lot of people simply keep it on their bookshelves like a totem.
GEB is filled from beginning to end with profound insights into the nature of intelligence. Let’s take the core of intelligence—consciousness. If we’re worried that AI may one day possess consciousness, before worrying about that, we first need to figure out what consciousness actually is.
Every morning when we wake up, consciousness begins to perceive the world: the sunlight is lovely today, I’m hungry, I had a nightmare last night…
Seen this way, consciousness might be defined as “the brain’s reflection of the physical world and of bodily and mental experience.”
But is that the essence of consciousness? No, it isn’t.
GEB says the most mysterious thing about consciousness is that it can also reflect consciousness itself—that is, self-reflection, or self-awareness.
In an article interpreting GEB called “An Eternal Golden Braid,” Chinese AI scholar Jiang Zhang wrote:
“The core of self-awareness is the self. The self is not an external thing I observe; it is this observation, consciousness itself. So my body, my memories, and even my feelings are not identical with me. The real me is probably not all these tangible things, but the ability to experience and recognize those tangible things themselves.”
In other words, the essence of consciousness may be “self-reference” (the “self” of oneself, the “reference” of pointing to oneself—when I, a southerner, say “self-reference” in Chinese, the retroflex and non-retroflex sounds make my tongue tie itself in knots. Understanding is everything).
Here is a famous painting by Escher, extensively quoted in GEB, called Gallery:
In this painting, a young man is admiring a painting in a gallery, and within the painting in front of him there is a gallery. In that gallery within the painting, another young man is admiring the work in front of him—which means that this young man is actually living inside the painting he is looking at.
The content of the painting ultimately points back to itself. Perhaps this is the most vivid depiction of the essence of consciousness.
So, is it possible to write a self-referential program in a computer? A self-referential program is one whose output is completely identical to its behavior while running. Although a “self-referential” program in the strict sense has not yet appeared, it is entirely feasible in theory.
So does that mean that if humanity really writes the first program capable of “self-reference” one day, AI’s self-awareness will awaken at the very moment the program runs? In theory, that is possible.
The first self-referential program written by humanity may also be the world’s first artificial consciousness.
So, although this approach of simulating human consciousness inside a computer seems pretty fringe for now, if it does become possible in the future, then the concern mentioned earlier—that AI might harm humanity—wouldn’t exactly be an unfounded worry. Because AI built along this technological route could very well possess the core of human thinking: self-awareness.
If a machine far smarter than humans were conscious, then of course we’d have reason to worry about whether it would be good or evil toward humanity.
So that’s my understanding of the two ways of building artificial intelligence: one simulates the effects of intelligence, while the other tries to crack the essence of intelligence. I hope this helps you understand AI a little better.
Leave a Reply