Shulou(Shulou.com)11/24 Report--

On weekdays, the world looks like a well-managed railway: things operate according to laws that we humans understand and can apply. We can accept occasional delays, which represent exceptions to these rules. But sometimes we think of what the world is going through as a multi-car collision on the road. Although the same physical or social laws are followed in this case, there are so many units of change that we cannot predict or explain the details of each collision-details that can cause only minor damage to one car, while another car explodes into a fireball.

The characteristics of an orderly train running along the track in a car accident are also applicable to walking on a leafy path on an ordinary autumn day. They are all events in which the interdependence of countless details is better than the explanatory power of the rules that determine them. All we can do seems to be resentful or amazed at a result. Now, our latest paradigm technology, machine learning, may reveal that the everyday world is dominated more by chance than by rules. If so, it is because machine learning can jump out of the pattern of human cognition and sum up laws that we cannot understand or apply.

The opacity of machine learning concept map machine learning systems has caused serious concern about their credibility and biased tendency. But the fact that they do work may give us a whole new understanding of what the world is and what role we play in it. Machine learning works in a fundamentally different way from traditional programming. In fact, traditional programming is the aggregator of our understanding of the world based on rules. To take the most representative example of machine learning: if you want to write software that recognizes handwritten numbers, programmers will traditionally tell the computer that "1" is made up of a vertical line, "8" is made up of a larger circle and a smaller circle above it, and so on. This method may work well, but its platonic ideal of handwritten numbers means that the program will misjudge a considerable proportion of handwritten numbers. Because the actual figures are made by mortals, it is impossible to be so "perfect".

A new mode of operation of machine learning the machine learning model knows how to learn from examples. In order to create a machine learning model that can recognize handwritten numbers, developers will not tell computers anything we know about the shape of numbers. Instead, developers provided it with sample images of thousands of handwritten numbers, each different and correctly marked as the number it represents. The system uses algorithms to find the statistical relationship between pixels with the same label image. A series of pixels on some vertical line will increase the statistical weight of the image as "1", reduce the probability that it is "3", and so on.

Unknowable but effective in machine learning applications in real life, the number of possible answers is hundreds of millions, the amount of data to be considered is very large, and the correlation between data points is so complex that we human beings usually can not understand it. For example, human metabolism is a series of extremely complex interactions and interdependent effects. Therefore, a machine learning system called DeepMetab, which can predict the response of human system to complex factors, has been created. It has become a place for doctors, researchers, non-professionals and hypochondriacs to ask questions about human organs and explore relevant ideas. Although we cannot understand how it produces output, DeepMetab remains the most important source of knowledge about the human body.

The combination of AI and Healthcare as we become more and more dependent on machine learning models (MLMs) that we do not understand, we may gradually accept the following two points of view: the first point of view is that in order to obtain the useful probabilistic output generated by the machine learning model, we must often tolerate this disadvantage of unexplained. The second view is that the difficulty to explain is not a weakness, but a real situation. Machine learning models are effective because they are better at reading the world than we are: they produce cognition beyond human beings by counting vast amounts of interrelated data without having to explain to humans how they got it. Whenever a citizen or regulator cries out in despair because he cannot understand how machine learning works, we can feel that these models do work.

Big data's concept map if the machine learning model works by abandoning the use of understandable rules to simplify and explain complexity, then it works! In the voice, we can feel that all small things interact with each other in their interdependence. And these small things are the real essence, and they rattle in the cosmic sound of harmony and law. The success of our technology is telling us that the world is a real black box.

Man-machine games from watches to cars, from cameras to thermostats, machine learning has been deeply embedded in our daily life. It is used to recommend videos, try to identify hate speech, guide vehicles, control the spread of disease, and is crucial to mitigating the climate crisis. It is not perfect and may magnify social prejudice, but we continue to use it because it is effective. Machine learning to do all this work without applying rules to specific things is surprising and even disturbing. We prefer rules to individual cases so much that we think it is crazy for a machine learning system to play go simply by analyzing a large number of games and moves without knowing the rules. But this is how machine learning becomes the best go player in history. In fact, when developers provide a system with data related to a domain, they usually deliberately hide the interrelationship between the data we already know from it.

Overly specific generalization? Now, even people who know a little bit about machine learning will feel creepy, because machine learning models are created by generalizing from data. For example, if a machine learning model of handwritten digit recognition is not generalized from the samples it learns, it will become a failed model because of over-fitting. However, the generalized description of machine learning model is different from the traditional generalization that we use to explain specific situations. We like traditional generalizations because (a) we can understand them; (b) they can often derive deductive conclusions; and (c) we can apply them to specific situations. However, (a) the generalized descriptions of machine learning models are not always easy to understand; (b) they are statistical, probabilistic and mainly inductive; and (c) unless we run the corresponding machine learning models, we usually cannot apply these descriptions.

In addition, the generalized description of a multi-level machine learning model may be very specific: for example, vascular patterns in retinal scans may predict the onset of arthritis, but only if 50 numerical indicators are met. and these 50 indicators may be interrelated. It's like you want to know how a car avoids serious damage in a multi-car collision: the vehicle has to overcome a lot of specific conditions, but this kind of event can't be summed up as an understandable rule. this complex rule cannot be migrated and applied to other events. Or it's like a clue in a murder that indicates the killer, but only in this one case.

The cue wall machine learning model does not deny the existence of rules or laws. It simply emphasizes that these rules alone are not enough to understand what is happening in our complex universe. Accidental details interact with each other, making the interpretation of the rules inadequate, even assuming that we can know all the rules in the world. For example, if you know the laws of gravity and air resistance, as well as the mass of the coin and the earth, and you know how high the coin falls, you can calculate how long it takes for the coin to land. This is usually enough to meet your practical purposes. But the traditional western scientific framework puts too much emphasis on rules. To fully apply these rules, we must know each factor that affects the fall, including which pigeons disturb the airflow around the coin and the simultaneous influence of the gravitational pull of distant stars. Do you remember adding the influence of distant comets? To apply these laws completely and accurately, we must have a comprehensive and unreachable knowledge of the universe like the Laplace demon.

The coin flip is not a criticism of the pursuit of scientific laws or the practice of science. Science is usually based on experience and is sufficient to meet our needs-although the actual achievable precision will lead us to make some concessions. But this should make us wonder: why does the western world regard the inexplicable chaos as a pure appearance and think that there are laws under it that can explain it? Why do we ontologically prefer immutable things to constantly flowing water or dust?

Rewriting the definition of knowledge these are common topics in the history of western philosophy, far beyond the scope of this article. But there is no denying that we are attracted to the world simplified by the eternal law, so we can understand the world and thus predict and control it. At the same time, these simple and wonderful laws hide from us the confusion of a particular situation, which is determined not only by the law itself, but also by the state of each other particular situation. But now, we have a prediction and control technology, which comes directly from many small factors that exist at the same time as a whole and influence each other. This technology gives us more control, but it does not improve our understanding. Its success makes us focus on things that are beyond our understanding.

The laws of physics at the same time, for the same reason, machine learning may break the obsession with certainty as a symbol of knowledge, because the result of machine learning is probability. In fact, a completely certain result from the machine learning model will cause people to doubt the model. The probability of the output of machine learning is inherently inaccurate; the real statement of probability is that it can correctly predict its error probability.

The butterfly effect now, we have a mechanism that shakes us, models that draw information from the many details connected to each other in an incomprehensible, subtle network. Perhaps we do not need to regard those chaotic swirls as mere appearances that have not yet been thoroughly understood. Perhaps the complexity and cognitive difficulty of the interaction between all factors will shake the cognitive basis of western science, that is, the most real is the most fixed, the most common and the most knowable.

The chaotic solution of the three-body problem suggests that perhaps we will eventually accept that the complex associations, accidents and coincidences of simple events are the true face of the world. We will also accept that a 1.4-kilogram brain is not enough to build a complete understanding of the world. The cruel unknowability of the world is blurring the boundaries of our understanding. If this is happening, it is because we hear more special, tiny, noisy signals through models such as machine learning. These signals are producing useful, amazing, probabilistic knowledge based on incomprehensible connections between things.

Author: David Weinberger

The cloud opens and the leaves fall.

Revision: miss circle pi

Original link: Learn from machine learning

This article comes from the official account of Wechat: Institute of Physics, Chinese Academy of Sciences (ID:cas-iop), author: David Weinberger

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.