In the tortuous development of artificial intelligence, from the initial glory to the sudden winter, every step reflects the hardships of scientific and technological exploration and the complexity of human nature. When the neural network fell into a trough due to the death of the representative figure and the limitations of technology, symbolism quietly took over the baton of the development of artificial intelligence with its rigorous logical reasoning and rule system. However, even this seemingly stable path has not escaped the storm of doubts and funding cuts from the outside world, and artificial intelligence has encountered an unprecedented winter for the first time on its way to the 2.0 era. This is not only a test of technology, but also a profound call for cooperation and understanding in the scientific research community.
Today is our second chapter, section 3, part 2: Logic Genius: The Fiery Symbols and the Frozen AI
The death of the representative figure and the failure of the perceptron quickly brought the research on neural networks to a low point.
In 1974, Paul · Warbers of Harvard University, in his doctoral dissertation, first proposed training artificial neural networks by error backpropagation (BP). The main idea of the BP algorithm is to backpropagate the error (gradient) signal, adjust the parameters in the neural network, and update the model weights to minimize the prediction error.
Although this won Warbers the IEEE Neural Network Society's Pioneer Award, it was not so easy to "break out" in the context of the suppression of connectionism at that time. In the five years that followed, the direction of neural network research became vague and progress was slow.
In 1988, when the influential Minsky republished his book Perceptron, he deleted the sentence in the first edition that personally attacked his younger brother Rosenblatt, and also wrote the discourse "In memory of Rosenblatt" himself.
At that time, Rosenblatt had been dead for a long time, and Minsky's apology was long overdue. And the scientists who have been suppressed by Minsky agree that Minsky is unforgivable. When connectionism made a comeback, these people vented all the grievances they had held back for ten years. This is called "treating others as they would have done their own thing."
From the publication of the book on perceptrons to the stagnation of connectionism in the past few years, people have turned to symbolism, just as the so-called symbolism stands on the shoulders of the pioneers of the artificial intelligence 1.0 era, ushering in a vigorous "spring".
· Search-based reasoning
Search reasoning is a method of reasoning in the field of artificial intelligence that finds an answer or solution to a problem by systematically searching for various possible solutions. We can call this process a "search tree".
Search-based reasoning is an extension and development of the "symbolism" of Sima He's master and apprentice. The search algorithms involved include depth-first search, breadth-first search, and heuristic search. Heuristic search is one of the ways for researchers to further improve the efficiency and accuracy of search-based reasoning.
Richard · Dahl and Peter ·Hart proposed the "A* algorithm", which is a commonly used heuristic search algorithm for solving problems such as path planning and graph search. In summary, the operation of search-based inference is to clearly define symbolic structures and rules.
The basic idea is as follows:
As you can see from the above, the advantage of search-based reasoning is that it has a wide range of applicability and is very systematic, allowing for a comprehensive consideration of all possible solutions, as well as a step-by-step search and evaluation. It's a lot like generative AI today. Of course, the advantages of this approach can also be its disadvantages.
It is difficult to fully cover all possibilities for complex problems. Even if we could cover so many candidates, it would take us a lot of time and resources to screen and evaluate. But even so, it does not affect that search-based reasoning is one of the core concepts and fundamental theorems of symbolism.
Of course, it would be fine if it were left until now. If you are not satisfied with the results given by the AI assistant, you can continue to ask for your requirements, and even leave the task of screening and evaluation to it as well, the key is that you have to master the skills of asking questions, which we will talk about in practice.
· SHRDLU
SHRDLU was developed in 1970 by MIA's Terry · Winoglag.
It is a natural language processing program that is able to understand and respond to instructions entered in natural language, thus manipulating virtual objects made up of blocks of various shapes. Its disadvantage is that the domain is too limited, the processing power is too low, and in the final analysis, it is actually a database problem.
For example, you can tell it, "Put the red square on top of the blue cylinder", and it will understand your instructions and perform this in a virtual environment. If you ask him to put the red square in his bag, he won't know how to respond to you.
· Expert system
An expert system is a computer system that mimics the decision-making ability of human experts. Answer questions in a particular area based on logical rules derived from expertise. The expert system consists of several subsystems, a knowledge base, an inference engine, and a user interface.
The founder of the expert system should be Edward · Feigenbaum, a professor at Stanford University in United States, who spearheaded the development of the DENDRAL system, which is the world's first relatively successful expert system. The DENDRAL system infers the molecular structure of a compound based on its molecular formula and mass spectrometry data.
It has the advantage of identifying and describing the structure of compounds through reasoning and search techniques, combined with a large amount of chemical knowledge and rules, but the disadvantage is that the application field is single, and it is highly dependent on predetermined rules, and the learning ability is not very good, and it cannot effectively deal with information with uncertainty or ambiguity.
However, even so, it is undeniable that it has reduced the burden on researchers to the greatest extent and laid the foundation for the development of machine learning and knowledge reasoning techniques in the future. More importantly, its success proves the value of symbolism in complex knowledge-intensive problems.
Time flashed to the seventies, and he was still this buddy.
GIVEN DENDRAL'S LIMITATIONS, HE AND A GROUP OF COLLEAGUES, INCLUDING JOSHUA · LEDBERG, DEVELOPED MYCIN, A MEDICAL EXPERT SYSTEM THAT BEGAN CONSTRUCTION IN 1972 AND TOOK SIX YEARS TO COMPLETE, WRITTEN IN THE INTER LISP LANGUAGE.
The MYCIN knowledge base has more than 200 rules, which can identify 51 kinds of germs, correctly handle 23 kinds of antibiotics, and use knowledge expression forms and reasoning rules to support machine perception and judgment of medical consultation results, and output conclusions with high accuracy of disease inference. It is mainly used to diagnose and treat patients with bacterial infections.
For example, one rule of MICIN might be: "If a patient has a fever and cough, there is a 95% probability that they will be diagnosed with tuberculosis." In its mode of operation, information such as patient symptoms and clinical signs is entered, and inference rules and probabilistic algorithms are used to find a set of diagnoses with the highest probability.
However, MYCIN still has limitations, unlike DENDRAL, there is no general theoretical model from which the MYCIN rule can be derived. What does that mean? Pure "handmade (input)". Even so, MYCIN has once again demonstrated the potential of expert systems to deal with complex issues and provide professional advice.
In 1973, United Kingdom applied mathematician James · Lighthill presented a report to the United Kingdom Research Committee, which assessed the current state of artificial intelligence research and came to pessimistic conclusions. How pessimistic? He twice called AI a "mirage". The opinion was immediately uploaded to the royal family, and the United Kingdom began to cut funding for AI research.
Also this year, the Defense Advanced Research Projects Agency (DARPA), a key agency for United States promoting AI research, also began to significantly reduce its funding. Its previous $3 million annual grant to the Carnegie Carnegie Program for Natural Language Systems is also gone. The incident dealt a major blow to confidence and funding flows across the AI field, and by the following year, funding for AI projects in the United States was already hard to find.
In 1977, the machine learning system developed by Hayes · Roth and others based on logic made great progress, but they could only learn a single concept and failed to put it into practical application. If you can't put it into practice, whether you are connectionist or semiotic, it will become a "utopian".
Although the expert system is improving, this achievement is far below investors' expectations for the development of artificial intelligence, and it is even more difficult to hedge against the AI cold wave caused by connectionism. At the same time, utopian ideas in the AI community are on the rise, leading to massive cuts in AI budgets around the world. In the 2.0 era, artificial intelligence has entered its first cold winter since its inception.
Together, there are two benefits, and fighting is two injuries. And how many people really understand? Due to the defects of human nature, people tend to be different, do not conspire with each other, and even poke each other with knives, and the knives are red. The same applies to AI.
Looking back at the development of artificial intelligence, from the silence of neural networks to the short-term prosperity of symbolism, to the advent of the AI winter, every step is full of challenges and reflections. However, it is these setbacks and reflections that have laid a solid foundation for the subsequent development of AI. Today, with the continuous breakthrough of technology and the wide expansion of applications, artificial intelligence is reshaping our world at an unprecedented speed. In the future, only by adhering to the spirit of openness and cooperation and jointly addressing challenges can we promote the sustainable and healthy development of AI technology and make greater contributions to the progress of human society.