Human learning is not knowledge, but the method of obtaining data, information, knowledge and experience; machine learning is learning data, information, and knowledge, but it will not learn and apply it lively.
There are similarities between different material systems; each subsystem of the same material system and the overall system also have similarities; material systems with different forms of motion and different properties obey similar physical laws. These facts all show that : Similarity is a basic characteristic of nature. For example, the mechanical system composed of mass-spring-damping and the circuit system composed of resistance-inductance-capacitance are similar systems, which reflect the similar relationship between physical phenomena (in general, similar relationships can be used to simplify complex systems for research ). It is easier for machines to learn and transfer such homogeneous and linear similar systems, but it is difficult to realize the analogy and conversion of heterogeneous and nonlinear similar systems. However, human learning can be based on symmetry and asymmetry, homogeneity and non-homogeneity, linearity and non-linearity, homology and non-homology, isomorphism and non-isomorphism, empathy and non-sympathy, sympathy and non-sympathy, Periodic and non-periodical, topological and non-topological, family and non-family gallop and stroll freely.
Machine learning is inseparable from time, space and symbols, while human learning is a system that changes with changes in value, facts, and emotions ; machine learning follows, follows, and relies on existing rules, while human learning is How to modify old rules, break conventional rules, and establish new rules. For example, the truly outstanding leaders and commanders are how to break the rules-reform, rather than step by step step by step steadily and steadily when the local area is going down and decayed, let alone watching the epidemic spread, but staring at the campaign and the black hat.
On March 16, 2017 , the Defense Advanced Research Projects Agency (DARPA) plans to launch the "Lifelong Learning Machines" (L2M) project, which aims to develop the next generation of machine learning technology and use it as a basis to promote the third AI Technology wave. DARPA believes that the development of AI technology has gone through the first and second wave, and is about to usher in the third wave. The first wave of AI technology was characterized by "rule knowledge". Typical examples include Windows operating systems, smartphone applications, and programs used by traffic lights. The second wave of AI technology is characterized by "statistical learning", typical examples are artificial neural network systems, and progress has been made in areas such as driverless cars. Although the above-mentioned AI technology has strong reasoning and judgment abilities for clear problems, it does not have the ability to learn, and the ability to deal with uncertain problems is also weak. The third wave of AI technology will be characterized by "adaptation to the environment". AI can understand the environment and discover logical rules, so as to conduct self-training and establish its own decision-making process. It can be seen from this that the continuous autonomous learning ability of AI will be the core driving force of the third wave of AI technology, and the goal of the L2M project is in line with the "adaptation to the environment" characteristics of the third wave of AI. By developing a new generation of machine learning technology to enable it to continuously learn from the environment and sum up general knowledge, the L2M project will lay a solid technical foundation for the third wave of AI technology. At present, the huge base of 30 performance groups is working through grants and contracts of different durations and scales.
In March 2019 , researchers from the University of Southern California (USC), a DARPA partner, published the results of exploring bionic artificial intelligence algorithms: Francisco J. Valero- L2M researcher and professor of biomedical engineering and biokinetics at the USC Viterbi School of Engineering. Cuevas, together with the college’s doctoral students AliMarjaninejad, Dario Urbina-Melendez, and Brian Cohn, published an article in the journal Nature Machine Intelligence, detailing the successful development of artificial intelligence-controlled robotic limbs . The limb is driven by animal-like tendons, can teach itself walking tasks, and can even automatically recover from balance disorders.
What drives USC researchers to develop this robot limb is a bionic algorithm that can learn walking tasks autonomously in just five minutes of "unstructured play"; that is, perform random movements so that the robot can learn Own structure and surrounding environment.
Current machine learning methods rely on pre-programming the system to handle all possible scenarios, which are complex, workload-intensive, and inefficient. In contrast, USC researchers revealed that it is possible for artificial intelligence systems to learn from relevant experience because over time they are dedicated to finding and adapting solutions to meet challenges.
In fact, for many infinite learning, it is difficult for people to achieve life-long learning. There are always some who can learn, and many others are half-knowledge or even ignorant. For the lack of "common sense" and "analogy" "For mechanical machines, lifelong learning may be a slogan!" The first thing that needs to be clarified should be: what can be learned? What can't be learned?
Human learning is omni-directional learning, learning from different angles. One thing can become multiple things, one relationship can become multiple relationships, and one fact can not only become multiple facts, but also multiple values. , What’s more interesting is that sometimes, people’s learning can turn multiple different things into one kind of things, multiple different relationships can become one relationship, multiple facts can become one fact, or even one. value. And machine learning is essentially the cognition explicit of people (one or some people). Strictly speaking, it is a kind of "self-righteous" and "yes", that is, people can only recognize things they are accustomed to or familiar with. Therefore, the limitations and narrowness of this or this group of people are unconsciously integrated into the model and the program. Therefore, this one-to-many transformation mechanism is often inherently inadequate from the beginning. Of course, machine learning is not useless. Although it is not good for intelligence, it should be good for computer or automation applications!
If the essence of learning is classification, then human learning is a method of obtaining and creating classification, and machine learning is simply a method of classification. DARPA’s "Lifelong Learning Machines" (L2M) project may be a beautiful bubble in essence. It will float up and down in the air when it blows. Although it will be colorful under the sun, it will eventually Shattered!