| Deep learning uses some simple,non-linear,multi-level representation models to transform data into higher-level and more abstract representation learning methods.At present,deep learning has developed rapidly in the fields of speech recognition,visual image recognition,and target detection,and has made great progress.Since Geoffrey Hinton and others proposed the concept of deep learning in 2006,and solved the problem of gradient disappearance in deep network training,the third artificial intelligence upsurge represented by deep learning has been set off around the world,and it continues to this day.In the past few years,Facebook,Google,Microsoft,Baidu,Tencent,and other startup companies have scrambled to conduct deep learning application research.In addition,inspired by the huge application prospects of deep learning,many countries have formulated future artificial intelligence development plans and launched a new generation of artificial intelligence research projects that cost a lot of money,trying to occupy the commanding heights of future technology.As deep learning has attracted more and more attention,society has higher and higher expectations for deep learning.This expectation is so high that it has even caused people to panic about artificial intelligence.Stephen William Hawking warned that "the development of artificial intelligence may mean the end of mankind",Elon Musk worried that the development of artificial intelligence may pose a huge threat to human survival,Bill Gates reminds people to be wary of artificial intelligence.In fact,it is too early to discuss whether deep learning threatens human beings,because deep learning faces insurmountable bottlenecks such as robustness problems,black box problems,and reliance on big data and computer computing power.One purpose of this article is to reflect on the deficiencies and limitations of deep learning,so that everyone has a rational and objective attitude towards artificial intelligence represented by deep learning.From a more macro perspective,since the concept of "artificial intelligence" was first proposed in 1956,artificial intelligence has experienced three ups and downs.The first wave of artificial intelligence was from 1956 to 1976,with abstract symbolic reasoning as the core.The biggest achievements include: proof of mathematical axioms,knowledge graphs and expert systems,but it fell into a low point due to problems such as the difficulty of obtaining and representing common sense;The second wave of artificial intelligence began in the early 1980 s,with probabilistic and statistical models as the core.Significant progress has been made in speech recognition and machine translation,but it has entered the winter due to difficulties in data acquisition and low learning efficiency.The third wave started in 2006 with the deep learning technology proposed by Geoffrey Hinton and others.After more than ten years of rapid development,deep learning is currently at its peak,and it has even begun to decline from its peak.At this time,discuss how long the third artificial intelligence craze lasts,how to start the fourth artificial intelligence craze,and more specifically,explore how to solve the major bottlenecks that plague deep learning,and how to build a safe,credible and reliable future artificial intelligence.Intelligence,how to explore a new path to promote the further development of artificial intelligence is a very urgent and very valuable subject.This is also the second purpose of this article.In order to achieve these two goals,this article has made two innovations: one is based on the research perspective of Kant’s philosophy;the other is the use of demonstration methods that combine engineering and cognitive science.Based on Kant’s philosophy,Kant’s philosophy reconciled rationalism and empiricism,and realized the philosophical "Copernician revolution".This is very important to the artificial intelligence that has suffered from the separation of the two research paths of symbolism and connectionism.It is instructive,and most of the current reflections on deep learning are based on deep learning algorithms,or based on engineering perspectives such as computers,and it is very easy to fall into the thinking limitation of "not knowing the true face of Lushan,only by being in this mountain";The method of argumentation uses a combination of engineering and cognitive science because some research paths are still in the early stages of exploration,and there are no specific research projects as arguments.In the case that the research project is not available,the author uses the research results of cognitive science as a supplementary argument.Based on the introduction and evaluation of the future artificial intelligence development plan proposed by Professor Marcus of New York University and Professor Zhang Bo of Tsinghua University,this article demonstrates the development of artificial intelligence from the perspective of engineering and cognitive science.In terms of direction,five conclusions are drawn: one is to reconcile the hybrid path of symbolism and connectionism,a relatively mature new research path;second,to build an innate framework is an important means to solve artificial intelligence’s dependence on data and improve learning efficiency;third,Establishing a knowledge base and causal reasoning is an important means to solve artificial intelligence’s dependence on data,"black box" problems,and robustness problems;fourth,establishing an internal cognitive model is important for solving robustness and establishing credible and safe artificial intelligence Means;Fifth,letting artificial intelligence understand the real world is an important means to fundamentally realize general artificial intelligence,and it is also a key means to solve the bottleneck faced by deep learning.All in all,during the transition from the third to the fourth wave of artificial intelligence,this article is based on Kant’s philosophy,reflects on the limitations of deep learning in the entire development of artificial intelligence,and explores the future research path of artificial intelligence,and finally develops Five conclusions about the future of artificial intelligence. |
PDF Full Text Request