The book "Questioning Artificial Intelligence-From Cambridge to Beijing" starts from the origin of artificial intelligence, tells the development and limitations of artificial intelligence, and proposes the future development direction of artificial intelligence-the concept of human-machine fusion intelligence. Human-computer integration starts from the three aspects of human, machine, and environment. It is the interactive integration of the human-computer environment (some people say that consciousness is experience or experience, in fact, consciousness is the product of interaction). The author talks about intelligence and logic from a philosophical level. The relationship between, as well as the influence of formalization and intentionality on artificial intelligence, puts forward the difficulty of human-machine integration-the concept and framework of deep situational awareness. By telling the application of ergonomics in education, military, intelligent communication and other directions, as well as exploring the current technical limitations of robots, the author has a detailed interpretation and interpretation of artificial intelligence from many aspects such as philosophy, ethics, cognition, and emotion. .
At present, although artificial intelligence is gaining momentum, there are still many difficulties, such as multi-modal information fusion, small sample learning, logical/non-logical reasoning, intuitive decision-making and logical decision-making, etc. Today's artificial intelligence is basically a formal factual calculation based on mathematical means, but lacks intentional value calculation, which means that the current artificial intelligence is more inclined to automation, and it is only by human programming to achieve its own Function, lack of autonomy or initiative, which is far from the expected intelligence in our hearts. At present, there is still a big gap between human and machine intelligence. For machines, there is a huge gap between the decision-making method based on rule conditions and probability statistics and the judgment mechanism based on emotional dysfunction and epiphany meditation. The selection and judgment mechanism of post-movement and epiphany meditation is unique to human beings. For example, from the perspective of internal origin, humans are motivated, but machines do not have it. However, it is difficult to characterize the motivation of machines. Another example is common sense (the high probability of many mixed things). It is easy for people to have common sense, but it is very difficult for machines to form. There are also decision-making. Human decision-making is divided into three parts: rational decision-making, descriptive decision-making and natural decision-making. If machines want to make “intelligent” decision-making, they need to integrate these kinds of decision-making modes of human beings in different times or situations. Establish different decisions under pressure. Therefore, the author proposes that human-computer fusion intelligence is the future development direction of artificial intelligence.
Human-computer integration intelligence originated in the two fields of human-computer interaction and intelligent science. Humans have the advantages of humans and their weaknesses, while machines have the advantages of machines, but they also have their shortcomings. Humans are not powerful in computing, but humans can break logic and use intuitive thinking to make decisions. Although machines have calculations but no calculations, they can detect various signals that humans feel cannot detect. Therefore, humans can process the subjective information that they are good at containing value orientation, and the machine can calculate the objective data related to rule probability statistics that they are good at. When intentional value is needed, it is processed by humans, and when formal facts are needed, it is shared by machines. . Only in this way can we better realize artificial intelligence by combining the advantages of humans with the advantages of machines and learning from each other.
Therefore, the author proposes that human-computer fusion intelligence is a new type of intelligent system produced by the interaction of human, machine, and environmental systems. At the intelligent input terminal, it combines the data collected objectively by the device sensor with the information and knowledge that people subjectively perceive to form a new input method; in the process of intelligent data/information intermediate processing, machine data calculation and human The information cognition (calculation) is integrated to construct a unique way of understanding: axiomatic reasoning + non-axiom reasoning mechanism; at the intelligent output end, it matches the result of machine logic operation with human value (intuition) decision-making to form Probabilistic, regularized, and inspired organically coordinated optimal judgments.
Here, the author has repeatedly emphasized the top priority of artificial intelligence-deep situational awareness. Starting from the existing concept of situational awareness, the author tells the enemy of situational awareness, and then leads to the concept and theoretical framework of in-depth situational awareness. Endsley's situational awareness is divided into three levels: the first level is the perception of each component in the environment, that is, the input of information, the second level is the comprehensive understanding of the current situation, that is, the processing of information, and the third level is the subsequent situation. Forecasting and planning, that is, the output of information. The author of this book has summarized 8 enemies of situational awareness in the book. In this regard, I have made a corresponding summary of each and put forward my own understanding, as follows:
The first is the tunnel effect of attention. I think this emphasizes that the perceiver is too focused on the aspect he is concerned about, while ignoring other global factors, which leads to a deviation in situational awareness. Therefore, in the process of situational awareness, not only limited concentration is required, but a comprehensive perception of the entire situation; the second is the unavoidable memory bottleneck, people often have some unavoidable mandatory memories, such as those complicated ones that must be remembered. Air traffic instructions. Too much reliance on human memory makes it easy to make mistakes in situational awareness, and most of the current systems are designed based on human memory. Therefore, considering that the part of human compulsory memory should be reduced in the system design, it may be better to perform in-depth situational awareness; the third is workload, anxiety, fatigue and other pressures, which will occupy a part of situational awareness to reduce human limitations. People’s ability to collect information effectively under pressure will be worse, which affects situational awareness; fourth is data overload. The author mentioned that humans can only receive and process a limited amount of information at a time (such as Miller). Theorem 7+-2), if we cannot arbitrarily change the human feeling and the size of the receiving information channel provided by the information processing mechanism, then we can reduce the impact of data overload on situational awareness by changing the rate of receiving information; the fifth is dislocation , Which means that in the process of designing the system, it is necessary to reduce the interference factors that attract attention elsewhere to prevent interference with situational awareness and generate deviation; the sixth is complexity creep, which emphasizes the excessive complexity of the system design It will affect the use. The overly complex system weakens and reduces people’s ability to obtain information, leading to wrong situational awareness. Seven is the wrong mental model. We often use experience psychology to deal with problems, and experience psychology often leads to errors. By analogy, we can use the standardization and effective use of automation models to minimize the occurrence of such errors; eight is that people are not abnormal in the loop. If they rely too much on the automation of the finished product, they will not be able to perceive loop failures in time, and automation will allow People leave the loop of control system functions, which will also lead to deviations in situational awareness.
Through the above analysis, we can see that people cannot leave the loop, so as not to be able to sense system failures in time, otherwise, when the system fails to switch back to manual mode, people often do not know which step the process has taken. And the system cannot rely too much on humans, because humans have memory bottlenecks, workload, anxiety, fatigue and other pressures, and even data overload. This means that we need to integrate human and machine intelligence, and hand over the computing and memory storage to the machine. There is no memory bottleneck work pressure, but the machine cannot completely leave the human being. It needs humans to perceive whether the current machine is working properly in real time. I think that if you want a machine to truly work independently from humans, you must wait until the machine can generate autonomous consciousness, can judge its own failures by itself, and have an intentional value orientation. Then it will be true wisdom.
Next, the author also mentioned: In-depth situational awareness is the perception of situational awareness, which is based on the situational awareness of the subject, coupled with the overall system trend analysis of the human-machine environment and their interrelationships. If situational awareness is a formal system, then in-depth situational awareness is a formal system with intentions added. Language and logic are a tool for formalizing intentionality. Deep situational awareness has two types of feedback mechanisms: soft and hard feedback mechanisms: both self-organization and self-adaptation, as well as other organization and other adaptation, including both local quantitative calculation and prediction, and global qualitative calculation and evaluation. It is a kind of autonomous and automatic convergence. The effect of information correction, compensation memory-expectation-selection-matching-prediction-control system.
我认为深度态势感知的难点是它存在一种自动反应,类似于跳蛙现象,即从信息输入阶段直接进入输出控制阶段,跳过了信息处理整合阶段,而且可能还存在多个跳蛙现象的混合并行。作者认为:这是由于任务主题明确、组织个体注意力的集中和长期针对性训练的条件习惯反射而引起的,因而可以无意识的协调各种自然活动的秩序,类似于形成了条件反射的状态。人类或者说其他具有智能的生物(比如狗狗)很容易通过训练产生这样的条件反射,但是机器是不能产生这样的条件反射的。就比如说当前制约机器人发展的一大瓶颈——莫拉维克悖论中就强调,对于计算机而言,实现逻辑推理等人类高级智慧只需要相对很少的计算能力,而实现感知、运动等低等级智慧却需要巨大的计算资源。也就是说最难以复刻的人类技能是那些无意识的技能,也就是俗话说的“熟能生巧”,那些已经形成类似不用过脑子的肌肉记忆一般的条件反射,是最难模拟实现的。因此如何将拥有具有类似“跳蛙”现象的深度态势感知融入到机器中,是值得我们探究的。作者提出,在深度态势感知形成跳蛙现象的过程中,类比发挥着非常重要的作用,这或许就是一大切入点了。
除了跳蛙现象,深度态势感知还具备了准确把握刺激客体的关键信息特征的能力,相较而言,普通的态势感知信息采样量大、信息融合慢、预测规划迟缓、执行力弱。深度态势感知则是基于以前经验的积累而进行的反应和行动,而不是通过常规统计概率的方法进行决策(或许有时也需要被迫对一些已经变化了的任务情境做有意识的分析决策),但很少把注意力转移到非主题或背景因素上,也就是说深度态势感知可以更好的抓住重点。深度态势感知系统通过大量的实践和训练经验,形成了一种内隐的动态触发情境认知阈值,即遇到对自己有用的关键信息特征就被激活,而不是只激活规定的信息特征。对于人类而言,学习最重要的就是忽略非关键数据、忘记不重要的信息,从而更好的抓住关键信息的特征、关联等来进行知识的获取,就好比人类有一个类似信息过滤的机制,知道什么重要什么不重要,能抓住中心思想和主题重点,能把握事物内在关联继而进行创造性分析,这或许就是价值决定的意向性吧。但是目前的机器学习是不会忽略、不懂忘记的,机器缺乏价值判断,尽管机器拥有大量的数据输入却不能整合出有用的信息,因而这也是深度态势感知的优势所在,它可以刺激客体把握关键信息,使机器更加的智能、有效。
作者表明,深度态势感知应该在情境中保持主动性,就类似于自动化的前馈控制系统。在多批量多目标多任务的情况下,想要快速有效的获取所需要的信息,就需要要建立适度规模的多模块深度态势感知技术系统,其次是要控制系统各功能模块间的整合与协调。深度态势感知系统在环境的刺激下,通过采集、过滤、改变态势分析策略,从动态的信息流中抽取不变性,在人机环境交互的作用下产生近乎知觉的操作或控制。这就要求深度态势感知中的计算是动态的非线性的,而且应该是具有自适应性的。
作者曾经说过,好的人机融合、深度态势感知思想,大都是从研究自己开始的,态势感知其实贯穿在我们生活中的方方面面。那么,研究人类的认知,研究人类的情感,研究意识的产生机理,或许就是研究深度态势感知好的开始吧。人类的认知和情感都是很复杂的,意识是如何产生的?可以把无关事物相关化的直觉是怎么来的?作为人类“初始化”的婴儿是如何认知的?是怎么进行抓取、行动、表达反抗与拒绝的?是怎么来学习、怎么来记忆、怎么来理解的?是怎么认识这个世界的?更重要的或许是人类的情感,情感影响着人类的态度、情绪继而影响到行为和其他认知,而目前的情感计算识别都很单一。人会开心会难过会爱会恨会嫉妒,那么这些情感又是怎样产生的?情感只是简单情绪的堆积吗?又该怎么判定自己的情感呢?当机器真的拥有了人类的情感,知道了通过自行更改自己的代码来进行欺骗、隐瞒,是否对于整个人工智能会产生伦理道德上的震撼考验。
当然,现在的人工智能远远不及需要人类担忧伦理的程度,想要到达那样的程度,或许需要人类从研究自己开始。知道自己如何认知,如何产生情感、意识,知道对于不同的事物怎么产生联系,然后将人类的意向性和机器的形式化融合,先产生人机融合智能,再来考虑要不要将智能完全赋予机器,使他们真的成为会思考会感动、会努力、会爱的“人类”吧!
数学在智能中的困窘是:一开始,数学就要求无矛盾性(无岐意二义性)。法国启蒙运动时期的著名哲学家、作家伏尔泰曾经说过:“不确定让人不舒服,而确定又是荒谬的。”例如:大嫂大姐大妈夫人根据不同的场合和任务可以变化性地同指一个人。