中國情報戰概覽 | 機器思維：軍事情報戰取勝的關鍵
現代英語：
Editor’s Note
In the 1950s, scientist Alan Turing first proposed the concept of “machine thinking.” With the advent of the intelligent era, the idea that machines can also possess “thinking” is gradually becoming a reality. In intelligent warfare, driven by machine thinking, some unmanned equipment and decision-making aids become “robot allies” and “intelligent advisors” fighting alongside humans. It is foreseeable that the relationship between humans and weapons will gradually shift from that of humans and tools to that of humans and intelligent partners with “limited subjective initiative.” A deep understanding and skillful application of machine thinking as a key will help people recognize the characteristics of intelligent warfare and seize the initiative in it.
In recent years, next-generation artificial intelligence technologies, represented by deep learning, have made groundbreaking progress, surpassing humans in many fields such as Go, speech recognition, and translation. More and more people are beginning to realize that the human brain is merely a highly advanced general-purpose intelligent agent; human intelligence is not the only form of intelligence in the world, nor is it the ultimate form of intelligence. Human society is entering an era of intelligent coexistence between humans and machines. All preparations for intelligent warfare, including exploring the mechanisms for winning intelligent warfare, developing intelligent weapons and equipment, developing intelligent combat forces, and innovating intelligent combat methods, should be based on a thorough understanding of how intelligent machines “think.”
Machine thinking is developing rapidly
From mechanical technology to information technology and then to artificial intelligence, technological development has progressed from simulating human limb functions, sensory functions, neural functions, and finally cognitive functions, gradually replacing, expanding, and amplifying various human abilities, progressing from simple to complex and from low to high levels. As a replacement for the human brain, the most complex organ in the human body, artificial intelligence must possess “thinking” abilities similar to those of the human brain in solving complex problems; we can call this “machine thinking.”
The new generation of artificial intelligence systems based on deep learning can be viewed as a “gray box” compared to the previous generation, with its “thinking” process and results exhibiting significant uncertainty and inexplicability. While people hope it can be explained, from another perspective, it is precisely this uncertainty and inexplicability that generates creativity and constitutes the true “source of wisdom.” Higher forms of human thought, besides logical reasoning, such as intuition, imagination, inspiration, and sudden insight, all possess a high degree of uncertainty and can only be understood intuitively, not explained in words. Just as the art of command in the military, where “the subtlety of application lies in the mind,” is difficult to explain.
Therefore, the uncertainty and inexplicability exhibited by machine thinking may precisely be the advanced and unique aspect of this breakthrough in artificial intelligence. No matter how fast a supercomputer or quantum computer is, or how powerful its computational intelligence is, because its computational principles are transparent and interpretable, its computational rules are pre-designed and deterministic, and its computational process is reversible and repeatable, people do not consider it creative or a challenge to human thinking abilities.
This breakthrough in artificial intelligence has significantly improved the “intelligence” of intelligent machines, with machine thinking demonstrating unique advantages in many fields that differ from and surpass human thinking. For example, after AlphaGo defeated the human world Go champion, some believed it was closer to the god of Go, creating a completely new school of Go like the “cosmic style,” and some Go players even began to learn from AlphaGo’s playing style. Furthermore, generative AI like ChatGPT, which has become incredibly popular in the last two years, already possesses a certain degree of creativity and human-like “subjective initiative,” enabling it to replace humans in many tasks.
Machine thinking is different from human thinking.
Currently, although artificial intelligence has made groundbreaking progress, it is still in the development stage of perceptual intelligence, weak AI, and specialized AI. Compared with human thinking, machine thinking still has obvious shortcomings. Experts have summarized its deficiencies into four points: First, it “has intelligence but lacks wisdom,” lacking intuition, inspiration, and other implicit human thinking abilities. Einstein once said that raising a question is often more important than solving a problem. ChatGPT is far better than the average person at answering questions, but it cannot raise a truly valuable scientific question. Second, it “has IQ but lacks EQ.” Intelligent machines themselves do not have, and find it difficult to, simulate human emotions such as anger, sadness, and joy, and therefore cannot truly understand these human emotions. Third, they are “good at calculation but not at scheming.” Although intelligent machines “think” very quickly, they are not good at taking roundabout ways or retreating to advance. They cannot pretend, deceive, or use tricks like humans. Fourth, they are “good at specialization but not generalization.” Intelligent machines have poor “learning by analogy,” that is, their ability to transfer learning is very poor. Although specialized artificial intelligence software can surpass human champions in Go, the “intelligence” of the most advanced general-purpose brain-like chips can only approach the level of a mouse brain.
Although machine thinking was created and designed by humans, it differs significantly from human thinking. There’s a Moraviek paradox in the field of artificial intelligence: for AI, achieving complex logical reasoning and other high-level human cognitive abilities requires minimal computation, while achieving unconscious skills like perception and movement, and simpler cognitive abilities like intuition, demands enormous computational power. AI can outperform humans in playing Go and solving equations, but tasks easy for humans, like driving a car or folding clothes, are very difficult for AI. Experts have outlined what AI currently cannot do, including: cross-domain reasoning, abstract thinking, self-awareness, aesthetics, and emotion. These are not difficult for humans, but are very challenging for AI.
Based on the differences between machine thinking and human thinking, in intelligent warfare, on the one hand, traditional strategies that work for humans, such as feints and diversions, are likely to be easily detected by machine thinking; the massive amounts of battlefield data, far exceeding the analytical processing capabilities of the human brain, will become the “thinking” material for machine thinking, allowing it to find clues about enemy actions and important targets. On the other hand, machine thinking also has some major flaws that seem utterly “idiotic” to humans. Foreign research teams have discovered that by changing just a few key pixels in a picture of a cat, an intelligent machine can identify the cat as a dog, while the human eye will not misidentify it due to this change. This illustrates a significant difference between deceiving humans and deceiving intelligent machines. The “calculations” used to deceive humans may be useless against the “calculations” of intelligent machines. Conversely, deception methods targeting machine thinking are very easy to use to fool intelligent machines, but may not be able to fool humans. With the deep application of artificial intelligence in intelligence analysis, further research is needed on how strategic deception is organized, how battlefield feints are implemented, how to deceive both human and computer brains, how to attack the weaknesses of adversary intelligent machines, and how to prevent one’s own intelligent machines from being deceived.
All of the above facts show that the complexity problems faced by humans and machines may be exactly opposite. Humans and machines each have their own advantages and disadvantages and are highly complementary. Through human-machine collaboration, humans can be responsible for judging whether they are “doing the right thing” while machines “do things correctly”.
Create machine thinking based on machine characteristics
The carrier of machine thinking is silicon-based chips, but it is not endogenous; rather, it is created by humans using innovative thinking. The level of human creators’ thinking determines the level of machine thinking. A key point to grasp in creating machine thinking is that it cannot be simply copied from human thinking methods based on carbon-based intelligence. Instead, it should be created according to the characteristics of silicon-based machines in terms of perception, judgment, decision-making, and action.
For example, how do cars pass through intersections? For manned vehicles, a complete set of mature rules has been established to avoid congestion and traffic accidents. But how can autonomous vehicles pass through without collisions? There are at least three solutions. First, the autonomous vehicle stops at the intersection and uses its onboard camera to mimic human eyes, automatically recognizing and judging traffic light changes, and only proceeding when the light turns green. Second, a signal generator is installed on the traffic light pole; when the green light is on, it directly emits a signal indicating passage, which the autonomous vehicle receives before proceeding. Third, traffic lights are eliminated; the autonomous vehicle uses sensors such as lidar, cameras, and millimeter-wave radar to detect passing vehicles at the intersection, employing collision avoidance algorithms and vehicle-to-vehicle cooperation to pass quickly and without interruption. The first approach is to design the driving method of autonomous vehicles according to human driving thinking and behavioral habits. The second approach is an improvement on the first approach. The third approach completely subverts the traditional mode of human vehicles relying on traffic lights and passing through intersections in a “stop-wait-go” manner, which greatly improves traffic efficiency and is equivalent to giving autonomous vehicles a machine thinking that truly suits their own characteristics.
Massively creating machine thinking to seize intelligent advantage
Machine thinking is essentially algorithmic thinking, digital thinking, and precision thinking. In intelligent warfare, in order to make one’s own intelligent machines “smarter” than the opponent’s and to seek to overwhelm the opponent’s intelligent advantage, we should create a large number of different types of high-level machine thinking and greatly improve the ability of intelligent machines to adapt to changing battlefield environments and solve complex combat problems.
For example, creating machine thinking that enables unmanned swarms to collectively understand the battlefield situation. A fundamental prerequisite for efficient collaborative operations between combat units is a shared understanding of the battlefield situation. For humans, the most intuitive and effective method is based on a unified battlefield situation map. However, this approach is unsuitable for collaborative operations between unmanned platforms within a swarm. This is because using visual diagrams as a medium for machine-to-machine communication is inefficient, and it is difficult for unmanned platforms to directly extract useful information from battlefield situation maps. Therefore, a dedicated battlefield situation sharing mechanism adapted to machine-to-machine communication is needed. For instance, leveraging the fact that intelligent machines are more efficient at “counting” than “viewing images,” the unmanned swarm can use software to create a virtual “bulletin board,” i.e., a shared data file. In collaborative operations, each drone platform promptly publishes its own location and status, as well as the nature, location, and environmental information of targets detected by its sensors, to the “bulletin board.” All drone platforms in the cluster can quickly read this shared data file to obtain near-real-time information on the enemy, ourselves, and the environment, thereby achieving a shared understanding of the battlefield situation.
Another example is the development of machine thinking for integrated offensive and defensive warfare using unmanned platforms. The basic principle of warfare, “eliminate the enemy, preserve yourself,” is easily understood by human soldiers, but enabling unmanned platforms to correctly balance avoiding enemy threats and engaging enemy targets requires a different approach. Utilizing artificial potential field algorithms might be one solution. Unmanned platforms could construct a repulsive potential field around targets that pose a threat, with stronger repulsion due to greater threat; and a gravitational potential field around targets intended for attack, with stronger attraction due to higher target value. Under the combined influence of these gravitational and repulsive potential fields, the unmanned system automatically generates the optimal attack path, thus maximizing the achievement of both eliminating the enemy and preserving itself.
現代國語：
編按
在1950年代，科學家艾倫·圖靈首次提出了「機器思維」的概念。隨著智慧時代的到來，機器也能擁有「思維」的概念逐漸成為現實。在由機器思維驅動的智慧戰爭中，一些無人裝備和決策輔助工具正成為與人類並肩作戰的「機器人盟友」和「智慧顧問」。可以預見，人與武器的關係將逐漸從人與工具的關係轉變為人與擁有「有限主觀主動性」的智慧夥伴的關係。深入理解並巧妙運用機器思維是關鍵，有助於人們認識智慧戰爭的特點，並在其中掌握主動權。
近年來，以深度學習為代表的新一代人工智慧技術取得了突破性進展，在圍棋、語音辨識、翻譯等諸多領域超越了人類。越來越多的人開始意識到，人腦只不過是一個高度發展的通用智能體；人類智能並非世間唯一的智能形式，也並非智能的終極形式。人類社會正步入人機智慧共存的時代。一切智慧戰爭的準備工作，包括探索智慧戰爭的致勝機制、研發智慧武器裝備、發展智慧作戰力量以及創新智慧作戰方法，都應建立在對智慧機器「思考」方式的透徹理解之上。
機器思維正在快速發展
從機械技術到資訊技術，再到人工智慧，技術發展經歷了從模擬人體肢體功能、感覺功能、神經功能，最終到認知功能的演進，逐步取代、擴展和增強了人類的各種能力，由簡到繁、由低到高不斷演進。作為人體最複雜器官——人腦的替代品，人工智慧必須具備與人腦類似的「思考」能力，能夠解決複雜問題；我們可以稱之為「機器思維」。
與上一代人工智慧系統相比，基於深度學習的新一代人工智慧系統可以被視為一個“灰箱”，其“思考”過程和結果都展現出顯著的不確定性和不可解釋性。人們雖然希望能夠解釋這些過程，但從另一個角度來看，正是這種不確定性和不可解釋性激發了創造力，構成了真正的「智慧之源」。除了邏輯推理之外，更高層次的人類思維，例如直覺、想像、靈感和頓悟，都具有高度的不確定性，只能透過直覺來理解，而無法用語言來解釋。正如軍事指揮的藝術一樣，“運用之妙在於心智”，難以言表。
因此，機器思維所展現出的不確定性和不可解釋性，或許正是人工智慧這項突破的先進之處與獨特之處。無論超級電腦或量子電腦的速度有多快，計算智慧有多強大，由於其計算原理透明且可解釋，計算規則預先設計且具有確定性，計算過程可逆且可重複，人們並不認為它具有創造性，也不認為它對人類思維能力構成挑戰。
人工智慧的這一突破顯著提升了智慧機器的“智慧”，機器思維在許多領域展現出與人類思維截然不同甚至超越人類思維的獨特優勢。例如，AlphaGo擊敗人類圍棋世界冠軍後，有些人認為它更接近圍棋之神，開創了「宇宙流」等全新圍棋流派，甚至有圍棋選手開始學習AlphaGo的棋風。此外，像ChatGPT這樣在過去兩年迅速走紅的生成式人工智慧，已經具備一定程度的創造力和類似人類的“主觀主動性”，使其能夠在許多任務中取代人類。
機器思維與人類思維截然不同。
目前，人工智慧雖然取得了突破性進展，但仍處於感知智慧、弱人工智慧和專業人工智慧的發展階段。與人類思維相比，機器思維仍有明顯的缺點。專家將其缺陷歸納為四點：首先，它“有智能但缺乏智慧”，缺乏直覺、靈感等人類固有的思維能力。愛因斯坦曾說過，提出問題往往比解決問題重要。 ChatGPT在回答問題上遠勝於一般人，但它無法提出真正有價值的科學問題。其次，它「有智商，但缺乏情商」。智慧機器本身並不具備，也很難模擬人類的情感，例如憤怒、悲傷和喜悅，因此無法真正理解這些人類情感。第三，它們「擅長計算，但不擅長規劃」。雖然智慧機器「思考」速度很快，但它們不擅長迂迴策略或退守後再前進。它們無法像人類那樣偽裝、欺騙或使用詭計。第四，它們「擅長專業化，但不擅長泛化」。智慧機器的「類比學習」能力很差，也就是說，它們的學習遷移能力非常弱。雖然專業的AI軟體可以在圍棋領域超越人類冠軍，但最先進的通用類腦晶片的「智慧」水平也只能接近小鼠大腦的水平。
儘管機器思維是由人類創造和設計的，但它與人類思維有顯著差異。人工智慧領域存在著一個莫拉維克悖論：對於人工智慧而言，實現複雜的邏輯推理和其他高級人類認知能力所需的計算量極少，而實現諸如感知和運動等無意識技能以及諸如直覺等更簡單的認知能力卻需要巨大的計算能力。人工智慧在圍棋和解方程式方面可以超越人類，但對人類來說輕而易舉的任務，例如開車或疊衣服，對人工智慧來說卻非常困難。專家們已經列出了人工智慧目前無法完成的任務，包括：跨領域推理、抽象思考、自我意識、美學和情感。這些對人類來說並不難，但對人工智慧來說卻極具挑戰性。
基於機器思維和人類思維的差異，在智慧戰爭中，一方面，對人類有效的傳統策略，例如佯攻和佯攻，很可能被機器思維輕易識破；另一方面，海量的戰場數據遠遠超過人腦的分析處理能力，將成為機器思維的「思考」素材，使其能夠從中發現敵方行動和重要目標的線索。另一方面，機器思維也存在著一些在人類看來極為「愚蠢」的重大缺陷。國外研究團隊發現，只要改變貓咪的圖片中幾個關鍵像素，智慧機器就能將貓辨識為狗，人眼卻不會因此而誤判。這說明欺騙人類和欺騙智慧機器之間存在顯著差異。用來欺騙人類的「計算」可能對智慧機器的「計算」毫無作用。反之，針對機器思維的欺騙方法很容易就能欺騙智慧機器，但卻可能無法欺騙人類。隨著人工智慧在情報分析領域的深度應用，我們需要進一步研究戰略欺騙的組織方式、戰場佯攻的實施方法、如何同時欺騙人類和電腦的大腦、如何攻擊敵方智慧機器的弱點以及如何防止己方智慧機器被欺騙。
以上種種事實表明，人類和機器面臨的複雜性問題可能截然相反。人類和機器各有優劣，高度互補。透過人機協作，人類負責判斷自己“是否在做正確的事”，而機器則負責“正確地做事”。
基於機器特性創造機器思維
機器思維的載體是矽晶片，但它並非內生的，而是由人類運用創新思維創造出來的。人類創造者的思維層次決定了機器思維的層次。創造機器思維的關鍵在於，它不能簡單地複製基於碳基智能的人類思維方式，而應該根據矽基機器在感知、判斷、決策和行動等方面的特性來創造。
例如，汽車如何通過十字路口？對於有人駕駛的車輛，已經建立了一套完整的成熟規則來避免擁擠和交通事故。但是，自動駕駛車輛如何才能無碰撞地通過十字路口呢？至少有三種解決方案。首先，自動駕駛車輛在十字路口停車，利用車載攝影機模擬人眼，自動辨識並判斷交通號誌的變化，僅在綠燈亮起時才通行。其次，在交通號誌桿上安裝號誌產生器；當綠燈亮起時，它直接發出通行訊號，自動駕駛車輛接收到該號誌後再通行。第三，取消交通號誌；自動駕駛車輛使用光達、攝影機和毫米波接收器等感測器進行通訊。ADA 系統用於偵測十字路口的過往車輛，利用防碰撞演算法和車對車協作實現快速無間斷通行。第一種方法是根據人類駕駛的思考和行為習慣來設計自動駕駛車輛的駕駛方式。第二種方法是對第一種方法的改進。第三種方法徹底顛覆了人類車輛依賴交通號誌、以「停-停-走」方式通過十字路口的傳統模式，大大提高了交通效率，相當於賦予自動駕駛車輛真正符合自身特性的機器思維。
大規模建構機器思維，奪取智慧優勢
機器思維本質上是演算法思維、數位思維和精確思維。在智慧戰爭中，為了使己方智慧機器比敵方更“聰明”，並力求壓倒敵方的智慧優勢，我們應該構建大量不同類型的高級機器思維，並大幅提升智慧機器適應不斷變化的戰場環境和解決複雜作戰問題的能力。
例如，創造一種機器思維，使無人集群能夠集體理解戰場態勢。作戰單位間高效率協同作戰的基本前提是對戰場態勢的共同理解。對人類而言，最直觀有效的方法是基於統一的戰場態勢圖。然而，這種方法並不適用於集群內無人平台之間的協同作戰。這是因為使用視覺化圖表作為機器間通訊的媒介效率低下，無人平台難以直接從戰場態勢圖中提取有用資訊。因此，需要一種專門針對機器間通訊的戰場態勢共享機制。例如，利用智慧機器更擅長“計數”而非“查看圖像”的特性，無人集群可以使用軟體創建一個虛擬的“公告板”，即共享資料檔案。在協同作戰中，每個無人機平台都會及時將自身位置和狀態，以及其感測器探測到的目標的性質、位置和環境資訊發佈到「公告板」上。集群中的所有無人機平台都能快速讀取共享數據文件，獲取近乎實時的敵我信息以及周圍環境信息，從而實現對戰場態勢的共同理解。
另一個例子是利用無人平台發展機器思維，以進行攻防一體化作戰。戰爭的基本原則「消滅敵人，保全自身」對人類士兵來說很容易理解，但要使無人平台能夠正確地平衡規避敵方威脅和攻擊敵方目標，則需要不同的方法。利用人工勢場演算法或許是一種解決方案。無人平台可以在構成威脅的目標周圍建構排斥位勢場，威脅越大，排斥力越強；在攻擊目標周圍建構重力位勢場，目標價值越高，引力越強。在這些重力位能場和排斥位勢場的共同作用下，無人系統能夠自動生成最佳攻擊路徑，從而最大限度地實現消滅敵人和保全自身的目標。
來源：中國軍網-解放軍報 作者：袁 藝 責任編輯：尚曉敏 發布：2024-02-27 06:xx:xx
袁 藝