“Is there a true AI?” This is one question that a lot of experts in the industry have tried to give answers to in different ways, but there seems to be conflicting response nonetheless. The concept of AI is increasingly being adopted in various spheres of life, which further suggests that it’s high time an answer is provided for the question.

“真的有AI吗？” 这个问题很多业内专家试图以不同的方式给出答案，但是似乎仍然存在矛盾的回答。 人工智能的概念正越来越多地应用于生活的各个领域，这进一步表明，现在是时候为这个问题提供答案了。

In this article, we’ll be exploring the future of AI and its ability to learn and evolve as proposed by experts and leading corporations in the industry.

在本文中，我们将探索AI的未来，以及该行业的专家和领先公司提出的AI的学习和发展能力。

什么是人工智能(AI)？ (What Is Artificial Intelligence (AI)?)

Artificial Intelligence is an extensive field of science that deals with Machine learning and the creation of smart machines that can act with little or no human interference. These machines have been supposedly tagged the future of technology since they will supposedly be able to think and act somewhat like humans.

人工智能是一门广泛的科学领域，涉及机器学习和创建可以在很少或没有人为干预的情况下运行的智能机器。 据说这些机器被标记为技术的未来，因为它们将能够像人类一样思考和行动。

A lot of big corporations are already diving into this field of technology, and one of the most common AI today is Apple’s Siri. Siri can do a whole lot of things, including calling your contacts, sending of SMS, show locations, play music, and a whole lot of other things.

许多大公司已经开始涉足这一技术领域，而当今最常见的AI之一是Apple的Siri。 Siri可以做很多事情，包括给您的联系人打电话，发送短信，显示位置，播放音乐以及很多其他事情。

The major question, however, is the possibility of having a true and independent AI. By true AI, I mean an AI that can completely think on its own without any form of human support. One of the foremost persons to raise this subject of this ability for an AI was the English mathematician, Alan Turing.

然而，主要的问题是拥有真正独立的AI的可能性。 真正的AI是指无需任何形式的人工支持就可以完全独立思考的AI。 英国数学家艾伦·图灵(Alan Turing)是提高人工智能这一能力的最重要人物之一。

History recognizes Turing for breaking the Nazi encryption machine, Enigma, during World War II. About a decade later, he posed the simple question that got the world thinking, “Can machines think?”

历史承认图灵在第二次世界大战期间破坏了纳粹加密机Enigma。 大约十年后，他提出了一个让世界思考的简单问题：“ 机器可以思考吗？ ”

In his 1950 paper, Turing charted a new course in the study of AI, which the world is still exploring today. While many believe that AI can be further developed to attain the full thinking dimension by man, others opine that it’s a futile dream.

图灵在其1950年的论文中绘制了AI研究的新课程，当今世界仍在探索这一课程。 尽管许多人认为可以进一步发展人工智能以实现人的全部思维能力，但其他人则认为这是徒劳的梦想。

真正的AI存在吗？ (Does True AI Exist?)

According to Yann LeCun (Professor of Computer Science at NYU and VP and Chief AI Scientist at Facebook), the technology of Generative Adversarial Networks (GANs) is one of the most “promising directions” in machine learning. LeCun is an authority in the field and certainly knows the depth of research that has been put in to the industry by different people and corporations.

根据纽约大学计算机科学教授，Facebook副总裁兼首席AI科学家Yann LeCun的说法，生成对抗网络(GANs)技术是机器学习中最有前途的方向之一。 LeCun是​​该领域的权威，并且当然知道不同的人和公司对该行业进行的深入研究。

However, I am an outlier on this topic as I don’t believe that true AI exists. Judging from his previous lectures, Professor LeCun also doesn’t believe that true AI exists, which is why I am perplexed by his comments on GANs being promising.

但是，由于我不相信真正的AI，所以在这个话题上我离奇。 从他以前的演讲来看， 勒昆教授也不相信真正的AI存在 ，这就是为什么我对他对GAN很有希望的评论感到困惑。

This idea of two machines facing off to sharpen each other skills sounds interesting indeed, but I think it ignores some of the staples of ML and Data Science. First, the tools don’t have inherent value on their own, meaning outside of human direction/guidance.

两台机器相互面对以提高彼此技能的想法听起来确实很有趣，但我认为它忽略了ML和Data Science的某些主要功能。 首先，这些工具本身并没有内在价值，这意味着超出了人类的指导/指导范围。

You cannot teach a machine to play like a master chess player by feeding it basic chess rules and facing it off against another machine that is a beginner chess player. I can see slight time savings by giving a GAN remedial work, but all the valuable ML we have now is always sourced and verified by human intelligence.

您不能通过提供基本的国际象棋规则并使它与另一台是初学者的象棋游戏机对峙的方式来教导一台象棋手一样的机器。 通过进行GAN补救工作，我可以节省一些时间，但是我们现在拥有的所有有价值的ML始终是由人类情报提供和验证的。

Even with all of the choices and all the data, “machines” still cannot predict, discern, qualify, or create much past the level of a 5-year-old child. This should not be surprising, it’s very much analogous to the way the study of thermodynamics and non-equilibrium thermodynamics are related, but the rules of the first seldom hold in the latter.

即使拥有所有选择和所有数据，“机器”仍然无法预测，辨别，鉴定或创造超过5岁孩子的水平。 这并不奇怪，它与热力学和非平衡热力学研究之间的联系非常相似，但后一种很少遵循第一个规则。

Simulations of intelligence have a long way to go, and pitting them against one another in GAN as a way of discovering something new, is at best a McNamara fallacy (quantitative fallacy): “making a decision based solely on quantitative observations and ignoring all others.” Having all the data is not the only factor in making a ‘good’ decision.

智力模拟还有很长的路要走，在GAN中将它们相互竞争，以此作为发现新事物的方式，充其量是McNamara谬论(定量谬论) ：“仅基于定量观察而忽略所有其他因素来做出决定。” 拥有所有数据并不是做出“良好”决策的唯一因素。

使用机器学习能否打破标准AES块密码加密？ (Is it Possible to Break Standard AES Block Cipher Encryption Using Machine Learning?)

My colleague once sought my opinion on the feasibility of training a machine learning system that could break standard AES block cipher encryption. The idea was to train an ML system with a set of AES ciphertexts labelled with the corresponding plaintexts (assume for simplicity that the encryption key is fixed). The hope was that it would generate an ML model that is able to invert AES, that is, decrypt ciphertexts into the corresponding plaintexts.

我的同事曾经就培训机器学习系统的可行性征询我的意见，该系统可能会破坏标准的AES块密码加密。 这个想法是用一组带有相应明文标记的AES密文来训练ML系统(为简单起见，假定加密密钥是固定的)。 希望它会生成一个能够反转AES的ML模型，即将密文解密为相应的明文。

If this works, the model will successfully guess future unobserved plaintexts by processing ciphertexts through the learned function, thus breaking all modern secure communication and storage solutions. Brilliant idea, it seemed.

如果可行，该模型将通过学习的功能来处理密文，从而成功猜测出未来未观察到的明文，从而打破所有现代的安全通信和存储解决方案。 似乎很棒的主意。

But, sadly, this idea won’t work because there is no pattern for the machine to learn as assumed by the statement: “let’s train an ML system with a set of AES ciphertexts labelled with the corresponding plaintexts.” The colleague is falsely assuming that ML can “find” key by looking at pairs of ciphertexts and plain texts.

但是，可悲的是，这个想法行不通，因为机器没有按照以下语句所假设的学习模式：“让我们训练带有一组标有相应明文的AES密文的ML系统。” 同事错误地认为ML可以通过查看成对的密文和纯文本来“查找”密钥。

However, there is no magic algorithm to undo a one-way trapdoor function. Of course, there is always brute force, trying all the combinations until you find the right one. But there is no way to “learn” from such a small dataset.

但是，没有魔术算法可以撤销单向陷门功能。 当然，总会有蛮力，尝试所有组合，直到找到合适的组合为止。 但是没有办法从如此小的数据集中“学习”。

Machine learning is about analyzing large data sets, training, and finding the rules that exist to tie things together with math. In this case, however, the math used to relate the ciphertext and the plain text is designed to be irreversible, and the variables can only be discovered with brute force.

机器学习是关于分析大型数据集，训练并找到将数学与事物联系在一起的规则。 但是，在这种情况下，用于将密文和明文联系起来的数学是不可逆的，并且只能用蛮力来发现变量。

如何重新训练Chatbot机器学习系统以更好地与用户互动 (How To Retrain A Chatbot Machine Learning System To Interact Better With Users)

Typical examples of chatbots are those used in banking sites for automated customer service. Usually, such bots are trained by looking at user chat records and discovering the highest mode among customer questions.

聊天机器人的典型示例是银行站点中用于自动客户服务的聊天机器人。 通常，通过查看用户聊天记录并发现客户问题中的最高模式来训练此类漫游器。

When we find the first-order mode for an array of questions, we can use the question as a first-order intent and use natural language processing to record all the possible ways a user can ask the same question.

当我们找到一系列问题的一阶模式时，我们可以将问题用作一阶意图，并使用自然语言处理来记录用户可以问相同问题的所有可能方式。

We can look and see if there is a high correlation between the satisfaction of the user and the agent providing anyone answer or solution. If there is a high correlation between the users, this would be the algorithmic way of determining the most common question and the most satisfying answer.

我们可以看看用户的满意度与提供答案或解决方案的代理商之间是否存在高度相关性。 如果用户之间的相关性很高，这将是确定最常见问题和最令人满意答案的算法方法。

Retraining such a chatbot would recommend a Supervised Learning Model (Discrete Variable Prediction) with a Decision Tree Classifier that is powered by a rich set of curated user intents.

对这样的聊天机器人进行再培训将建议使用带有决策树分类器的监督学习模型(离散变量预测)，该决策树由一组丰富的策划用户意图提供支持。

Retraining the bot based on unedited chat transcripts is fine (and is probably best) as long as the models are preprocessed to dealing with missing data, handle data imputation, and handle categorical data.

只要对模型进行预处理以处理丢失的数据，处理数据估算和分类数据，就可以基于未编辑的聊天记录对机器人进行再培训是很好的(并且可能是最好的)。

The model should also be trained to encode class labels for classification problems, induce features to transform the data, and deal with dimensionality reduction, such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA).

还应该训练模型以对分类问题的类标签进行编码，引入特征以转换数据以及处理降维，例如主成分分析(PCA)和线性判别分析(LDA)。

结论 (Conclusion)

Artificial intelligence can certainly do several things, but it has limitations. More so, its evolution still subjects it to the dependence on humans. True AI does not exist, at least for now. More so, the industry needs more and more research if further advancements can be achieved.

人工智能当然可以做几件事，但是它有局限性。 更重要的是，它的进化仍然使它依赖于人类。 真正的AI至少在现在还不存在。 因此，如果要实现进一步的发展，行业需要越来越多的研究。