走向乌托邦的未来 (Towards a utopian future)

The past decade has seen an intense hype about the promising future of artificial intelligence (AI). A future where automation would create a dearth of human jobs by replacing them with AI based systems. The objective of making this utopian dream a reality is encouraging industry as well as academia, to invest towards the research and development of AI based systems. Terms like Machine Learning (ML), Deep Learning(DL) and AI can be frequently encountered in application and research based publications. Improvements in computing power and storage, and the undergoing developments of GPU based libraries have facilitated the use of these methods. However, these terms muddle a lot of people and are sometimes used interchangeably. Therefore, before delving into the amazing prospects of these automated systems, one needs to understand what these terms mean and how are they related/correlated to one another.

在过去的十年中，人们对人工智能(AI)的光明前景进行了大肆宣传。 未来，自动化将通过基于AI的系统代替人工来创造大量人工工作。 实现这个乌托邦梦想的目标是鼓励行业和学术界投资于基于AI的系统的研发。 诸如机器学习(ML)，深度学习(DL)和AI之类的术语在基于应用程序和研究的出版物中经常遇到。 计算能力和存储的改进以及基于GPU的库的不断发展促进了这些方法的使用。 但是，这些术语使很多人感到困惑，有时可以互换使用。 因此，在深入研究这些自动化系统的惊人前景之前，需要了解这些术语的含义以及它们之间的关系/相关性。

迈向自动化智能：过去，现在和未来 (Towards Automated Intelligence: Past, present and future)

The term ‘Artificial Intelligence’ came into existence in the 1950s when researchers in the domain of computer science starting pondering over the possibility of “making computers think”. In other words, would it be possible to automate intellectual tasks performed by humans? Most of the research in AI is centered around five different components: learning, reasoning, problem-solving, perception, and language-understanding. To achieve these objectives researchers have proposed a number of tools, some of which we are going to discuss in this article:

“人工智能”一词在1950年代问世，当时计算机科学领域的研究人员开始思考“让计算机思考”的可能性。 换句话说，是否有可能使人类执行的智力任务自动化？ 人工智能的大多数研究都围绕五个不同的方面：学习，推理，解决问题，感知和语言理解。 为了实现这些目标，研究人员提出了许多工具，本文将讨论其中一些：

Symbolic AI

象征性AI

Symbolic AI (Haugeland, John (1985)) was the earliest forms is AI (1950–1980) that researchers believed could help achieve human-level intelligence by explicitly hardcoding the rules to manipulate data (IBM’s Deep Blue vs chess champion Kasparov, 1997). But could machines go beyond performing tasks which we already know how to perform? In other words, can a machine learn different patterns and behaviours by themselves without any set of hardcoded rules? This question paved way for a new

象征性AI( Haugeland，John (1985))最早的形式是AI(1950–1980)，研究人员认为它可以通过对规则进行硬编码来操纵数据来帮助实现人类智能(IBM Deep Blue vs国际象棋冠军Kasparov，1997)。 。 但是，机器能否超越执行我们已经知道如何执行的任务呢？ 换句话说，一台机器无需任何硬编码规则就能自己学习不同的模式和行为吗？ 这个问题为新方法铺平了道路

Machine Learning

机器学习

Machine learning focusses more on learning rules rather than simply executing pre-defined rules. Rules are learnt by exposing the model to relevant examples. For example, in order to understand social sentiment of a product, businesses try to classify online opinions given by their customers into positive or negative. For this task, firstly a set of customer online reviews are manually tagged as positive (denoted by 1) or negative (denoted by 0). These tagged reviews are then input to the model. The model then tries to learn mathematical/statistical rules that enable it to predict the nature of a review (positive or negative) based on its text.

机器学习更多地关注学习规则，而不是简单地执行预定义的规则。 通过将模型暴露给相关示例来学习规则。 例如，为了理解产品的社会情感，企业试图将其客户给出的在线意见分为正面还是负面。 对于此任务，首先将一组客户在线评论手动标记为正面(由1表示)或负面(由0表示)。 这些标记的评论然后输入到模型中。 然后，该模型尝试学习数学/统计规则，以使其能够根据其文本预测评论的性质(正面或负面)。

In order to learn these rules, a ML model needs the following three informations:

为了学习这些规则，ML模型需要以下三个信息：

1. Data (Input): For a task of sentiment analysis, data includes different textual reviews submitted by the users数据(输入)：对于情感分析任务，数据包括用户提交的不同文本评论
2. Answers(Output): Each of the reviews in step1, tagged with their sentiments (positive=1, negative=0)答案(输出)：步骤1中的每条评论，都标有他们的观点(正面= 1，负面= 0)
3. Measure of performance (Loss function): Function which quantifies model performance (loss score) by comparing ‘true answers’ with ‘model predicted answers’. The parameters (weights, bias) are then adjusted in the direction of minimization of loss score. This is the step where the model “learns” the rules based on examples it is exposed to and updates its parameters accordingly. For sentiment analysis task, the loss function is log-sigmoid function.性能度量(损失函数)：通过将“真实答案”与“模型预测答案”进行比较来量化模型性能(损失得分)的功能。 然后在损失分数最小化的方向上调整参数(权重，偏差)。 这是模型根据其所暴露的示例“学习”规则并相应更新其参数的步骤。 对于情感分析任务，损失函数是对数乙状结肠函数。

So what does a machine learning model do? It transforms the given ‘data’(input) into meaningful ‘answers’ (output) by learning the rules (optimizing the objective function) through the relevant examples it is exposed to. The term “transforming the data” means representing the data in a form where the task at hand becomes easy to achieve. For example, in figure 3 below, the task is to linearly separate red points from blue. In 2D space (left figure), it is clearly not possible to do so. However, if we transform this data using the rules shown in figure 3(b), the data becomes linearly separable (right figure). The transformation shown in figure 3(b) has facilitated in finding the separating hyperplane.

那么机器学习模型有什么作用？ 通过学习规则(优化目标函数)并通过暴露的相关示例，它将给定的“数据”(输入)转换为有意义的“答案”(输出)。 术语“转换数据”是指以容易实现手头的任务的形式表示数据。 例如，在下面的图3中，任务是线性地将红色点与蓝色分开。 在2D空间(左图)中，显然不可能这样做。 但是，如果我们使用图3(b)中所示的规则转换此数据，则数据将变为线性可分离的(右图)。 图3(b)所示的变换有助于找到分离的超平面。

In conclusion, a ML model ‘searches’ for pertinent representations of the data within a pre-defined space of possibilities (hypothesis space) while being directed by the objective function.

总之，一个ML模型在目标功能的指导下，在可能性的预定义空间(假设空间)内“ 搜索 ”数据的相关表示。

Deep learning

深度学习

Deep learning is a specific form of machine learning where layers of meaningful representations are stacked one after the other. As information passes through each layer, the representation of the input data become more helpful in predicting the output. Each layer can be considered as a filter that purifies the data coming to it so that output becomes increasing clear. The following figure clarifies this process.

深度学习是机器学习的一种特殊形式，其中有意义的表示层一层又一层地堆积在一起。 随着信息穿过每一层，输入数据的表示形式在预测输出中变得更加有用。 每一层都可以看作是过滤器，用于净化到达其的数据，从而使输出变得越来越清晰。 下图阐明了此过程。

The term “deep” in DL is a reference to the presence of successive layers. The depth of a model is equal to the number of successive layers it constitutes. Some ML models use one or two layered representations (shallow networks) whereas modern deep learning models use up to tens or hundreds of layers.

DL中的“深层”一词是指连续层的存在。 模型的深度等于其构成的连续层数。 一些ML模型使用一层或两层表示(浅层网络)，而现代深度学习模型则使用多达数十或数百层。

So how does an iteration of a DL model look like? Flowchart shown in figure 5 depicts the different steps involved in each iteration. Each layer is characterized by weights. Weights can be thought of as projections of input data into higher dimensions. For example, in figure 4, the input image of digit 4 was projected into a 4 dimensional space by layer1. So after data flows through layer 1 we end up with 4 different representations of the image.Once data flows through all of the 4 layers and predicts the output, a loss score is evaluated quantifying the performance of the model. The optimizer then updates the weights in the direction of decreasing loss score. This process is iterated until the loss score reaches a stage where it doesn’t change significantly over iterations.

那么DL模型的迭代看起来如何？ 图5所示的流程图描述了每次迭代所涉及的不同步骤。 每层均以重量为特征。 权重可以看作是输入数据到更高维度的投影。 例如，在图4中，数字1的输入图像被layer1投影到4维空间中。 因此，在数据流过第1层之后，我们最终得到了图像的4种不同表示形式。一旦数据流过所有4层并预测了输出，就会对损失评分进行评估，从而量化模型的性能。 然后，优化器会按照降低损失评分的方向更新权重。 反复进行此过程，直到损失分数达到在迭代中不会发生明显变化的阶段为止。

Having understood the mechanism around DL models, the next question that arises is that what makes DL better than ML? DL models make problem solving much easier by completely automating “Feature engineering”. ML models needs human intervention in engineering good layers of representation of the input data (e.g., kernel functions in SVM, transformation shown in figure 3(b)). Complex problems require finer representation of input data which cannot be attained manually. DL models, on the other hand, completely automate this step by learning successive layers of data representations (or weights) in one pass. This simplifies the workflow by creating a single end-to-end model where all the features are jointly learnt. In summary, the efficiency of DL models could be attributed to two factors:

了解了DL模型的机制后，出现的下一个问题是，什么使DL优于ML？ DL模型通过完全自动化“特征工程”使解决问题变得更加容易。 ML模型需要人为干预来设计输入数据表示的良好层(例如，SVM中的内核功能，图3(b)所示的转换)。 复杂的问题需要无法手动获得的输入数据的更好表示。 另一方面，DL模型通过一次通过学习连续的数据表示层(或权重)来完全自动化此步骤。 通过创建可共同学习所有功能的单个端到端模型，这简化了工作流程。 总之，DL模型的效率可以归因于两个因素：

1. incremental layer-by-layer mechanism that enables incorporating complex representations of input data and,渐进的逐层机制，可以合并输入数据的复杂表示，并且
2. joint learning of all intermediate layers so that the dependent layers are updated simultaneously.联合学习所有中间层，以便同时更新相关层。

结论 (Conclusion)

In this post we covered a brief history of AI and how it evolved over the years through symbolic AI, ML and DL. We also tried to understand how AI, ML and DL are related/correlated with one other (figure 6). These terms often muddle a lot of people and are often used interchangeably.

在这篇文章中，我们简要介绍了AI的历史，以及多年来AI通过符号AI，ML和DL演变的过程。 我们还试图了解AI，ML和DL如何相互关联/关联(图6)。 这些术语经常使很多人感到困惑，并且经常互换使用。

There is a lot of hype around DL models due to their ability to better deal with complex problems such as image classification, text processing etc. Additionally these models also remove the necessity to manually engineer features, amenable to the objective at hand. Therefore, in the next post we shall be delving into the details of DL models and try to understand its various components along with examples.

由于DL模型能够更好地处理诸如图像分类，文本处理等复杂问题，因此它们有很多炒作。此外，这些模型还消除了手动设计功能的必要性，以适应当前的目标。 因此，在下一篇文章中，我们将深入研究DL模型的细节，并尝试与示例一起理解其各个组成部分。