数据科学家告诉你：深度学习和机器学习不同

GoodData数据科学和机器学习高级总监Arvin Hsu 认为，尽管深度学习和机器学习两个术语可互换使用，但深度学习和机器学习在解决的业务问题和它们的结果方面有很大的不同。Arvin在数据科学和数据建模方面拥有超过15年的经验，其中包括6年的开发机器学习的数据产品，包括像迪士尼和初创公司这样的企业公司。他热衷于在大数据、机器学习和企业数据交叉点进行创新，还着迷于新技术与传统的智慧融合，用技术改变人们的工作方式。

随着人工智能进入主流技术运用当中，很多人都在讨论怎样才能最有效的利用该技术。在观察一些数据科学家的对话时，我注意到一些人经常互换使用术语机器学习（ML）和深度学习（DL）。其实两者在解决业务问题及其所需的资源方面实际上是不同的概念，混淆它们可能会导致不必要且成本高昂的结果。让我们花一点时间来具体了解一下。

我们最近总能看到一些关于的人工智能头条新闻，比如关于苹果使用面部识别，iPhone安全或制作相关资讯，模仿奥巴马总统的演讲模式的视频，这些应用使用的程序通常属于深度学习的类别。深度学习实际上已经存在了数十年，但是在过去几年中，必须在足够大的数据范围内，它才能在计算上变得可行，从而使其成为有效的选择。

深度学习被认为是机器学习整体的一个子集，这是人工智能的一种方法，使应用程序能够更准确地预测结果，而而且无需特别编程。电子邮件垃圾邮件过滤器是一个很好的例子，过滤器后面是一个算法，可以不断地“学习”关于红旗邮件的信息，指示可能的垃圾邮件或网络钓鱼邮件。因此，大多数应用程序都能够将垃圾邮件减少到收到的所有电子邮件的1-3％。大约15年前，垃圾邮件过滤器开始从基于规则的系统转移到基于机器学习的过滤器。然而一个简单的贝叶斯机器学习算法可以从一个大型的“垃圾邮件”培训里集中学习，其中单词，标题和IP地址最有可能表明电子邮件是垃圾邮件。

为了区分深度学习和机器学习之间不同的使用目的，我把在过去15到20年里，在商业上可行的简单的机器学习算法作为“经典机器学习”。这些算法包括数据科学家可以在一个小数据集上运行的一套机器学习算法，这些算法能够相对容易地创建预测和预测，集群，检测异常值等。

深度学习发挥作用，实现期望的目标，需要分析大量相互关联的复杂网络的因素。要了解两者的不同之处，请思考，一辆驶近交叉路口的汽车，使用经典的机器学习算法，即使在不同的天气条件下也可以确定交通信号是红色、黄色还是绿色。但是，正如每一个司机所知道的，在交叉路口做出决定需要理解的不仅仅是光是红色还是绿色，我们还必须考虑行人，其他汽车，我们所在的车道等，以及所有这些因素之间的关系。吸收和处理所有这些数据进行优化决策是深度学习的一个工作，这就是为什么它被用于无人驾驶当中的原因。

虽然深度学习已经进入主流商业领域，但事实上实现起来既不便宜也不简单。在人事方面，你将需要一支经过专门培训的数据科学家和工程师团队，具备深度学习技术的高级专业知识。目前这类专家人数并不是很多。在硬件方面，你将需要一台具有高端图形处理单元（GPU）的计算机，这将大大的提高成本。

幸运的是，对于大多数商业公司来说，经典机器学习将为企业提供了完美的服务 我喜欢将经典机器学习视为“80/20”解决方案，即通过20％的成本，你可以实现80％的深度学习能力。

让我们回到垃圾邮件过滤器的电子邮件的案例当中，正如我前面提到的，大多数经典机器学习功能的过滤器都能够将垃圾邮件发送率降低至1-3％。最近Google的领导层提出，这个速度很低，对Gmail用户来说是不能接受的。他们发起了一个将深度学习方法整合到Gmail过滤器中的举措，Gmail过滤器的垃圾邮件发送率为0.1％。这个结果是否值得大量投入人力，资源和资金呢？Google认为是完全值得的。

所以在你决定使用深度学习是否是明智方向之举前，花一些时间和精力来思考你最终想要实现什么。时代和技术给与我们很好的机会，经典的机器学习将让你到达你想去的地方。如果不能，深度学习的门总是开放的。

原文

Deep Learning vs. Machine Learning for Business Outcomes: A Data Scientist’s Perspective

In this special guest feature, Arvin Hsu, Senior Director of Data Science and Machine Learning for GoodData, discusses that despite the two terms being used interchangeably, deep learning and machine learning are very different in terms of the business problems they solve and the outcomes they enable. Arvin has over 15 years of experience in the field of Data Science and Data Modeling, including 6 years building Machine Learning based data products with both enterprise companies like Disney and startups. He’s passionate about the innovations being created at the intersection of Big Data, Machine Learning, and Enterprise Data. He’s also fascinated by how new technology will merge with ancient wisdoms to shift the way the world works.

As artificial intelligence (AI) works its way into mainstream business practices, various different applications are coming up in conversations about how to best leverage the technology. In observing these conversations, I notice some writers using the terms machine learning (ML) and deep learning (DL) interchangeably. The two are actually different concepts in terms of the business problems they solve and the resources they require, and confusing them could lead to unwanted — and costly — results. Let’s take a moment to set the record straight.

When we see AI making headlines — for things like Apple using facial recognition for iPhone security or the fabricated videos that mimic President Obama’s speech patterns — those applications usually fall into the category of deep learning. DL has actually been around for decades, but only in the last few years has it become computationally feasible on a large enough scale to make it an effective option.

Deep learning is considered a subset of machine learning as a whole, an approach to AI that enables applications to more accurately predict outcomes without being specifically programmed. A good example of ML at work is your email spam filter. Behind the filter is an algorithm that continuously “learns” about red flags that indicate possible spam or phishing messages. As a result, most apps are able to reduce spam to 1–3 percent of all emails received. About 15 years ago, spam filters started shifting from a rules-based system (e.g. “Move emails from Nigerian Princes into the spam folder.”) to machine learning–based filters. A simple Bayesian ML algorithm could learn from a large “spam” training set in which words, headlines, and IP addresses were most likely to indicate that an email was spam.

For differentiation purposes, I’ll refer to simple ML algorithms that have been commercially feasible for the past 15–20 years as “classic machine learning.” These comprise a set of machine learning algorithms that a data scientist can run on a small data set with relative ease to create predictions and forecasts, cluster, detect outliers, and more.

Deep learning comes into play when the desired objective requires analyzing a massive number of factors linked by a complex web of interrelationships. To understand the difference, think about a car approaching an intersection. A classic machine-learning algorithm can determine whether the traffic signal is red, yellow, or green, even under different weather conditions. But as any driver knows, making decisions at an intersection requires understanding much more than whether the light is red or green — we must also consider pedestrians, other cars, which lane we’re in, etc., and how all these factors relate to each other. Absorbing and processing all this data to make an optimized decision is a job for deep learning, which is why it’s being used for self-driving cars.

While deep learning has worked its way into the mainstream business world, it’s neither cheap nor simple to implement. On the personnel side, you would need a team of specially trained data scientists and engineers with advanced expertise in deep learning techniques. There aren’t many of these specialists around, and those who are available command top salaries. On the hardware side, you would need a host of computers with high-end graphics processing units (GPUs), which drives up the cost exponentially.

Fortunately, for most business purposes, classic machine learning serves us perfectly well. I like to think of classic ML as an “80/20” solution — it lets you achieve 80 percent of what you could do with deep learning at just 20 percent of the cost. It all depends on your objectives.

For a real-world example, let’s go back to email spam filters. As I mentioned earlier, most classic machine learning–enabled filters are able to get spam rates down to 1–3 percent. Recently the leadership at Google decided that this rate, low as it may be, was unacceptable for Gmail users. They launched an initiative to integrate deep-learning approaches into the Gmail filter, which now boasts a spam rate of 0.1 percent, with a false-positive rate of 0.05 percent. Was this outcome worth the huge investment of people, resources, and budgetary dollars it required? Google believes it was; another organization may have seen it differently.

So before you launch a task force to decide whether deep learning is a wise direction for your organization, devote some time and energy to deciding what it is you want to achieve. Chances are good that classic machine learning will get you where you want to go. And if it can’t, the deep learning door is always open.