How does artificial intelligence learn? - Briana Brownell

翻译人员: Yuwei Wu 校对人员: Helen Chang

Today, artificial intelligence helps doctors diagnose patients,

如今的人工智能 可以帮助医生诊断病人，

pilots fly commercial aircraft, and city planners predict traffic.

飞行员驾驶商业飞机， 城市规划员预测交通。

But no matter what these AIs are doing, the computer scientists who designed them

但不论人工智能如何运作

设计它们计算机科学家 却可能不知道其具体运作方法。

likely don’t know exactly how they’re doing it.

This is because artificial intelligence is often self-taught,

这是因为人工智能常有自学能力，

working off a simple set of instructions

它们会根据设定好的程序指令

to create a unique array of rules and strategies.

创造一套独特的规则和策略。

So how exactly does a machine learn?

那么机器到底是如何自主学习的？

There are many different ways to build self-teaching programs.

创建自学程序的方法有很多种。

But they all rely on the three basic types of machine learning:

但它们都依赖三种基础机器学习方法：

unsupervised learning, supervised learning, and reinforcement learning.

无监督学习、监督学习，和强化学习。

To see these in action,

来看看它们的真实运作情况，

let’s imagine researchers are trying to pull information

先想象一下，研究员正尝试 从一组包含了几千名患者的

from a set of medical data containing thousands of patient profiles.

医疗资料中提取信息。

First up, unsupervised learning.

首先，是无监督学习。

This approach would be ideal for analyzing all the profiles

此方法适用于通过分析所有资料

to find general similarities and useful patterns.

来获取大体的相似性和有用的模式。

Maybe certain patients have similar disease presentations,

也许某些患者有类似的疾病表现，

or perhaps a treatment produces specific sets of side effects.

抑或治疗会产生特定的副作用。

This broad pattern-seeking approach can be used to identify similarities

这种广泛的模式寻找方法

可以用来识别患者 档案之间的相似之处，

between patient profiles and find emerging patterns,

并发现新出现的模式，

all without human guidance.

所有这些都无需人类的指导。

But let's imagine doctors are looking for something more specific.

但让我们想象一下医生 在寻找更具体的东西。

These physicians want to create an algorithm

这些医生想要创建一种算法

for diagnosing a particular condition.

来诊断特定的疾病。

They begin by collecting two sets of data—

他们首先收集两组数据：

medical images and test results from both healthy patients

健康患者和确诊患者的 医学图像和测试结果。

and those diagnosed with the condition.

Then, they input this data into a program

然后，他们将这些数据输入一个程序，

designed to identify features shared by the sick patients

该程序旨在识别患病患者

but not the healthy patients.

而非健康患者共有的特征。

Based on how frequently it sees certain features,

根据看到某些特征的频率，

the program will assign values to those features’ diagnostic significance,

该程序将对这些特征的诊断意义赋值，

generating an algorithm for diagnosing future patients.

生成一种用于诊断未来患者的算法。

However, unlike unsupervised learning,

然而，与无监督学习不同的是，

doctors and computer scientists have an active role in what happens next.

医生和计算机科学家在接下来 发生的事情中扮演着重要的角色。

Doctors will make the final diagnosis

医生将做出最终诊断，

and check the accuracy of the algorithm’s prediction.

并检查算法预测的准确性。

Then computer scientists can use the updated datasets

然后，计算机科学家 可以使用更新的数据集

to adjust the program’s parameters and improve its accuracy.

来调整程序的参数，并提高其准确性。

This hands-on approach is called supervised learning.

这种亲身实践的方法被称为监督学习。

Now, let’s say these doctors want to design another algorithm

现在，假设这些医生想设计另一种算法

to recommend treatment plans.

来推荐治疗方案。

Since these plans will be implemented in stages,

由于这些计划将分阶段实施，

and they may change depending on each individual's response to treatments,

并可能根据每个人 对治疗的反应而改变，

the doctors decide to use reinforcement learning.

医生决定使用强化学习。

This program uses an iterative approach to gather feedback

这个程序使用一种迭代的方法 来收集关于哪种药物、

about which medications, dosages and treatments are most effective.

关于哪种药物、剂量和治疗 最有效的反馈。

Then, it compares that data against each patient’s profile

然后，它将这些数据 与每个患者的档案进行比较，

to create their unique, optimal treatment plan.

以创建他们独特的、最佳的治疗方案。

As the treatments progress and the program receives more feedback,

随着治疗的进展 和程序收到更多的反馈，

it can constantly update the plan for each patient.

它可以不断更新每个患者的计划。

None of these three techniques are inherently smarter than any other.

这三种技术中没有哪一种 天生就比其他的更聪明。

While some require more or less human intervention,

虽然有些需要或多或少的人工干预，

they all have their own strengths and weaknesses

但它们都有自己的优缺点，

which makes them best suited for certain tasks.

这使它们最适合某些任务。

However, by using them together,

然而，通过将它们一起使用，

researchers can build complex AI systems,

研究人员可以构建 复杂的人工智能系统，

where individual programs can supervise and teach each other.

其中单个程序可以相互监督和指导。

For example, when our unsupervised learning program

例如，当我们的非监督学习程序

finds groups of patients that are similar,

发现相似的患者群体时，

it could send that data to a connected supervised learning program.

它可以将这些数据发送到 连接的监督学习程序。

That program could then incorporate this information into its predictions.

然后该程序将这些信息纳入其预测。

Or perhaps dozens of reinforcement learning programs

或者，数十个强化学习项目

might simulate potential patient outcomes

可能会模拟潜在的患者结果，

to collect feedback about different treatment plans.

以收集关于不同治疗计划的反馈。

There are numerous ways to create these machine-learning systems,

有许多方法可以创建 这些机器学习系统，

and perhaps the most promising models

而最有希望的模型可能是

are those that mimic the relationship between neurons in the brain.

那些模拟大脑神经元之间关系的模型。

These artificial neural networks can use millions of connections

这些人工神经网络 可以使用数百万个连接

to tackle difficult tasks like image recognition, speech recognition,

来处理图像识别、语音识别

and even language translation.

甚至语言翻译等困难任务。

However, the more self-directed these models become,

然而，这些模型越自我指导，

the harder it is for computer scientists

计算机科学家就越难确定

to determine how these self-taught algorithms arrive at their solution.

这些自学的算法是如何得到 它们的解决方案的。

Researchers are already looking at ways to make machine learning more transparent.

研究人员已经在寻找 让机器学习更透明的方法。

But as AI becomes more involved in our everyday lives,

但随着人工智能越来越多地 参与到我们的日常生活中，

these enigmatic decisions have increasingly large impacts

这些神秘的决定对我们的工作、

on our work, health, and safety.

健康和安全产生越来越大的影响。

So as machines continue learning to investigate, negotiate and communicate,

因此，在机器继续学习调查、 谈判和沟通的同时，

we must also consider how to teach them to teach each other to operate ethically.

我们也必须考虑如何教它们 彼此如何合乎道德地操作。