How does artificial intelligence learn? - Briana Brownell

譯者: Jing-Ai Huang 審譯者: Helen Chang

Today, artificial intelligence helps doctors diagnose patients,

現在人工智慧幫醫生診斷病人，

pilots fly commercial aircraft, and city planners predict traffic.

幫飛行員駕駛商用飛機， 幫規劃師預測交通。

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

但是不管人工智慧在做什麼，

設計它們的電腦科學家們 可能並不知道它們到底在做什麼。

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

This is because artificial intelligence is often self-taught,

這是因為人工智慧是自學成才的，

working off a simple set of instructions

從一組簡單的指令

to create a unique array of rules and strategies.

創造出一組獨特的規則和策略。

So how exactly does a machine learn?

那麼機器到底是如何學習的呢？

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

建立自學程式有很多不同的方式。

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

但它們都依賴於三種 基本類型的機器學習。

unsupervised learning, supervised learning, and reinforcement learning.

非監督式學習、監督式學習 和強化式學習。

To see these in action,

舉例來說，

let’s imagine researchers are trying to pull information

讓我們想像研究人員

從一組包含數千份患者資料的 醫療數據中蒐集數據。

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

First up, unsupervised learning.

首先是非監督式學習，

This approach would be ideal for analyzing all the profiles

這種方法常被用來分析所有的檔案

to find general similarities and useful patterns.

找規律性和有用的特徵。

Maybe certain patients have similar disease presentations,

也許某些患者有相似的臨床表現

or perhaps a treatment produces specific sets of side effects.

或一種治療方法會產生特定的副作用。

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

這種廣泛的搜索模式

在沒有人的指導下

between patient profiles and find emerging patterns,

可以在患者的檔案中 找出新興的特徵。

all without human guidance.

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

但我們想像一下， 醫生要找的是更具體的東西。

These physicians want to create an algorithm

這些醫生希望建立一個

for diagnosing a particular condition.

可用於診斷某一特定症狀的演算法。

They begin by collecting two sets of data—

他們首先蒐集了兩組數據，

medical images and test results from both healthy patients

健康的人和那些被診斷出 有病情的患者的

and those diagnosed with the condition.

醫學影像和檢驗結果。

Then, they input this data into a program

然後，他們將這些數據輸入到一個

designed to identify features shared by the sick patients

設計於辨別患者有

but not the healthy patients.

但是健康的人沒有共同特徵的程式。

Based on how frequently it sees certain features,

根據看到某些特徵的頻率，

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

程式將為這些特徵的診斷意義賦值。

generating an algorithm for diagnosing future patients.

產生用於診斷未來的病人的算法。

However, unlike unsupervised learning,

但是與非監督式學習不同的是 醫生和計算機科學家

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

對接下來發生的事情 扮演著重要的角色。

Doctors will make the final diagnosis

醫生會做出最終診斷

and check the accuracy of the algorithm’s prediction.

並檢查算法預測的準確性。

Then computer scientists can use the updated datasets

然後，計算機科學家 可以使用更新的數據集

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

調整程式的參數，提高準確性。

This hands-on approach is called supervised learning.

這種實踐的方法叫做監督式學習。

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

假設這些醫生想設計另一種算法

to recommend treatment plans.

來建議治療方法，

Since these plans will be implemented in stages,

由於這些計劃將分階段實施，

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

而且它們可能會根據每個人 對治療的反應而改變，

the doctors decide to use reinforcement learning.

醫生們會決定使用強化式學習。

This program uses an iterative approach to gather feedback

這個程式使用迭代法來收集反饋意見，

about which medications, dosages and treatments are most effective.

哪些藥物、劑量和治療方法最有效。

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

然後，它將這些數據 與每個病人的檔案進行比較。

to create their unique, optimal treatment plan.

來製造出獨特的最佳治療方案。

As the treatments progress and the program receives more feedback,

隨著治療的進展和項目 收到更多的反饋。

it can constantly update the plan for each patient.

它可以不斷更新每個病人的計劃。

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

三種技術之一並不比其他兩種聰明。

While some require more or less human intervention,

雖然有些需要或多或少的干涉，

they all have their own strengths and weaknesses

但是尺有所短，寸有所長，

which makes them best suited for certain tasks.

這讓它們各有自己適合的任務。

However, by using them together,

然而，一起使用它們的話

researchers can build complex AI systems,

研究人員可以構建 複雜的人工智慧系統，

where individual programs can supervise and teach each other.

讓它們互相監督和教導。

For example, when our unsupervised learning program

例如，當我們的非監督式學習程式

finds groups of patients that are similar,

找到一群相似的患者，

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

它可以將這些數據發送到 一個有連結的監督式學習程式中。

That program could then incorporate this information into its predictions.

然後，那個程式就可以 將資訊納入預測中。

Or perhaps dozens of reinforcement learning programs

或許這幾十個強化式學習程式

might simulate potential patient outcomes

可以模擬病患會有的結果

to collect feedback about different treatment plans.

以收集對不同治療方案的反饋。

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

創建這些機器學習系統的方法有很多，

and perhaps the most promising models

而最有前途的系統

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

是那些能模仿大腦神經元之間關係的。

These artificial neural networks can use millions of connections

這些人工神經網絡 可以使用數以百萬計的連接。

to tackle difficult tasks like image recognition, speech recognition,

以應對圖像識別、 語音識別等困難任務，

and even language translation.

甚至語言翻譯。

However, the more self-directed these models become,

然而，這些模式越是自我導向。

the harder it is for computer scientists

計算機科學家越難以確定

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

這些自學算法是如何得出 其解決方案的。

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

研究人員已經在研究 如何讓機器學習更加透明。

But as AI becomes more involved in our everyday lives,

但隨著人工智慧越來越頻繁得 參與在我們的日常生活中。

these enigmatic decisions have increasingly large impacts

這些神祕的決定

on our work, health, and safety.

對我們的工作、健康和安全 產生了越來越大的影響

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

所以隨著機器不斷學習、 調查、協商和交流。

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

我們必須考慮該如何教導它們， 讓它們互相教導經營道德。