Today, artificial intelligence helps doctors diagnose patients, pilots fly commercial aircraft, and city planners predict traffic. But no matter what these AIs are doing, the computer scientists who designed them likely don’t know exactly how they’re doing it. This is because artificial intelligence is often self-taught, working off a simple set of instructions to create a unique array of rules and strategies. So how exactly does a machine learn?
現在人工智慧幫醫生診斷病人, 幫飛行員駕駛商用飛機, 幫規劃師預測交通。 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。 這是因為人工智慧是自學成才的, 從一組簡單的指令 創造出一組獨特的規則和策略。 那麼機器到底是如何學習的呢?
There are many different ways to build self-teaching programs. But they all rely on the three basic types of machine learning: unsupervised learning, supervised learning, and reinforcement learning. To see these in action, let’s imagine researchers are trying to pull information from a set of medical data containing thousands of patient profiles.
建立自學程式有很多不同的方式。 但它們都依賴於三種 基本類型的機器學習。 非監督式學習、監督式學習 和強化式學習。 舉例來說, 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。
First up, unsupervised learning. This approach would be ideal for analyzing all the profiles to find general similarities and useful patterns. Maybe certain patients have similar disease presentations, or perhaps a treatment produces specific sets of side effects. This broad pattern-seeking approach can be used to identify similarities between patient profiles and find emerging patterns, all without human guidance.
首先是非監督式學習, 這種方法常被用來分析所有的檔案 找規律性和有用的特徵。 也許某些患者有相似的臨床表現 或一種治療方法會產生特定的副作用。 這種廣泛的搜索模式 在沒有人的指導下 可以在患者的檔案中 找出新興的特徵。
But let's imagine doctors are looking for something more specific. These physicians want to create an algorithm for diagnosing a particular condition. They begin by collecting two sets of data— medical images and test results from both healthy patients and those diagnosed with the condition. Then, they input this data into a program designed to identify features shared by the sick patients but not the healthy patients. Based on how frequently it sees certain features, the program will assign values to those features’ diagnostic significance, generating an algorithm for diagnosing future patients. However, unlike unsupervised learning, doctors and computer scientists have an active role in what happens next. Doctors will make the final diagnosis and check the accuracy of the algorithm’s prediction. Then computer scientists can use the updated datasets to adjust the program’s parameters and improve its accuracy. This hands-on approach is called supervised learning.
但我們想像一下, 醫生要找的是更具體的東西。 這些醫生希望建立一個 可用於診斷某一特定症狀的演算法。 他們首先蒐集了兩組數據, 健康的人和那些被診斷出 有病情的患者的 醫學影像和檢驗結果。 然後,他們將這些數據輸入到一個 設計於辨別患者有 但是健康的人沒有共同特徵的程式。 根據看到某些特徵的頻率, 程式將為這些特徵的診斷意義賦值。 產生用於診斷未來的病人的算法。 但是與非監督式學習不同的是 醫生和計算機科學家 對接下來發生的事情 扮演著重要的角色。 醫生會做出最終診斷 並檢查算法預測的準確性。 然後,計算機科學家 可以使用更新的數據集 調整程式的參數,提高準確性。 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 to recommend treatment plans. Since these plans will be implemented in stages, and they may change depending on each individual's response to treatments, the doctors decide to use reinforcement learning. This program uses an iterative approach to gather feedback about which medications, dosages and treatments are most effective. Then, it compares that data against each patient’s profile to create their unique, optimal treatment plan. As the treatments progress and the program receives more feedback, it can constantly update the plan for each patient. None of these three techniques are inherently smarter than any other. While some require more or less human intervention, they all have their own strengths and weaknesses which makes them best suited for certain tasks. However, by using them together, researchers can build complex AI systems, where individual programs can supervise and teach each other. For example, when our unsupervised learning program finds groups of patients that are similar, it could send that data to a connected supervised learning program. That program could then incorporate this information into its predictions. Or perhaps dozens of reinforcement learning programs might simulate potential patient outcomes to collect feedback about different treatment plans.
假設這些醫生想設計另一種算法 來建議治療方法, 由於這些計劃將分階段實施, 而且它們可能會根據每個人 對治療的反應而改變, 醫生們會決定使用強化式學習。 這個程式使用迭代法來收集反饋意見, 哪些藥物、劑量和治療方法最有效。 然後,它將這些數據 與每個病人的檔案進行比較。 來製造出獨特的最佳治療方案。 隨著治療的進展和項目 收到更多的反饋。 它可以不斷更新每個病人的計劃。 三種技術之一並不比其他兩種聰明。 雖然有些需要或多或少的干涉, 但是尺有所短,寸有所長, 這讓它們各有自己適合的任務。 然而,一起使用它們的話 研究人員可以構建 複雜的人工智慧系統, 讓它們互相監督和教導。 例如,當我們的非監督式學習程式 找到一群相似的患者, 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。 然後,那個程式就可以 將資訊納入預測中。 或許這幾十個強化式學習程式 可以模擬病患會有的結果 以收集對不同治療方案的反饋。
There are numerous ways to create these machine-learning systems, and perhaps the most promising models are those that mimic the relationship between neurons in the brain. These artificial neural networks can use millions of connections to tackle difficult tasks like image recognition, speech recognition, and even language translation. However, the more self-directed these models become, the harder it is for computer scientists to determine how these self-taught algorithms arrive at their solution. Researchers are already looking at ways to make machine learning more transparent. But as AI becomes more involved in our everyday lives, these enigmatic decisions have increasingly large impacts on our work, health, and safety. So as machines continue learning to investigate, negotiate and communicate, we must also consider how to teach them to teach each other to operate ethically.
創建這些機器學習系統的方法有很多, 而最有前途的系統 是那些能模仿大腦神經元之間關係的。 這些人工神經網絡 可以使用數以百萬計的連接。 以應對圖像識別、 語音識別等困難任務, 甚至語言翻譯。 然而,這些模式越是自我導向。 計算機科學家越難以確定 這些自學算法是如何得出 其解決方案的。 研究人員已經在研究 如何讓機器學習更加透明。 但隨著人工智慧越來越頻繁得 參與在我們的日常生活中。 這些神祕的決定 對我們的工作、健康和安全 產生了越來越大的影響 所以隨著機器不斷學習、 調查、協商和交流。 我們必須考慮該如何教導它們, 讓它們互相教導經營道德。