15%的錯誤率，是比較好的學習方法

如果人工智能可以作為參考的話，那么15%的錯誤率，是最好的學習方法

Getting everything right all of the time might sound like the ideal scenario, but such a perfect success rate can mean you're not actually learning anything new.

在任何時候都做正確的事情可能聽起來像是一個理想的場景，但是如此完美的成功率可能意味著你實際上沒有學到任何新東西。

To make sure you're learning at the optimal rate, new research finds you should be aiming to fail around 15 percent of the time – or 15.87 percent of the time, to be exact.

為了確保你以最佳的速度學習，新的研究發(fā)現(xiàn)，你應該把失敗的目標定在15%左右，確切地說，是15.87%。

These findings could have implications for training courses, teaching in classrooms, and everywhere that learning happens. It's that sweet spot between finding something too easy and too difficult.

這些發(fā)現(xiàn)可能會對培訓課程、課堂教學以及學習發(fā)生的任何地方產(chǎn)生影響。這是在太容易和太困難之間找到平衡點。

"These ideas that were out there in the education field – that there is this 'zone of proximal difficulty', in which you ought to be maximising your learning – we've put that on a mathematical footing," says psychologist Robert Wilson from the University of Arizona.

亞利桑那大學的心理學家羅伯特·威爾遜說:“在教育領域，有這樣一個‘最近困難區(qū)’，在這個區(qū)域里，你應該最大限度地提高你的學習能力。我們已經(jīng)把它建立在數(shù)學的基礎上。”

To come up with the 15/85 percent split, Wilson and his colleagues ran a series of machine learning experiments. The experiments were designed to teach computers how to do simple tasks, such as putting patterns into categories, or recognising the difference between odd and even numbers.

為了得出15/85%的比例，威爾遜和他的同事進行了一系列的機器學習實驗。這些實驗的目的是教計算機如何做一些簡單的任務，比如將模式分類，或者識別奇數(shù)和偶數(shù)之間的區(qū)別。

The computer systems learnt fastest, the researchers found, when they were making the right call 85 percent of the time. That figure seems to match up with previous studies carried out with animals, too.

研究人員發(fā)現(xiàn)，當計算機系統(tǒng)有85%的正確率時，學習速度最快。這一數(shù)字似乎與之前對動物進行的研究相符。

According to the team, this sort of split is most likely to apply to humans when it comes to perceptual learning, where we gradually learn through experience and examples (not unlike a machine learning algorithm).

據(jù)研究小組稱，在感知學習方面，這種分裂最有可能適用于人類，在感知學習中，我們通過經(jīng)驗和例子逐步學習(與機器學習算法并無不同)。

Take a radiologist learning to tell the difference between images of tumors and non-tumors, for example: at a level that's too easy, the radiologist would identify 100 percent of the images correctly. At a level that's too difficult, that might drop to somewhere around 50 percent.

以一位放射科醫(yī)生學習區(qū)分腫瘤和非腫瘤圖像為例：在一個過于簡單的級別上，放射科醫(yī)生可以100%正確地識別圖像。在一個太難的水平上，可能會下降到50%左右。

But if the radiologist is correctly identifying 85 percent of the images and making mistakes with the other 15 percent, that could be the spot where the learning rate is the fastest.

但是，如果放射科醫(yī)生正確識別85%的圖像，而對另外15%的圖像出錯，那么這可能是學習速度最快的地方。

Of course, as we gain more knowledge, that difficulty level needs to be adjusted again, to keep the learning task at just the right level in terms of how challenging it is.

當然，隨著我們獲得更多的知識，這個難度水平需要再次調(diào)整，以使學習任務在多大的挑戰(zhàn)性方面保持在正確的水平。

The researchers are also keen to point out that their study only covers basic, binary choices – it doesn't necessarily follow that we should all be aiming for an 85 percent grade in our future exams.

研究人員還指出，他們的研究只涵蓋了基本的二元選擇——這并不一定意味著我們在未來的考試中都應該爭取85%的分數(shù)。

More research is going to be needed to figure out how this applies more broadly to education, outside of computer algorithms. For now though, it's a good starting point for finding that balance between something that's so easy we get bored, and so difficult we give up – a quandary that educators have been thinking about for a long time.

我們還需要更多的研究來弄清楚這是如何在計算機算法之外更廣泛地應用于教育的。不過，就目前而言，這是一個很好的起點，有助于在容易讓我們感到無聊、也很難讓我們放棄的事情之間找到平衡——這是教育工作者長期以來一直在思考的一個窘境。

"If you are taking classes that are too easy and ace-ing them all the time, then you probably aren't getting as much out of a class as someone who's struggling but managing to keep up," says Wilson.

威爾遜說：“如果你所上的課程太簡單，而且總是名列前茅，那么你在一門課上獲得的收益可能不如一個正在努力奮斗但又設法跟上進度的人。”。

"The hope is we can expand this work and start to talk about more complicated forms of learning."

“希望我們能擴大這項工作，開始討論更復雜的學習形式。”

The research has been published in Nature Communications.

這項研究已發(fā)表在《自然通訊》雜志上。