如果要選2016年的熱詞，“機(jī)器學(xué)習(xí)”肯定當(dāng)仁不讓。似乎每家公司都在自我介紹里沾點(diǎn)機(jī)器學(xué)習(xí)的邊，而且確實(shí)效果不錯(cuò)。

據(jù)云安全公司CloudPassage的卡森·斯威特介紹，很多企業(yè)都想用機(jī)器學(xué)習(xí)的工具解決問題，雖然并不十分了解工具的作用。

不久前在舊金山，斯威特跟另外兩家網(wǎng)絡(luò)安全公司管理者在結(jié)構(gòu)安全論壇上發(fā)言，解釋了一些機(jī)器學(xué)習(xí)方面常見的誤解。其中之一就是將機(jī)器學(xué)習(xí)等同于“人工智能”（也是今年熱詞榜的大熱門）。

威脅監(jiān)測(cè)公司Sqrrl的馬克·特倫佐尼解釋說，人工智能相當(dāng)于造一個(gè)大腦，但并不能產(chǎn)生確定結(jié)果（即產(chǎn)生可預(yù)期的結(jié)果），所以惡作劇的人通過故意挑逗微軟旗下的人工智能聊天軟件，能讓機(jī)器人說出種族歧視言論。

另一方面，機(jī)器學(xué)習(xí)會(huì)產(chǎn)生可控的回應(yīng)和有效的預(yù)測(cè)。機(jī)器學(xué)習(xí)能從海量數(shù)據(jù)中找出規(guī)則，甚至能將結(jié)果以可視化圖標(biāo)方式呈現(xiàn)并突出最重要的信息。

但機(jī)器學(xué)習(xí)也有重要限制，其中最明顯的一個(gè)就是仍然需要人類提出合適的問題。

“機(jī)器學(xué)習(xí)鋒利如矛之尖，但得先建立模型才能為安全分析師所用，”特倫佐尼表示。

凱文·馬哈菲任職于移動(dòng)安全公司Lookout（曾找出臭名昭著的iPhone漏洞），他也表示企業(yè)要在機(jī)器學(xué)習(xí)算法中輸入“干凈的數(shù)據(jù)”。如果把一堆堆的隨機(jī)信息扔進(jìn)去，結(jié)果只會(huì)是“垃圾進(jìn)去，垃圾出來”。

回答論壇主持人《財(cái)富》雜志的喬納森·瓦尼恩提問時(shí)，馬哈菲還解釋了“機(jī)器學(xué)習(xí)”與“深度學(xué)習(xí)”的區(qū)別。關(guān)鍵在于規(guī)模：深度學(xué)習(xí)是指最近計(jì)算機(jī)能力方面的突破，實(shí)現(xiàn)深度學(xué)習(xí)的花費(fèi)巨大，可供機(jī)器學(xué)習(xí)工具利用成百上千萬的參數(shù)探索可能性。

不過馬哈菲也提醒道，雖然深度學(xué)習(xí)是一項(xiàng)精深的技術(shù)，很多企業(yè)還是應(yīng)該先了解些機(jī)器學(xué)習(xí)的基礎(chǔ)。

“我們已經(jīng)開始問‘今早的冰沙里要加多少甘藍(lán)’，而多數(shù)企業(yè)還停留在起床抽根煙的階段，”馬哈菲打趣說。（財(cái)富中文網(wǎng)）

譯者：Pessy

審校：夏林

If you had to pick a tech industry buzzword for 2016, “machine learning” would be a good choice. Every other company, it seems, is packing the phrase into their pitches, and it’s having an effect.

According to Carson Sweet of cloud security firm CloudPassage, many companies are asking for machine learning tools to solve problems—even if they don’t have a clear idea of what these tools can do.

Speaking at the Structure Security conference on Tuesday in San Francisco, Sweet and executives from two other cyber-security firms explained some common misconceptions about machine learning. One of these is that machine learning is the same thing as “artificial intelligence” (another top candidate for buzzword of the year).

As Mark Terenzoni of threat detection firm Sqrrl explained, AI is like building a brain, but one that is unable to produce deterministic outcomes (ones that will produce a predictable outcome) — that’s why mischief makers were able to manipulate Microsoft’s AI chat bot into spewing racist comments.

Machine learning, on the other hand, results in predictable responses and useful predictions. It can detect patterns in giant amounts of data and even present the results in visual graphics that highlight the most salient information.

But there are important limits to machine learning, and the biggest of these is that it still requires humans to frame the right question.

“Machine learning is the tip of spear, but you have to do a lot of curating to create a model that makes sense to a security analyst,” said Terenzoni.

Kevin Mahaffey of mobile security firm Lookout (which helped expose that notorious iPhone bug) likewise noted that firms need “clean data” to feed machine learning algorithms. Simply shoveling random stacks of information, he said, will produce a “garbage in, garbage out” result.

Mahaffey, in response to a question from moderator Jonathan Vanian of Fortune, also clarified the difference between “machine learning” and “deep learning.” It turns to be a question of scale: deep learning describes the recent breakthroughs in computer power and cost that makes it possible for machine learning tools to explore millions of parameters.

Mahaffey, however, cautioned that while deep learning represents a remarkable technology, many firms still need to learn the basics of machine learning.

“We’re asking ‘how many grams of Kale do you want in your smoothies this morning’—while most organizations are still smoking a pack of cigarettes a day,” joked Mahaffey.