原創譯文|如何區分人工智慧、機器學習與深度學習？

轉載聲明

本文為燈塔大數據原創內容，歡迎個人轉載至朋友圈，其他機構轉載請在文章開頭標註：

現在科技行業的術語產生速度令人驚愕，幾乎每天都有新名詞被創造出來。最近一段時間人們討論最多的莫過於人工智慧、機器學習和深度學習了。

很多公司已經果斷採取措施，開發人工智慧、機器學習和深度學習方面的應用。這種商業氛圍由來已久，驅動這些公司爭先恐後開發新技術的既有擔心落後於人的恐懼，也有對技術改變世界的憧憬和希望。

市場競爭的巨大壓力，讓每個公司都感受到了一股強烈的緊張感和急迫感，所有商業領袖都害怕一個不小心就錯過了「風口」，落後於行業發展趨勢。

人工智慧和機器學習的概念並不是「新鮮玩意」，現在它們已經成為計算機行業最令人興奮的名詞，也似乎將給整個商界帶來顛覆性改變。

但是為什麼現在人工智慧會這麼火呢？不少專家相信，這是因為50年前人工智慧就宣稱能夠幫助人類解決現實問題，而50年後的今天，它真的做到了。

人工智慧、機器學習和深度學習正在改變整個科技世界，但是這些技術的發展全都得益於數據學的發展和過去在數據儲存、計算和分析上的巨大投入。現在的技術進步一部分原因在於學習演算法能夠在越來越多的數據中發現模型。

那麼，我們應該如何區分人工智慧、機器學習和深度學習呢？解釋這三者之間關係最簡單、最直接的方法就是下圖這張同心模型。這張圖中還解釋了三者的定義。

人工智慧所包含的範圍最廣，其次是機器學習，機器學習是人工智慧的子領域，最後是深度學習，就是驅動現在人工智慧蓬勃發展的技術。

人工智慧：三者中含義最廣泛的術語，包括使用邏輯、如果-那麼規則、決策樹的能夠模擬人類智力的所有技術（包含機器學習和深度學習）

機器學習：人工智慧的子領域，包括了能夠使機器改進任務體驗的所有深奧統計技術，包含了深度學習

深度學習：機器學習的子領域，由多種演算法組成，能夠讓軟體的多層級神經網路接受大量數據的「訓練」，提高在語音、圖像識別等任務上的表現

機器學習是人工智慧技術的核心概念，通過模擬人類的決策過程來搭建神經網路，解決現實世界中的問題。

而深度學習是機器學習工具和技術的子領域，深度學習的應用非常廣泛，幾乎涵蓋了所有需要「思考」的應用情景，既包括人類的思考，也包括虛擬的「思維」。

機器學習的應用情景更適合公司和企業，因為機器學習能夠解決一些實際的商業問題，如利用監督式學習模型（如回歸模型和分級模型）來做出預測，或者利用無監督模型（如群集模型）來發現未知領域。

深度學習是機器學習的一部分，在一些應用領域也取得了令人矚目的發展，如模式識別、圖像分類、自然語言處理、自動駕駛等等。

相較於深度學期，隨機森林（random forests）和梯度推進（gradientboosting）等機器學習技術在解決商業問題時表現更好。

深度學習是嘗試學習具有多層神經網路的大數據集多層級特徵，然後做出預測性決策。這就意味著深度學習包含了兩個階段：第一步，要利用大量輸入數據來「訓練」神經網路；第二步，就是用這個接受過「訓練」的神經網路來「推測」，預測新數據的分佈。

由於涉及的參數非常多，訓練數據集規模也相當大，所以在神經網路的訓練階段，對參與訓練的計算機算力要求非常的高。

下圖總結了深度學習「訓練」和「推測」階段的流程：

英文原文

Differentiating between AI, MachineLearning and Deep Learning

Many companies are moving decisively todevelop capabilities based on AI, machine learning and deep learning. Intime-honored business fashion, the motivation is a combination of fear andhope.

Competitive pressures are spurring companies on, and there is a sense ofurgency amongst many enterprise thought leaders about not falling behind.

Artificial intelligence (AI) and machinelearning-decades-old technologies that are now electrifying the computingindustry-for all intents and purposes, seem to be in the process oftransforming corporate America. But why is AI so hot right now?

Many expertsbelieve it』s because, after 50 years of promises that AI was going to solvecritical problems, it』s finally working.

AI, machine learning and deep learning aretransforming the entire world of technology, but these technologies are onlymaking headway now due to the proliferation of data and the investments beingmade in storage, compute and analytics solutions.

Much of this progress is dueto the ability of learning algorithms to spot patterns in larger and largeramounts of data.

At first glance, when looking out over theglobal business landscape, some companies might be considered as「under-investing」 in computer systems for AI.

Companies first steps should include the 「Five-stepenterprise AIstrategy」. To learn more about ridingthe wave of machine learning and deep learning download this insideHPC specialreport.

With all the quickly evolving jargon in theindustry today, it』s important to be able to differentiate between AI, machinelearning and deep learning. T

he easiest way to think of their relationship isto visualize them as a concentric model, as depicted in the figure to theright, with each term defined.

Here, AI—the idea that came first—has thelargest area, followed by machine learning—which blossomed later and is shownas a subset of AI. Finally, deep learning—which is driving today』s AIexplosion—fits inside both.

Machine learning takes some of the coreideas of AI and focuses them on solving real-world problems with neuralnetworks designed to mimic our own decision-making.

Deep learning focuses evenmore narrowly on a subset of machine learning tools and techniques, and appliesthem to solving just about any problem which requires 「thought」—human or artificial.

Machine learning takes some of the coreideas of AI and focuses them on solving real-world problems with neuralnetworks designed to mimic our own decision-making.

Machine learning is well-suited for problemdomains typically found in the enterprise, like making predictions withsupervised learning methods (e.g. regression and classification), and knowledgediscovery with unsupervised methods (e.g. clustering).

Deep learning is an areaof machine learning that has achieved significant progress in certainapplication areas that include pattern recognition, image classification,natural language processing (NLP), autonomous driving, and so on.

Machinelearning techniques like random forests and gradient boosting often performbetter in the enterprise problem space than deep learning.

Deep learning attempts to learn multiplelevels of features of large data sets with multi-layer neural networks and makepredictive decisions for the new data.

This indicates two phases in deeplearning: first, the neural network is 「trained」 with a large number of inputdata; second, the trained neural network is used for 「inference」 designed tomake predictions with new data.

Due to the large number of parameters andtraining set size, the training phase requires tremendous amounts ofcomputation power.

The figure below summarizes the roles oftraining versus inferencing in deep learning.

閱讀原文了解更多詳