TensorFlow的未來？2018.03.NO.1(Note )

www.quora.com

note：

I am going to take the contrarian viewpoint here.

作者是一個機器學習方面的工程師，在他的眼中，TensorFlow在今天的確是一個應用廣泛的深度學習平台。而TensorFlow今天的輝煌，是和全世界的學術界和工業研究屆早先對TensorFlow的接受有著巨大的關聯的。而至於作者，他最初使用TensorFlow的原因在於，當時，他需要調試出一個外包程序；而這個外包程序屬於一些科研人員，他們的論文上個星期剛剛發表在arXiv上。順便一提，（利用TensorFlow）這也是很多計算機視覺方面的研究人員，一直工作事倍功半，宵衣旰食的原因了。

Yes,

TensorFlow is superior today, largely because of early adoption by academic and

不過，TensorFlow在機器學習模型的應用方面有這樣一個特點，他通過大量的已知的係數集代替了大規模的代碼庫。在探索和嘗試以後，一旦你通過TensorFlow找到了你理想中的深度學習模型，那麼，接下來你可以通過CNTK和MXNET非常方便的再次實現它。

But.

到目前為止，絕大多數大型企業甚至是創業者們，都不會在自己的實體伺服器中跑他們的深度學習網路。取而代之的是第三方的雲服務（third party cloud services），而在雲應用方面，Amazon和微軟相對於Google有著非常明顯的優勢，而且這一種優勢還在不斷發展。不僅如此，在企業競爭中，Google在工程實踐方面的話語權也非常小，反之，微軟簡直是這方面的王者。

Now,

基於上面的兩個事實，咱們就可以看到TensorFlow的最大的隱憂啦！

Combine the above two facts, and you begin to see a major threat to TensorFlow!

深度學習，對於企業級（enterprise-level）的機器學習還不是那麼重要，在這個領域中，邏輯方程（logistic regression）仍然統治著這一片江山，這樣一來，成熟的團隊就會離開Amazon或者微軟的云為Google的雲服務。事實上，只有很少的雲工程真的完全只應用機器學習理論。所以，當一個團隊真的需要一些機器學習模型代碼來完成某些功能的時候，他們會更願意在TensorFlow上寫出一個原型，然後再用CNTK或者MXNET來進行工程實現。非常明顯，同時三個公司一起為某個目標專門設計最好的演算法框架或者計算機硬體，用MXNET寫的模型便可以更好的利用AWS的機器。他們甚至會通過低價提供本機DL框架的方式，來鼓勵他們這樣做。

Deep learning isn』t so important for enterprise-level machine learning yet, where logistic regression still rules, that development teams will want to leave Amazon or Microsoft』s superior cloud offering for Google』s. In fact, machine learning as a whole provides a small fraction（分數，份額） of the utility（應用，實用） of cloud services to businesses. Thus for the deep learning components that most teams do need in their systems, they might very well want to prototype（原型，樣板） on TensorFlow but shift（移動） to CNTK/MXNET for production. Obviously, in the meanwhile all three companies will try to specialize their framework and hardware（硬體） to work best together, such that MXNET models might be able to best utilize（利用） AWS machines. They might even incentivize（以物質激勵） this, by offering lesser prices for using their native DL framework.

兩三年後，一旦圍繞著深度學習的科研狂熱冷卻一點，開發者們就會學會企業們在招聘啟事上所希求的人才所熟悉的小把戲。這樣一來，便會開始一個更新迭代開發平台的新的正反饋循環（當然新的平台會比TensorFlow好多啦！），無論怎麼說，深度學習領域內，新的事物一個比一個驚艷的出現在我們眼前，而我們無疑是這之中的見證者。也許這之中的某一者，一個全新全新的架構，會取得令人驚訝的成就，直接成為人類知識體系皇冠上的明珠。

Once the research environment around deep learning cools down a bit in a couple of years, developers will learn the tools that most enterprises put in their job adverts. This could start a feedback cycle that starts improving these rival frameworks at a faster pace than TF - anyways, we keep seeing deep learning ideas get superseded by newer ones, which may enable even an entirely new framework to skip a couple generations and get in the top league!

所以，我感覺現在說誰會是架構戰爭中的勝利者為時過早。斷言（claim）戰爭已經結束而TensorFlow贏得了這場戰爭是膚淺的，事實上，戰爭才剛剛開始；而且這些戰爭，不會繼續發生在實驗室或者那些狹小的宿舍內，不會取決於某些個體或者小群體偶然的奇思妙想。而在這些商業活動中，毫無疑問會有一定數量的資金被應用在雲服務的開發上，而這些雲服務會變得更加完備和系統化，當構架的拼圖遊戲變得越來越方便，DL主流框架便會越來越少的進入革新者們的視野中。

So, my feeling is that it』s too early to say who will win in the framework wars. It』s easy to claim that the wars are over and TF has won. But in fact, the wars have only just begun; and those wars won』t be fought in academic labs and dorm（宿舍） rooms. It』s the businesses that will eventually dole out（少量發放） the money for cloud services - and those services will likely be bought as a package - and the ease of use of a tiny piece of the puzzle, the DL framework will be a non-issue（不被考慮） in that decision.

兩個我看好的方向，都包括：

1. 要麼雲開發商在他們的原始架構上支持一個更加頂層，更加普適的標準化的結構，而我們已經在Keras上見到這一點了。當然這對於TensorFlow和其他專註底層構建的結構絕不會是好消息。
2. 而另一個可能的方向上，邊緣集成智能架構可能會取得對雲智能架構的優勢地位，而英偉達也已經把他們的結構呈現給了世人，更可貴的是，他們並不想把自己完全局限在Google的架構內，雖然Google已經通過TPU業務開始逐漸影響他們。

Two alternative universes that I completely neglect（忽略，忽視） in the above include ： one in which every cloud provider starts supporting a higher, relatively 「neutral」（原意中立，譯者理解為普適） layer on top of their native framework — which gets standardized. We are beginning to see that happen with Keras. But then again, that means lesser importance for TensorFlow and all the other lower-level frameworks.

Another possibility is that edge intelligence becomes a bigger deal than cloud intelligence, in which case we might well see NVidia coming up with their own framework, as they might not want to limit themselves to Google』s framework, which has already attempted to enter NVidia』s business through its TPUs!

Original answer（原文如下）：

I am going to take the contrarian viewpoint here.

Yes,TensorFlow is superior today, largely because of early adoption by academic and

But，One aspect of machine learning models is that they replace huge codebases with huge sets of coefficients. After all the research and hammering, once you arrive at your final deep learning model, it doesn』t take much time to reimplement it with CNTK or MXNET.

Now,most big enterprises and even startups won』t be running their deep networks on premises. They will be using third party cloud services, and in cloud Amazon and Microsoft have a significant lead over Google, and that lead is growing. Not only that, Google has relatively little experience talking to the enterprise market, whereas Microsoft is king when it comes to enterprise.

Combine the above two facts, and you begin to see a major threat to TensorFlow!

Deep learning isn』t so important for enterprise-level machine learning yet, where logistic regression still rules, that development teams will want to leave Amazon or Microsoft』s superior cloud offering for Google』s. In fact, machine learning as a whole provides a small fraction of the utility of cloud services to businesses. Thus for the deep learning components that most teams do need in their systems, they might very well want to prototype on TensorFlow but shift to CNTK/MXNET for production. Obviously, in the meanwhile all three companies will try to specialize their framework and hardware to work best together, such that MXNET models might be able to best utilize AWS machines. They might even incentivize this, by offering lesser prices for using their native DL framework.

Once the research environment around deep learning cools down a bit in a couple of years, developers will learn the tools that most enterprises put in their job adverts. This could start a feedback cycle that starts improving these rival frameworks at a faster pace than TF - anyways, we keep seeing deep learning ideas get superseded by newer ones, which may enable even an entirely new framework to skip a couple generations and get in the top league!

So, my feeling is that it』s too early to say who will win in the framework wars. It』s easy to claim that the wars are over and TF has won. But in fact, the wars have only just begun; and those wars won』t be fought in academic labs and dorm rooms. It』s the businesses that will eventually dole out the money for cloud services - and those services will likely be bought as a package - and the ease of use of a tiny piece of the puzzle, the DL framework will be a non-issue in that decision.

Two alternative universes that I completely neglect in the above include … one in which every cloud provider starts supporting a higher, relatively 「neutral」 layer on top of their native framework — which gets standardized. We are beginning to see that happen with Keras. But then again, that means lesser importance for TensorFlow and all the other lower-level frameworks. Another possibility is that edge intelligence becomes a bigger deal than cloud intelligence, in which case we might well see NVidia coming up with their own framework, as they might not want to limit themselves to Google』s framework, which has already attempted to enter NVidia』s business through its TPUs!