深度学习 准备_为企业准备好深度学习

深度学习 准备

计算的历史是一个不断增加抽象的故事 (The history of computing is a tale of increasing abstraction)

No matter which technology you look at, the same pattern somehow repeats itself: a breakthrough happens in a research environment, public or private. At this stage just a very few people are able to understand this discovery and contribute to its refinement, this crowd tends to sign starting with PhD before his name. Expectations are high at this point.

不管你看看哪种技术，同样的模式在某种程度上重演：突破发生在研究环境，公共或私人。 在这个阶段，只有很少的人能够理解这个发现并为它的完善做出贡献。这个人群倾向于以博士的名字开始签字。 在这一点上的期望很高。

After a viable business use of this technology emerges, one might expect a few companies trying to jump in and tame and understand this novel discovery. However the real widespread adoption only happens when the technology is democratised and “ordinary” IT people are able to squeeze the value with relative low effort.More often than not, this democratization force is in the form of abstraction.

在这项技术可行的商业用途出现之后，人们可能会期望一些公司试图介入并驯服并理解这一新发现。 但是，只有在技术民主化之后，真正的广泛采用才会发生，并且“普通” IT人员能够以相对较低的努力来压缩价值。这种民主化力量通常以抽象的形式出现。

Pick SQL, the data “lingua franca”. With a simple (and very expressive) “SELECT FROM WHERE” statement a “plain” data analyst is able to orchestrate hundreds of processes, data files, memory pools, latches and locks interacting simultaneously to retrieve data from a RDBMS. If you are interested in all the things that happen in a database under cover, have a look at this video I created explaining it.

选择SQL，数据为“通用语言”。 借助简单(且非常有表达力的)“ SELECT FROM WHERE”语句，“普通”数据分析人员便能够协调数百个进程，数据文件，内存池，闩锁和锁，同时进行交互以从RDBMS检索数据。 如果您对隐藏在数据库中的所有事情都感兴趣，请观看我创建的解释该视频的视频 。

This level of abstraction finally derives on industry standardisation generating strong network effects.

这种抽象水平最终源自产生强大网络效应的行业标准化。

Deep learning is without any doubt the coolest new kid on the data block. This amazing bio inspired technology has proven to excel at solving some of the more pressing problems, from human level image recognition to voice generation. For instance, have a look at this video I recorded about GPT-3, an impressive language model from openAI with mind blowing results.

毫无疑问，深度学习是数据块上最酷的新手。 事实证明，这项令人惊叹的受生物启发的技术擅长解决一些更迫切的问题，从人类图像识别到语音生成。 例如，看一下我录制的有关GPT-3的视频 ，这是来自openAI的令人印象深刻的语言模型，结果令人震惊。

When it comes to developing these amazing models, Tensorflow has become the de-facto framework for defining and executing deep learning models, and truth to be said, it is a bit complex to handle. First you have to define a computational graph where each node is an operation and the edges are multidimensional arrays referred as tensors.This graph has to be built before running it in a session and then feed the data to previously defined placeholders, this somehow cumbersome way of developing requires advanced mathematical skills that the majority of developers simply do not have. When, back in 2017, Tensorflow 1.0 was conceived at Google as the evolution of the DistBelief framework, the end user persona were groups of highly trained research engineers.

在开发这些惊人的模型时，Tensorflow已成为定义和执行深度学习模型的事实上的框架，说实话，处理起来有点复杂。 首先，您必须定义一个计算图，其中每个节点都是一个操作，并且边缘是称为张量的多维数组。必须在运行会话之前将其构建，然后再将数据馈入先前定义的占位符，这种方式有点麻烦开发工作需要大多数开发人员根本不具备的高级数学技能。 早在2017年，Tensorflow 1.0就在Google构想为DistBelief框架的演进时，最终用户角色是一群训练有素的研究工程师。

Seduced by this hype, I have seen many companies sitting in traditional industries trying to jump in this deep learning wagon with a bit “desperate” moves, from hiring an expensive team of Machine Learning experts with little to none knowledge of the business, to invest in “magical black box” solutions that nobody end using.

在这种炒作的诱惑下，我看到许多传统行业的公司试图通过一些“绝望的”举动跳入这个深度学习的行列，从雇用对业务一无所知的昂贵的机器学习专家团队进行投资。没有人最终使用的“神奇黑匣子”解决方案中。

As we have seen with the SQL example, one of the secret ingredients to enterprise technology adoption is democratization, in other words, this can be expressed by the following question: can your IT engineers develop a deep learning model effectively?

正如我们在SQL示例中看到的那样，采用企业技术的秘密要素之一就是民主化，换句话说，这可以用以下问题来表达：您的IT工程师可以有效地开发深度学习模型吗？

进入Keras：将深度学习从研究转移到企业 (Enter Keras: moving deep learning from research to the enterprise)

Back in 2015, Francois Chollet was, in his free time, doing natural language processing research and found out that there was no reusable open-source implementation of LSTM, so he decided to make one, then he started to code more and more reusable layers, eventually that group of components turned into a framework. Keras was born. Francois decided to open-sourced it and the rest of history.

早在2015年，Francois Chollet在业余时间从事自然语言处理研究，发现没有可重用的LSTM开源实现，因此他决定编写一个，然后开始编写越来越多的可重用层代码，最终那组组件变成了一个框架。 凯拉斯(Keras)出生。 弗朗索瓦(Francois)决定将其以及其他历史资料开源。

Think of the following problem statement: “You need to build a system that given a video stream and a question about what object is in the video, it generates a response”. No easy stuff, right? Frame extraction, NLP question analysis, image recognition, text generation and stacking all those pieces together.

考虑以下问题陈述： “您需要构建一个给出视频流的系统，并询问视频中存在什么对象，它会生成响应”。 没有简单的东西，对吧？ 帧提取，NLP问题分析，图像识别，文本生成以及将所有这些片段堆叠在一起。

Thanks to the incorporation Keras in Tensorflow 2.0, you can solve this problem with just 13 lines of code.

多亏了Tensorflow 2.0中的Keras合并，您只需13行代码即可解决此问题。

Figure 1. Tensorflow 2.0 code with Keras

图1.使用Keras的Tensorflow 2.0代码

Today Keras is the preferred Tensorflow API, it offers a higher-level, more intuitive set of abstractions that make it easy to develop deep learning models regardless of the computational backend used. The critical part is that it opens the deep learning potential to a wider population of developers. Believe me, you are good to go with medium python (more common SDK) knowledge and basic concepts of Machine Learning.

今天，Keras是首选的Tensorflow API，它提供了更高层次，更直观的抽象集，无论使用哪种计算后端，都可以轻松开发深度学习模型。 关键部分是它为更广泛的开发人员打开了深度学习的潜力。 相信我，您会精通python(更常见的SDK)知识和机器学习的基本概念。

If you are a “hardcore” ML engineer, no problem, Keras is just built on top of the Tensorflow backend, as the next image shows, you are free to go as deep as you want to.

如果您是“核心” ML工程师，那么没问题，Keras只是建立在Tensorflow后端之上，如下图所示，您可以随意深入。

Figure 2. Tensorflow building blocks

图2. Tensorflow构建块

Is not only because of Keras that the time for Deep Learning enterprise adoption has finally arrived, with the advent of cloud technology, companies can spin off development sandbox environments in seconds and get access to hyper-specialized deep learning specific hardware such as TPUs without incurring in upfront costs.

不仅是因为Keras，深度学习企业采用的时机终于到了，随着云技术的出现，公司可以在几秒钟内剥离开发沙箱环境，并获得对超专业深度学习特定硬件(例如TPU)的访问权前期费用。

摘要 (Summary)

In summary, as with any technology, we are recently seeing a spike on the use of Deep Learning in the traditional enterprise space, thanks in part to the democratization of its use because of high level and more expressive ways of developing neural networks.

总之，与任何技术一样，我们最近看到了在传统企业空间中使用深度学习的高峰，这在一定程度上要归功于其使用的民主化，这是因为开发神经网络的水平更高且表达方式更丰富。

翻译自: https://medium.com/@velascoluis/deep-learning-ready-for-the-enterprise-76bff91cf6d7

深度学习 准备