The danger of the hype over Deep Learning (Machine Learning)

The danger of the hype over Deep Learning (Machine Learning) is that most of the recent breakthroughs in various application domains are achieved through the improvement of hardware processing capability in combination with the big data and not the true leapfrog development of the science itself. The fundamental of the neural network science probably hasn't changed much in the past 30 years. It is the powerful hardware that has drastically reduced the time taken for the computation of big data. Besides, the result of computation probably will only show you the correlation and not causal relation.

Reference:
[1] https://www.technologyreview.com/s/608911/is-ai-riding-a-one-trick-pony/
[2] https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/

有兴趣钻研或投资人工智能商业项目的朋友，务必理解它的局限性。现在市面上最火热的人工智能分支是深度学习（机器学习）。大概在30年前，当我们还没有这些时髦的名称时，我们称这项技术为人工神经网络。

近年来，深度学习技术所带来的丰硕科研成果和商业应用，基本上是建基于电脑处理器和相关硬件的提升，以及丰厚的大数据 － 因此人工智能系统能够在极短的时间内演算出结果。

然而，在过去的30年里，人工神经网络科学原理本身，并没有太大的改变或呈质的飞跃。

人工智能科研人员坦诚，他们只能在事后下结论，人工智能的演算非常具有洞见、非常精准。但，科研人员无法完全推导出人工智能之所以精准的科学逻辑。

换句话说，迄今为止，由于人们只能靠人工智能来推断出数据的相关性，而非因果关系，我们并没有100巴仙的把握来确定人工智能运算结果的准确性。

工智能两种技术Deep Learning和Reinforcement Learning项目的发展条件

1) 强大的电脑处理器和相关的周边硬体设备 - 否则我们演算一年都还没完成运算结论。
2) 创投领域的大数据 - 所有的电脑技术都遵循Garbage In Garbage Out的原理。没有充裕的相关领域数据，就是在烧钱。
3) 严谨的AB Testing - 由于没有标准演算结果/答案，运用benchmark baseline system的AB testing 非常关键。新的演算法加数据模型的运算结果如果优于baseline system，就算是测试过了第一关。当然还有其他测试标准。
4) 技术人才 - 慎防空有prototype，没有实际技术的团队。一个团队可以拿其他技术演算的结果来充当人工智能的成果。

The Prerequisites to start a large scale Deeping Learning/Reinforcement Learning AI Project

1) Powerful computer processors and the associated peripheral hardware infrastructure - otherwise it is very unlikely we would ever have the chance to complete the computation of the incoming big data stream which continues to grow at exponential rate.

2) Big data of the application domain - All IT systems comply with the Garbage In Garbage Out principle. Without meaningful data, an AI application is with no any superiority.

3) Rigorous AB Testing - Due to the constraint that there is very rarely we could anticipate a standard outcome from the AI computation, AB testing which benchmarks the outcome of the latest result against the baseline system becomes the prudent method to validate the performance of AI system under development. If the result of the new algorithm plus the data model appears to be better than the baseline system, we could consider it has passed the meaningful threshold to kick off the test against a set of comprehensive metrics.

4) Technical talents - Be wary of being confused by a AI demo prototype that was developed by using other technology. The project team that has developed the AI prototype is probably truly talented but in reality the prospect for them to deliver a revolutionary AI product is in doubt. This is important as the team are asking for expensive budget based on AI tagline and therefore their skillet in AI domain needs to be scrutinized.