为什么机器学习如此神奇

I have been working with Machine Learning for a couple of years and now as an instructor at School of Data Science. At first it felt like magic: you have some data (in my case computer game data), insert it in a model (in my case a Neural Network) and you get a good result! Just like with “real” magic, the trick does get less magical when you know what is going on. However, just like with “real” magic, this doesn’t make it less amazing!

我已经在机器学习领域工作了两年，现在担任数据科学学院的讲师。 刚开始感觉就像魔术：您有一些数据(在我的情况下是计算机游戏数据)，将其插入模型(在我的情况下是神经网络)，您将获得很好的结果！ 就像使用“真正的”魔术一样，当您知道发生了什么时，该技巧的魔术性也会降低。 但是，就像使用“真正的”魔术一样，这并没有使它变得那么神奇！

机器学习的魔力 (The magic of Machine Learning)

It’s extremely powerful to draw the right conclusion given a certain situation. This could be: selling or buying stocks at the right time, determining if a patient has a certain disease or correctly guessing the content of a piece of text.

在特定情况下得出正确的结论非常强大。 这可能是：在正确的时间买卖股票，确定患者是否患有某种疾病或正确猜测一段文字的内容。

To draw the correct conclusions you have to extract the relevant patterns out of the data. You could manually obtain these features by handcrafting them from the data. This is quite a decent approach when you understand the data extremely well; it might outperform a Machine Learning approach. However, this is not practical when dealing with tougher sets of data like text in natural language processing or images in image processing.

为了得出正确的结论，您必须从数据中提取相关模式。 您可以通过从数据中手工制作这些功能来手动获得这些功能。 当您非常了解数据时，这是一种不错的方法。 它可能胜过机器学习方法。 但是，当处理更复杂的数据集(例如自然语言处理中的文本或图像处理中的图像)时，这是不切实际的。

The magic of Machine Learning is its ability to find these patterns in the data by itself. These patterns are then used to draw a conclusion. Most of the time models are trained to give the correct conclusion right away. The job you have as the creator of artificial intelligence is to choose/create a model that is able to find these patterns.

机器学习的魔力在于它能够自行在数据中找到这些模式。 然后使用这些模式得出结论。 大多数时间模型都经过训练，可以立即给出正确的结论。 作为人工智能的创造者，您要做的工作是选择/创建能够找到这些模式的模型。

Asimovinstitute Asimovinstitute上的神经网络动物园

当机器学习开始失去魔力时(When Machine Learning starts losing its magic)

Just like with “real” magic this feeling of awe is replaced with disappointment when you try this Machine Learning magic yourself and the trick doesn’t work… Now you have to figure out why copy pasting code from the internet and applying it on your data gives such a bad result.

就像使用“真正的”魔术一样，当您自己尝试使用机器学习魔术时，这种敬畏感被失望所取代，并且该技巧不起作用……现在您必须弄清楚为什么要从Internet复制粘贴代码并将其应用到您的数据中给出了如此糟糕的结果。

Image by anncapictures on Pixabay

该图片由 anncapictures在 Pixabay上 发布

The data might simply not contain any relevant patterns for your conclusion. For example, try to predict someone’s personality or what someone has to do that week solely based on someone’s date of birth, the position of the planets and rotation of the earth! That would be pure insanity, right?

数据可能根本不包含任何相关的结论模式。 例如，尝试仅根据某人的出生日期，行星的位置和地球的旋转来预测某人的性格或某周该做什么？ 那纯粹是精神错乱吧？

Maybe your model is not able to get the patterns out of the data. For example, trying to predict whether a picture is displaying a cat or a dog would not work very well if your model is trying to fit some curve through the pixel values (like logistic regression). In that case you should look at a bit more sophisticated models.

也许您的模型无法从数据中提取模式。 例如，如果您的模型试图通过像素值拟合某些曲线(例如逻辑回归)，则试图预测图片是显示猫还是狗的效果就不太好。 在这种情况下，您应该考虑一些更复杂的模型。

Another seemingly weird problem that sometimes occurs, is that the data is too rich. For example, inserting text symbol for symbol in your model will make it lose track of previous words very fast. A better alternative would be to insert whole words or parts of words.

有时会出现的另一个看似奇怪的问题是数据太丰富。 例如，在模型中为符号插入文本符号将使其很快失去对先前单词的跟踪。 更好的选择是插入整个单词或单词的一部分。

When trying to fix the problem you start to learn more and more about your data and your model. You start to see that:

在尝试解决问题时，您开始越来越了解有关数据和模型的知识。 您开始看到：

• your large Neural Network is actually an absurd way of drawing a straight line through some points;
  您的大型神经网络实际上是通过某些点绘制直线的荒谬方式；
• your natural language model is performing at about the same level as just simply counting the number of words in a text with an additional simple calculation;
  您的自然语言模型的表现与仅通过简单的计算即可简单地计算文本中的单词数大致相同。
• your image recognition model actually doesn’t know the concepts that well by giving it a few slightly off images.
  您的图像识别模型实际上不了解这些概念，因此给它一些稍微偏离图像的模型。

那么，机器学习的魔力是什么？ (So what is the magic of Machine Learning?)

I find Machine Learning fascinating. It’s a continuous roller coaster of feeling in awe by the magical results of a model followed by a slight contempt when you find the model’s boundaries. However, everytime the magic disappears it’s because you learned something new.

我发现机器学习很有趣。 令人惊奇的是，模型的神奇结果使人惊叹不已，而当您发现模型的边界时，便会轻视它。 但是，每一次魔术消失，都是因为您学到了新东西。

My background is in physics with also some cell biology on the side. And I’m very used to the feeling of describing something that feels magical by maths. Machine Learning gives me the feeling that one day we can also understand the human brain that way. I think that this will probably be the biggest magic trick of Machine Learning…

我的背景是物理学，同时也有一些细胞生物学。 而且我已经习惯了用数学来描述一些神奇的东西。 机器学习给我的感觉是，有一天我们也可以那样理解人的大脑。 我认为这可能是机器学习的最大魔咒...

And o yeah, it actually works!

是的，它实际上有效！

我更多 (More of me)

My website

我的网站

My own projects on youtube:

我在youtube上的项目：

Our Machine Learning and data science courses:

我们的机器学习和数据科学课程：

Meetups about Machine Learning in The Netherlands:

在荷兰举行的有关机器学习的聚会：

翻译自: https://medium.com/the-innovation/why-is-machine-learning-so-magical-332a1f32f9b