介绍 (Introduction)

The special thing I found when I first started diving into the field of Artificial Intelligence was the infinite amount of parallels between how neural networks learn and my subjective experience of my own intelligence.

我刚开始涉入人工智能领域时发现的特别之处是，神经网络的学习方式与我对自己的智能的主观体验之间有着无数的相似之处。

So, I decided to write a series of posts about such parallels. I would summarize this first one with the following proposition:

因此，我决定写一系列有关这种并行的文章。 我将用以下命题来总结第一个：

We can learn new skills faster if we frame the process as a supervised learning task.

如果我们将流程视为监督学习任务，则可以更快地学习新技能。

监督学习 (Learning as Supervised Learning)

When I reference supervised learning here, I am talking about the idea of models learning by leveraging exposure to relevant examples using an advanced pattern recognition system.

当我在这里引用监督学习时，我在说的是通过使用高级模式识别系统利用相关示例的知识来进行模型学习的想法。

If you are involved with AI you have seen a picture like this:

如果您参与了AI，您将看到如下图片：

Yes, the inputs, the outputs, the cost functions…

是的，投入，产出，成本函数……

But, on its core, supervised learning fascinated me because it directly spoke to the way I learn all sorts of different subjects.

但是，从本质上讲 ，有监督的学习使我着迷，因为它直接影响了我学习各种不同学科的方式 。

This becomes clearer if we break down the intuitions behind the main concepts involved in a basic neural network.

如果我们打破了基本神经网络所涉及的主要概念的直觉，这一点将变得更加清楚。

Inputs: data with patterns one wishes to grasp.

输入：具有希望掌握的模式的数据。

A forward pass: transformation of this input

前向传递：此输入的变换

A cost function representing how far the output is from an ideal goal and a backward pass where the model can update its strategies to learn better.

成本函数代表输出与理想目标的距离，以及模型可以更新其策略以更好地学习的反向传递 。

It is easy to see how such steps relate to our human experience of learning. You dive into a new field like math, combat sports, or cooking and you don’t know much about it, but the more you are exposed to examples of these topics, the more you learn even without consciously trying to!

很容易看出这些步骤与我们人类的学习经验之间的关系。 您进入数学，格斗运动或烹饪等新领域，对此一无所知，但是您接触这些主题的例子越多，即使没有自觉地尝试也学到更多 ！

通过曝光学习 (Learning through exposure)

Take kids for example,

以孩子们为例，

in this very enlightening talk, psychologist Chris Lonsdale points out that babies learn to speak partially because they are constantly exposed to adults who are experts at that particular language and can provide them with helpful feedback, so they pick up on the language just by basic trial and error.

在这个非常有启发性的演讲中 ，心理学家克里斯·朗斯代尔(Chris Lonsdale)指出，婴儿之所以学习部分语言，是因为他们不断接触成年人，他们是该特定语言的专家 ，可以为他们提供有用的反馈，因此，他们只是通过基本的试验就可以选择该语言。和错误。

Now, I don’t want to get into what kind of complicated loss function these kid’s brains are trying to optimize,

现在，我不想了解这些孩子的大脑正在尝试优化的复杂损失函数，

because that is beside the point (although a very interesting one).

因为那是重点( 尽管很有趣 )。

The point is: they manage to learn without having to appeal to some complicated system or rule.

关键是：他们设法学习而不必诉诸某些复杂的系统或规则。

They learn it through their exposure.

他们通过接触来学习。

Consider how many adults who start to learn a language try to do it by attempting to learn the rules for making grammatic expressions rather than just exposing themselves to the target language and allowing that exposure combined with an acute awareness guide their learning process.

考虑多少个开始学习语言的成年人尝试通过尝试学习语法表达的规则来尝试做到这一点，而不仅仅是将自己暴露于目标语言并允许其暴露与敏锐的意识共同指导他们的学习过程。

At this point you might be thinking: “ok, we learn from examples, so what? For complicated stuff we still need to sit down and read countless books, process an infinite amount of information, get bored and procrastinate.”.

此时，您可能会想：“好吧，我们从示例中学到了，那又如何呢？ 对于复杂的东西，我们仍然需要坐下来阅读无数的书，处理无数的信息，感到无聊和拖延。”

That might be true but there is an interesting way to tinker with our brain’s innate capacity to make associations, which means treating ourselves, to some limited degree, like a learning machine, meaning (for the purpose of this article) to automate exposure just as we would do with a machine learning task.

可能是正确的，但是有一种有趣的方式可以改善我们大脑的天生建立联系的能力，这意味着在某种程度上像对待学习机那样对待自己，这意味着(对于本文而言)意味着可以自动进行曝光我们将完成机器学习任务。

通过程序化学习来学习法语 (Learning French through programmatic exposure)

To clarify, I’ll give you an example. When I was in Brazil, one of my big dreams was to learn French. I, like many Brazilians, did not have the opportunity to go to France and hang out there to learn it so, for a while, I was quite upset.

为了澄清，我举一个例子。 当我在巴西时，我最大的梦想之一就是学习法语。 我像许多巴西人一样，没有机会去法国并在那里闲逛学习，所以有一段时间，我很沮丧。

However, I came upon this idea of systematic immersion in the environment of the target language, so I wondered: if I programmatically expose myself to French will I learn enough to acquire the language?

但是，我想到了一种系统地沉浸在目标语言环境中的想法，所以我想知道：如果我以编程方式使自己接触法语，我是否会学到足够的知识来学习该语言？

Upon some reflection and realizing I did not have enough time to just study French all day, I decided to do something a bit unconventional: I gave a “pythonesque” twist to things and decided to automate part of this process.

经过一番思考和意识到，我没有足够的时间整天学习法语，所以我决定做一些与众不同的事情：我对事情采取了“怪异”的态度，并决定使该过程的一部分自动化。

自动曝光的步骤 (Steps to automate exposure)

As I quickly mentioned before, in most deep learning projects you usually deal with a pipeline composed of the following: a dataset (with the input data the model needs to learn the task), a forward pass where the data goes through the model and it is transformed to an output representing the model prediction, a cost function that tells you how wrong the output of your model is, a backward pass to update the units allowing the model to learn and some type of evaluation process enabling you to know how good the model is when applied to data it has never seen before.

正如我之前很快提到的，在大多数深度学习项目中，您通常会处理由以下内容组成的管道： 数据集 (使用输入数据，模型需要学习任务)，数据通过模型的正向传递 ，转换为代表模型预测的输出，指示您模型输出错误的成本函数 ， 向后传递以更新允许模型学习的单位以及某种类型的评估过程，从而使您知道模型的质量模型是应用于从未有过的数据时。

Given these structural components, what I did was to use python to build a simple system that resembled this type of pipeline so that I could automate my exposure to relevant content.

有了这些结构组件，我要做的就是使用python构建类似于这种类型的管道的简单系统， 以便我可以自动展示相关内容。

Let’s go through each step one by one.

让我们一步一步地完成每个步骤。

Dataset

数据集

I built a learning curriculum made out of links to videos, articles, textbooks with all the content from the web I thought could help me understand the language (I suggest around 20 to 40 links covering video, audio, and text).

我建立了一个学习课程，该课程由指向视频，文章，教科书的链接组成，并包含我认为可以帮助我理解该语言的所有内容(我建议大约20至40个包含视频，音频和文本的链接)。

https://www.youtube.com/watch?v=6TUNC31t73w https://www.youtube.com/watch?v=vXVg-XjmY-w https://www.youtube.com/watch?v=ujDtm0hZyII https://www.youtube.com/watch?v=X8gno6Uzuo8 https://tel.archives-ouvertes.fr/file/index/docid/825854/filename/1996_Framling_Kary.pdf https://www.persee.fr/doc/intel_0769-4113_1987_num_2_1_1804 https://lejournal.cnrs.fr/dossiers/comment-lintelligence-artificielle-va-changer-nos-vies ....

I wrote a script to transfer the links to a .csv file and create the following columns:

我编写了一个脚本，将链接转移到.csv文件并创建以下列：

• links with all the sources to learn

  与所有资源links学习

• attention_level how much attention you are putting in a given source (from just background noise to focused) as a number from 0 to 10

  attention_level您将给定来源中多少注意力(从背景噪声到聚焦)从0到10的数字

• session_time to record how long you spend on each link

  session_time记录您在每个链接上花费的时间

• date to record the date of the session

  date记录会议日期

• session_score to record a subjective score of your understanding of that particular source

  session_score记录您对特定来源的理解的主观评分

• last_index for when you stop learning in the middle of the dataset and do not cover all the sources. This way, in your next training session you can pick up from where you left off

  last_index用于您何时停止在数据集中学习并且不涵盖所有资源。 这样，在下一个培训课程中，您可以从上次停学的地方接起

Forward pass

前传

I set up a script to go systematically through the entire list of sources in the .csv file, using python’s webbrowser framework to access the URLs.

我设置了一个脚本，以使用python的webbrowser框架访问URL来系统地webbrowser .csv文件中的整个源列表。

• At each step, the user is prompted to give a score to its attention_level to that content which is subjective but it gives you the opportunity to grade your attention as you move through the content

  在每个步骤中，系统都会提示用户为该内容的attention_level等级打分，这是主观的，但是当您浏览内容时，它为您提供了对attention_level进行评分的机会

• At the end of each link the user is prompted to give a score to its “performance” (session_score)

  在每个链接的末尾，提示用户为其“表现”( session_score )评分。

• One epoch would be going through the entire material on a batch of the dataset. Meaning, going through one or a few links per epoch depending on where you stop

  一个时期将遍历一批数据集上的整个材料。 意思是，根据每个停站的位置，每个时期都要经过一个或几个链接

Cost Function

成本函数

I see a cost function as a proxy for a learning goal to which you compare the output of your model at each step of the way to know if you are doing well.

我认为成本函数可以作为学习目标的代理，您可以在模型的每个步骤中与模型输出进行比较，以了解自己是否做得很好。

In this case, the cost would be a metric of performance at a language-task which varied depending on the type of skill at hand: reading, listening, writing, speaking.

在这种情况下，成本将是语言任务的绩效指标，该任务根据手头的技能类型而有所不同：阅读，听力，写作和口语。

I measured each individually.

我分别进行了测量。

• Listening: I measured comprehension by listening to videos and assessing how much I understood by rewatching it with subtitles.

  听力 ：我通过听视频来衡量理解力，并通过重新观看字幕来评估我的理解程度。

• Speaking: I measured only pronunciation (this is the easiest thing to measure when you are starting out) so I would repeat certain chosen phrases from the material until I felt like I sounded reasonably well.

  口语：我只测量发音(这是刚开始时最容易测量的东西)，所以我会重复从教材中选择的某些短语，直到感觉听起来还不错。

• Writing: I practiced on Duolingo for a few minutes because it is easy and it gives you quick feedback. In addition, I would finish the day with a little writing sample written in French summarizing what I had learned and then compare it with what I meant using google translate.

  写作：我在Duolingo上练习了几分钟，因为它很容易并且可以给您快速的反馈。 此外，我将用法语写的一个书面示例来结束这一天，总结我所学的内容，然后将其与使用Google翻译的含义进行比较。

• Reading: I would read an article and again double-check the meaning.

  阅读 ：我会阅读一篇文章，然后再次仔细检查其含义。

I graded all these skills and stored them in thesession_scorevariable that I collected with the script.

我对所有这些技能进行了评分，并将它们存储在脚本收集的session_score变量中。

Backward pass (or updating your strategies online)

向后传递(或在线更新策略)

Although we lack access to the gradients of our neurons, we do have access to our thoughts and mental awareness as we learn, which we can use to assess what could be updated to increase learning.

尽管我们无法获得神经元的梯度 ，但是我们在学习过程中仍然可以获取思想和意识，可以用来评估可以进行哪些更新以增加学习。

Below are examples of my mental process for updating my learning strategies:

以下是我更新学习策略的心理过程的示例：

• For the low-attention sessions, I would just leave the audio in the background and repeat the words I could understand out loud to practice pronunciation as I mentioned before.对于注意力不集中的课程，我只是将音频留在背景中，然后重复我大声理解的单词来练习发音，就像我之前提到的那样。
• If I was sensing that the low attention sessions were working poorly I would do more focused ones (with a higher attention level).如果我感觉到注意力不集中的会议效果不佳，我会做重点更突出的会议(注意力水平更高)。
• If the high attention ones required too much motivation, I would switch back.如果高度关注者需要太多动力，我会转回去。

• This procedure did not guarantee an “increase in accuracy” but it guaranteed constant awareness of my mental state as I learned the content. This became a lateral skill that improved my learning process.

  此程序不能保证“准确性”的提高，但可以保证我在学习内容时不断意识到自己的心理状态 。 这成为一项横向技能，改善了我的学习过程。

The main idea here was to be mindful of what was working and what was not, so in this case the idea of a backward pass just serves as a euphemism for updating your strategies according to your online perception of your learning performance.

这里的主要思想是要记住什么在起作用 ，什么没在起作用 ，因此在这种情况下， 向后通过的想法只是根据在线对学习成绩的看法来更新策略的委婉说法。

Evaluation

评价

Evaluation, in this case, is tricky, because I was not paying full attention to everything in the same way. Therefore, any feedback mechanism here would have been noisy so I used a few different approaches that did not involve any automation per see but, again, only a guided self-awareness:

在这种情况下，评估很棘手，因为我没有以相同的方式全神贯注于所有事情。 因此，这里的任何反馈机制都将是嘈杂的，因此我使用了几种不同的方法，这些方法在每次观看时都没有涉及任何自动化，但是，再次，它只是一种引导式的自我意识：

• Language parent. Another great technique from Mr. Lonsdale, a language parent would be someone that helps you learn without negative feedback. For me this was a friend who was a native French speaker. In the absence of someone qualified, you can use foreign meetups to be around people that can give you feedback on how well you are doing based on how much you can communicate with them (expose yourself!).

  语言父母 。 朗斯代尔先生的另一项很棒的技巧是语言父母，它可以帮助您在没有负面反馈的情况下学习。 对我来说，这是一个以法语为母语的朋友。 在没有合格人员的情况下，您可以使用外国聚会在周围的人周围，这些聚会可以根据您与他们进行交流的程度(向您展示自己！)向您反馈您的表现。

• Watching novel content with French subtitles. This was an easier and effective evaluation technique. I would watch something new with subtitles in French and I would do just a personal evaluation of how much I felt I could understand by rewatching it with the original subtitles.

  观看带有法文字幕的新颖内容。 这是一种更简单有效的评估技术。 我会用法语字幕看一些新的东西，而我只是通过个人对原始字幕的重新观看，对自己的理解程度进行个人评估。

回顾周期 (Reviewing the Cycle)

At this point the idea was to use the same dataset for two or three weeks, doing one full run of the entire dataset every two or three days. In this way, at the end of the cycle you feel that you have captured the maximum that those sources could provide.

此时的想法是使用相同的数据集两到三周，每两三天对整个数据集进行一次完整运行。 这样，在周期结束时，您会感觉到已经获得了这些资源可以提供的最大数量。

After going through the entire dataset of sources at least 20 times, I re-evaluated if it was worth to add more content or to remove it.

在遍历整个源数据集至少20次之后，我重新评估了是否值得添加或删除更多内容。

个人成绩 (Personal results)

I was quite pleased that after around one year of implementing this method I learned to speak enough French to communicate with any French-speaking person without relying on English.

实施此方法大约一年后，我学会了足够的法语，可以与任何说法语的人进行交流，而无需依靠英语，对此我感到非常高兴。

The fascinating thing that I learned with this experience was how much I could absorb just by automating the learning process to remove the mental energy involved with the entry barrier to start a study session (in this case just running a single line of code in the terminal):

我从这次经历中学到的令人着迷的事情是，我可以通过自动化学习过程来消除与进入障碍有关的精神能量，从而开始学习课程 ，从而可以吸收多少(在这种情况下，只需在终端中运行一行代码即可) )：

(env_1) C:\\path\\to\\folder\\> py -i learn_this.py

结论 (Conclusion)

The most important point is that my main source of learning was the exposure.

最重要的一点是，我的主要学习来源是接触知识 。

The amazing advantage for humans in comparison to neural networks is that we can create our own datasets for our own learning goals and we can control what goes inside our minds and how.

与神经网络相比，人类的惊人优势是我们可以为自己的学习目标创建自己的数据集，并且可以控制脑海中的内心世界和方式。

By combining programmatic exposure to acute awareness, I was able to learn a complex language without the hurdles of a classical chapter-by-chapter learning procedure.

通过将程序性接触与急性意识相结合，我能够学习复杂的语言，而没有经典的逐章学习程序的障碍。

When you are exposed to anything in your life, you are, even in some minimal sense, learning something, be it the patterns that make up the events that happen before your eyes or the chains of potential causes that might be responsible for each of them.

当您接触到生活中的任何事物时，即使是从最小的意义上来说，您也将学到一些东西，无论是构成眼前发生的事件的模式还是可能对每个事件负责的潜在原因链。

When we try to learn by going through the full rule structure that makes up that process, like learning mathematics by memorizing all its theorems, we are implicitly telling ourselves that understanding that subject involves the exhaustive mapping of all its rules, which is something that carries within itself a tiresome psychological load that can wear out even the most prolific learner.

当我们尝试通过构成该过程的完整规则结构进行学习时，例如通过记忆其所有定理来学习数学时，我们在暗中告诉自己，对主题的理解涉及其所有规则的详尽映射，这是在进行的在自己内部，即使是最有能力的学习者，也会承受沉重的心理负担。

As humans, we have the disadvantage of being wired for constant self-assessment which inherently leads us towards paths of self-sabotage when we decide to tackle a novel subject.

作为人类，我们的缺点是经常进行自我评估 当我们决定解决一个新颖的主题时，这从本质上引导我们走上了自我破坏的道路。

However, it is this same self-assessing nature that allows us to implement powerful changes to guide us towards higher peaks in our learning ventures.

但是，正是这种自我评估的性质使我们能够进行有力的改变，以引导我们迈向学习事业的更高高峰。

Learning to deal with this mental trade-off is the key to becoming skilled at anything.

学会应对这种精神上的权衡是掌握任何技能的关键。

There is a lot we can take from this simple idea that our brains are wired for pattern recognition. The main of which is that potentially countless subjects (not all!) can be framed as a supervised learning task.

这个简单的想法可以为我们带来很多好处，我们的大脑被连接起来进行模式识别。 其主要目的是可以将潜在的无数主题(不是全部！)构筑为监督学习任务。

All we need to do is set up our training dataset (the exposure to examples) and make sure we don’t second-guess our updating mechanisms as well as our performance throughout the process (meaning we don’t think too hard about how well or how fast we learn).

我们需要做的就是建立训练数据集(示例展示)，并确保我们不会在整个过程中对更新机制和性能进行第二次猜测(这意味着我们不会对效果如何进行过分思考)或我们学习的速度有多快 )。

The source code for this project can be found here.

该项目的源代码可以在这里找到。

In my next post, I intend to go further into the many parallels between learning as a human and learning as a machine using the literature available as my guide and practical examples to clarify.

在我的下一篇文章中，我打算进一步深入研究人类学习和机器学习之间的相似之处，其中使用可用的文献作为指导，并举例说明一些实际的例子。

Leave a comment to let me know what you think!

发表评论，让我知道您的想法！

That’s all for today, thanks, see you next time!

今天就这些，谢谢，下次见！