ai策略机器人研究a50_跟上AI研究的策略

最新推荐文章于 2024-10-02 00:23:27 发布

weixin_26632369

最新推荐文章于 2024-10-02 00:23:27 发布

阅读量383

点赞数

文章标签：人工智能 java 算法机器学习

原文链接：https://towardsdatascience.com/strategies-for-keeping-up-with-ai-research-7861037c7247

版权

ai策略机器人研究a50

人工智能的速度 (The Speed of AI)

For anyone that is following the fields of Artificial Intelligence (AI), Deep Learning (DL) or Machine Learning (ML) it can seem that research is speeding past you like a race car sometimes. A quick search on arXiv.org shows that 2,683 new articles related to AI, DL, or ML were announced between April 1st and May 1st of this year (2020). The research community is constantly being flooded with new work (both good and bad) and keeping up can seem like an impossibility. This is only further exacerbated by the coverage that AI is receiving in the media. Whether it is the imminent arrival of self-driving cars, advancements in natural language processing, automated medical diagnosis, or the fear of the cybernetic robot uprising, AI has been getting plenty of attention.

对于那些关注人工智能(AI)，深度学习(DL)或机器学习(ML)领域的人来说，研究似乎有时像赛车一样飞速驶向您。在arXiv.org上的快速搜索显示，在今年(2020年)的4月1日至5月1日之间宣布了2683篇与AI，DL或ML相关的新文章。研究社区不断充斥着新的工作(无论好坏)，跟上工作似乎是不可能的。 AI在媒体中获得的报道只会进一步加剧这一点。无论是自动驾驶汽车即将到来，自然语言处理的进步，自动医疗诊断，还是担心控制论机器人起义，AI都引起了广泛关注。

There are, of course, many benefits to the myriad of exposure that AI is receiving; however, for those working in the field, there can also be some negative effects. First, the fire hose of information from new applications, better architectures, emerging sub-fields, etc…, can become a distraction. There is a countless number of rabbit holes to jump down at any moment. While it is important to explore new and emerging solutions that may provide better solutions, if we spend all of our time chasing shiny objects, then we will never make any progress of our own.

当然，人工智能正在获得无数的曝光有很多好处。但是，对于那些在野外工作的人来说，也会产生一些负面影响。首先，来自新应用，更好的体系结构，新兴的子领域等的信息的消防水带可能会分散注意力。随时都有无数的兔子洞跳下来。尽管探索可能提供更好解决方案的新兴解决方案很重要，但是如果我们将所有时间都花在寻找有光泽的物体上，那么我们将永远不会取得任何进展。

Second, the barrage of new research and the growing AI industry can create unwanted anxiety. The pressure to succeed and succeed quickly can shake the confidence of even the most experienced researchers and developers. It can seem like other researchers are publishing twice as fast or companies are developing twice as many applications. This can lead to poor design decisions and wasted time, which undermines our ability to apply AI effectively. Furthermore, this can make us susceptible to group thinking and limit our ability to approach problems with novel and creative solutions.

其次，大量新研究和不断发展的AI行业可能会造成不必要的焦虑。成功和快速成功的压力甚至可以动摇最有经验的研究人员和开发人员的信心。看起来其他研究人员的发布速度是其他研究人员的两倍，或者公司开发的应用程序数量是两倍。这可能会导致糟糕的设计决策和浪费的时间，从而破坏我们有效应用AI的能力。此外，这可能使我们容易受到群体思维的影响，并限制了我们以新颖和创造性的解决方案来解决问题的能力。

Ultimately, it is necessary to keep up with the research related to your application area or industry, if you want to stay relevant. The trick is finding the ways that don’t consume all of your time and work best for you. The rest of this article identifies strategies to do just that. It is also important that we allow the state-of-the-art research to inspire us rather than just influence us, but that will be a topic for a future article.

最终，如果您想保持相关性，就必须跟上与您的应用领域或行业相关的研究。诀窍是找到不会浪费您所有时间并最适合您的方式。本文的其余部分确定了做到这一点的策略。同样重要的是，我们应允许最新的研究启发我们而不是仅仅影响我们，但这将是以后文章的主题。

保持策略 (Strategies for Keeping Up)

So the good news is there is a plethora of resources available to help us keep up with the research that is relevant to our interests. In this article, I share a number of the resources that I use and discuss some of my strategies for staying current with AI research. The items below are ordered based on my personal priority but everyone learns differently, so I encourage you to try them out and see what works well for you.

因此，好消息是，有大量资源可用来帮助我们跟上与我们的兴趣相关的研究。在本文中，我分享了许多我使用的资源，并讨论了一些与AI研究保持同步的策略。下面的项目是根据我的个人优先顺序排列的，但是每个人的学习方法都不同，因此，我鼓励您尝试一下，看看哪些对您有效。

0.现实和一致 (0. Be Realistic and Consistent)

We all have a limited amount of time to spend on keeping up with AI, so be realistic and consistent. Dedicate a fixed amount of time that fits within your schedule. Recognizing your time limitations should help you to choose the appropriate strategies that will be the most effective for you. Work smart and hard.

我们所有人都有有限的时间花在跟上AI上，所以要现实和一致。安排一个适合您计划的固定时间。认识到您的时间限制应该可以帮助您选择最适合自己的策略。聪明而努力地工作。

1.阅读论文 (1. Read Papers)

So as obvious as it may sound it, it is very important to spend some of your time reading papers. This will provide you with the obvious benefit of learning about new and emerging work, but it will also strengthen your ability to distill and understand complex concepts. Reading technical papers is a skill that is crucial for engineering and science, and like all other skills, practice makes perfect. Bonus: critically reading papers is also a great way to come up with new research/algorithm ideas by really challenging everything the authors are presenting.

听起来很明显，所以花一些时间阅读论文非常重要。这将为您提供学习新兴工作的明显好处，但也将增强您提炼和理解复杂概念的能力。阅读技术论文是一项对工程和科学至关重要的技能，与所有其他技能一样，实践也很完美。奖励：认真阅读论文也是通过挑战作者提出的一切来提出新研究/算法思想的好方法。

Here are some of my tips for reading papers.

这是我阅读论文的一些技巧。

1A. Stay organized.

1A。 保持井井有条。

So the first tip I have is to keep your papers organized. In my case, I have “_ReadMe_” folder on my cloud drive that has all of the papers that I would like to read. So, when you have a little bit of time you can go straight to your folder and start reading. Once you have read the paper you can either move it to somewhere else in your digital library or simply delete it.

因此，我的第一个技巧是使您的论文井井有条。就我而言，我的云驱动器上有“ _ReadMe_”文件夹，其中包含我想阅读的所有论文。因此，如果您有一点时间，可以直接进入文件夹并开始阅读。阅读论文后，您可以将其移至数字图书馆中的其他地方，也可以将其删除。

1B. Read smart.

1B。 精读。

As I learned quickly during my Ph.D, reading papers can take a lot of time and if you don’t have a good strategy you will never get through all of the papers that you need to. Personally, I have found that a three-phase approach works very well for me. Each phase has a specific purpose and the subsequent stages build on the previous ones. This can greatly speed up the process, because you will often find that getting through phase 2 is sufficient in many cases.

正如我在攻读博士学位期间很快学到的那样，阅读论文可能会花费很多时间，而且如果您没有好的策略，您将永远无法读完所有需要的论文。我个人发现，三阶段方法对我来说非常有效。每个阶段都有特定的目的，后续阶段则基于先前的目的。这可以大大加快该过程，因为您通常会发现在很多情况下，通过阶段2就足够了。

This is my general approach.

这是我的一般方法。

Phase 1: Read the abstract and the conclusion.
阶段1：阅读摘要和结论。
Phase 2: Read through the entire paper, but skip any technical details that require lots of mental effort.
阶段2：通读整篇论文，但跳过所有需要大量精力的技术细节。
Phase 3: Read critically and challenge the authors assumptions and assertions.
阶段3：认真阅读并挑战作者的假设和主张。
Phase 4 (Bonus!): Code it up! This is obviously more than just reading, but ultimately the best way to truly understand new concepts.
阶段4(奖金！)：编写代码！这显然不仅仅是阅读，而是最终真正理解新概念的最佳方法。

I will be publishing a more detailed article on my approach in the near future. I would also encourage you to read several other strategies and find what works for you.

我将在不久的将来发布有关我的方法的更详细的文章。我也鼓励您阅读其他几种策略，并找到适合您的方法。

1C. Set up alerts.

1C。 设置警报。

Google Scholar is an amazing resource for many reasons, but the automated alerts are a great way to keep up with what’s new. Alerts can be created based on an author or a search string. Author alerts are great for following the heavy hitters in the field and search string alerts are good for finding new authors.

出于多种原因， Google学术搜索是一个了不起的资源，但是自动警报是跟上最新消息的好方法。可以基于作者或搜索字符串创建警报。作者警报非常适合跟踪该领域的热门人物，而搜索字符串警报则有助于查找新作者。

To create an author alert:

要创建作者警报：

Search for an author and click on their name to get to their Google Scholar profile.
搜索作者，然后单击其姓名以进入其Google学术搜索个人资料。
Select the “Follow” button in the top right corner of the profile.
选择配置文件右上角的“关注”按钮。
Enter your e-mail address and select the alert options.
输入您的电子邮件地址，然后选择警报选项。

To create a search string alert:

要创建搜索字符串警报：

Log into Google Scholar and select “Alerts” from the side menu.
登录Google学术搜索，然后从侧面菜单中选择“警告”。
Click “Create Alert”.
点击“创建警报”。
Enter your e-mail address, the search string that should be queried and the total number of results that should be included.
输入您的电子邮件地址，应查询的搜索字符串和应包括的结果总数。

Pro Tip: If you use GMail then you can take advantage of these fancy filtering tips to automagically organize the alerts in your in-box. You can read about the fancy filtering here.

专业提示：如果您使用GMail，则可以利用这些精美的过滤提示来自动组织收件箱中的警报。您可以在此处阅读有关花式过滤的信息。

2.观看YouTube (2. Watch YouTube)

YouTube is one of the most exciting places that I am starting to see really good content. There are a lot of really smart people that are publishing paper reviews, online lectures and tutorials. All of these are excellent sources for quickly coming up to speed on new topics and gaining intuition. I found that these videos have saved me time and made me aware of new topics that I wouldn’t have otherwise seen.

YouTube是我开始看到非常好的内容的最令人兴奋的地方之一。有很多非常聪明的人正在发表论文评论，在线讲座和教程。所有这些都是极好的资源，可帮助您快速掌握新主题并获得直觉。我发现这些视频为我节省了时间，并使我意识到了原本不会看到的新主题。

Rather than providing a long list of potential channels, I have included just three YouTube channels that I would recommend starting with. I have included a little blurb about each channel and I why I think it is worthwhile. After checking out these channels, I would encourage you to check out some more channels to find the ones that suit your needs.

我没有提供一长串潜在频道，而是只推荐了三个YouTube频道。我对每个频道都做了一些介绍，为什么我觉得这值得。在检查了这些渠道之后，我建议您进一步检查一些渠道以找到适合您需求的渠道。

Yannic Kilcher: Yannic posts really good paper reviews. He walks through newly published papers and provides most of the required background material necessary to get a good grasp on the papers contribution. His videos are straight-and-to-the-point which allows him to post a lot of videos, most of which are less than 30 minutes. Nowadays, if there is a new paper that I am interested in, I will do a quick search to see if Yannic has covered it yet.
Yannic Kilcher ：Yannic发表了非常好的论文评论。他浏览了新发表的论文，并提供了必要的大多数必要的背景材料，以使您能够更好地掌握论文的贡献。他的视频内容直截了当，这使他可以发布很多视频，其中大多数视频不到30分钟。如今，如果有我感兴趣的新论文，我将进行快速搜索以查看Yannic是否已覆盖它。
Henry AI Labs: This channel is run by Connor Shorten and also provides a lot of paper reviews and explanation videos. Like Yannic, Connor is very knowledgeable and does an excellent job explaining AI concepts. The real golden nugget of this channel (for me at least) is the “AI Weekly Update” series. Each week Connor provides a high-level review of the latest news in AI research The update videos are less than 30 minutes and include a ton of content.
亨利AI实验室：该频道由康纳·肖顿(Connor Shorten)运营，还提供了大量纸质评论和说明视频。像Yannic一样，Connor知识渊博，在解释AI概念方面做得很出色。这个频道(至少对我而言)的真正金块是“ AI每周更新”系列。康纳每周都会对AI研究中的最新新闻进行高层审查。更新视频不到30分钟，并且包含大量内容。
Two-Minute-Papers: This channel is run by Karoly Zsolnai-Feher and has already amassed over 600k subscribers. Karoly produces short (~5 minutes) videos that are very well produced. The videos are typically very high level, but very clearly illustrate the topics. These videos are easy to watch and can be addictive (in a good way).
两分钟文件：该频道由Karoly Zsolnai-Feher运营，已经积累了60万订户。 Karoly制作了很短的视频(〜5分钟)，效果非常好。这些视频通常水平很高，但是非常清楚地说明了主题。这些视频易于观看，并且会令人上瘾(以一种很好的方式)。

Pro Tip: YouTube settings allow you play videos back at up-to 2x speed, which is extremely helpful for getting through AI content. I have found that 2x speed is tolerable, once you get used to it. For even finer control, you can also use plugins like Enhancer for YouTube.

专家提示：YouTube设置可让您以高达2倍的速度播放视频，这对于获取AI内容非常有帮助。我发现，一旦习惯了2倍的速度，它是可以容忍的。为了获得更好的控制效果，您还可以使用YouTube增强器之类的插件。

3.订阅新闻快报和订阅 (3. Subscribe to News Letters and Subreddits)

Another great resource is to use news letters and Subreddits to help filter out content. Admittedly, I don’t subscribe to many news letters, but I find The Batch by deeplearning.ai to be very informative. Andrew Ng is one of the biggest names in AI and for good reasons. Each week he provides a curated list of hot topics in AI, along with the relevant context and why this research might be important. I also personally, think that Andrew covers topics that are interesting for both researchers and application developers. I would highly recommend subscribing to this news letter, as well as find other news letters that match your interests.

另一个很棒的资源是使用新闻快报和Subreddits帮助过滤内容。诚然，我没有订阅太多新闻快报，但是我通过deeplearning.ai发现The Batch非常有用。吴安德(Andrew Ng)是AI中最著名的人物之一，这有充分的理由。每周他都会提供AI热门话题的精选列表，以及相关的背景以及这项研究为何重要的原因。我个人也认为Andrew涵盖了研究人员和应用程序开发人员都感兴趣的主题。我强烈建议您订阅此新闻通讯，以及查找其他符合您兴趣的新闻通讯。

I have been a big fan of Reddit for many years. I have always viewed Reddit as a community based filter for the internet. Reddit hosts a number of sub-communities can Subreddits, that are dedicated to a particular topic or theme. Members can post content that can be either up or down voted based on how well it is received. The hope being that the most useful and informative content is promoted to the top of the list. There are several different Subreddits related to AI and ML, but I would recommend starting out with Deep Learning. This community has just under 50k members and the Subreddit is very active.

多年来，我一直是Reddit的忠实拥护者。我一直将Reddit视为基于社区的Internet过滤器。 Reddit托管有许多子社区Sub Subddits，专门针对特定主题或主题。成员可以根据收到的内容来发布可以投票赞成或反对票的内容。希望将最有用和最有用的内容提升到列表的顶部。有几种与AI和ML相关的Subreddit，但我建议从Deep Learning开始。这个社区的成员不到5万，Subreddit非常活跃。

As an aside, Reddit is not only a good source of information, but it is also a place to interact with like-minded AI enthusiasts. I think this can be especially useful for those that are new in the field and would like to get some feedback, or those who don’t interact regularly with AI experts. I don’t personally post in many of these Subreddits currently, but hopefully that will change in the near future.

顺便说一句，Reddit不仅是很好的信息来源，还是与志趣相投的AI爱好者进行交流的地方。我认为这对于本领域的新手，想要获得一些反馈的人，或者那些不定期与AI专家互动的人尤其有用。我目前没有亲自在其中许多Subreddit中发布信息，但希望不久的将来会有所改变。

4.回到基础 (4. Go Back to the Basics)

I found that it is useful to spend some of my time reviewing the basics. Strengthening your foundations will make you more efficient at reading papers within your field. I also try to choose these topics based on what I have been reading. For example, say that you are trying to learn about speech recognition algorithms, then it would be worthwhile to review your basic signal processing and spectral analysis. This will give you an edge to better understand and perhaps improve on what you are reading. It will also help you to challenge assumptions that you may otherwise take at face value.

我发现花一些时间来复习基础知识很有用。加强基础将使您在阅读该领域的论文时更有效率。我也尝试根据我一直在阅读的内容选择这些主题。例如，假设您正在尝试学习语音识别算法，那么值得回顾一下基本信号处理和频谱分析。这将使您有更好的理解力，甚至可以改善阅读的内容。它还将帮助您挑战您可能会以其面值承担的假设。

While it is impossible to cover everyone’s interests, I would recommend the following general categories as topics to consider.

虽然不可能涵盖每个人的利益，但我建议将以下一般类别作为要考虑的主题。

Calculus
结石
Linear Algebra
线性代数
Optimization
优化
Programming
程式设计
Digital Signal Processing
数字信号处理
Random Processes, Statistics, and Information theory
随机过程，统计和信息论
Domain Knowledge (depends on your field of interest)
领域知识(取决于您的兴趣领域)

There are a number of great free and paid resources online for learning these types of topics. I would recommend that you take a look at Coursera, edX, Standford Online and MIT OpenCourseware for quick start.

在线上有许多免费的付费资源，可用于学习这些类型的主题。我建议您看一下Coursera ， edX ， Standford Online和MIT OpenCourseware，以快速入门。

5.阅读技术博客 (5. Read Technical Blogs)

There is definitely no shortage of AI and ML blogs these days. Content can vary widely based on the site and purpose of the blog. I typically end up on most blogs through google searches when I am trying to find information on a specific topic. However, there are a few blogs that I browse regularly to look for new content. Here are the top three that I find informative, but as usual this will depend on your interests.

如今，AI和ML博客绝对不缺。内容可以根据博客的站点和目的而有很大差异。当我尝试查找有关特定主题的信息时，通常会通过Google搜索进入大多数博客。但是，我会定期浏览一些博客以寻找新内容。以下是我认为能提供信息的前三名，但通常情况下，这取决于您的兴趣。

Pro Tip: You can use a news reader tool like Inoreader or Feedly to gather all of your blog content in one place.

专家提示：您可以使用Inoreader或Feedly等新闻阅读器工具将所有博客内容收集到一个位置。

6.收听播客 (6. Listen to Podcasts)

Podcasts can be hit or miss depending on who you are. There are a number of different AI and ML related podcasts out there, many of which follow fairly similar formats. Most of them include interviews with researchers and developers, talk about emerging issues, and cover the fundamentals of the fields. I have found myself spending less time with podcasts lately, but I still think they are great resources and worth a review. Some of these podcasts can be very lengthy, so it is probably best to choose a few and to stick with those for a while.

取决于您是谁，播客可能会受到打击或错过。那里有许多与AI和ML相关的不同播客，其中许多遵循相当相似的格式。其中大多数包括对研究人员和开发人员的采访，谈论新出现的问题以及涵盖该领域的基础知识。我发现自己最近花在播客上的时间减少了，但是我仍然认为它们是很棒的资源，值得一读。这些播客可能有些冗长，因此最好选择一些播客，并坚持一会儿。

Similar to the section on YouTube, rather than providing a long list of possibilities, I have provided a few recommendations below.

与YouTube上的部分类似，我没有提供很多可能性，而是在下面提供了一些建议。

Linear Digressions: This podcast offers bite-sized doses of technical machine learning content. The episodes are short (less than 20 minutes) and cover a wide range of topics.
线性题外话：此播客提供一口大小的技术机器学习内容。剧集很短(不到20分钟)，涵盖了广泛的主题。
Talking Machines: This podcast features longer episodes (about an hour) and is less technical than Linear Digressions. The format is based on interviewing working AI practitioners and discussing relevant topics in a very practical sense.
会说话的机器：该播客的节目较长(大约一个小时)，并且比线性传讯的技术性差。该格式基于采访实际AI工作者并在非常实际的意义上讨论相关主题。
Artificial Intelligence Podcast: This podcast is solely based on interviews and the episodes range from one to three hours. Full transparency, this podcast covers a broad range of topics, many of which are non-technical and even philosophical in nature. I personally find many of the episodes interesting and the many of the guests are very well respected researchers and developers in the AI community.
人工智能播客：此播客仅基于采访，播出时间从一小时到三个小时不等。此播客内容完全透明，涵盖了广泛的主题，其中许多主题本质上不是技术性的，甚至是哲学性的。我个人发现许多情节很有趣，并且许多来宾都是AI社区中非常受人尊敬的研究人员和开发人员。

7.监控社交媒体 (7. Monitor Social Media)

Now, I will admit that this recommendation is a bit hypocritical, since I abstain almost entirely from social media. However, I recognize that it can be a great tool for tracking what’s happening with your favorite authors or generally what is happening in your field. Many researchers and developers will often Tweet about new papers or developments in industry. This can be used as another source of filtering out important information. So if social media is your thing that find ways to make it productive for you.

现在，我承认这个建议有些虚伪，因为我几乎完全不使用社交媒体。但是，我认识到它可能是跟踪您最喜欢的作者正在发生的事情或总体上您所在领域正在发生的事情的很好的工具。许多研究人员和开发人员经常会在推特上发表有关行业的新论文或发展的推文。这可以用作过滤掉重要信息的另一个来源。因此，如果社交媒体是您的理想之选，便会找到使之高效的方法。

结论 (Conclusion)

AI (and related fields) are experiencing a period unprecedented growth which fantastic for the field, but it can certainly be overwhelming. With so much happening all the time it is easy to get distracted and that can be a detriment to your work. Take some solace in knowing that you are not alone. The most important thing is to recognize what time you have and to use it as wisely as possible.

人工智能(及相关领域)正经历一段空前的增长，这对于该领域而言是惊人的，但它肯定会是压倒性的。这么多事情一直发生，很容易分心，这可能会损害您的工作。知道自己并不孤单，请安慰一下。最重要的是要认识到自己有什么时间，并尽可能明智地利用它。

This article presented a handful of strategies that I use for keeping up research. Pick a few of the strategies from the list and give them shot. Don’t try to implement too much at once or it will quickly become overwhelming. See what works and what doesn’t, as you slowly find your way of learning. Ultimately you will need to find what works best for you and that may change over time. Keep exploring, keep learning, and grow your understanding.

本文介绍了一些我用来跟进研究的策略。从列表中选择一些策略，然后进行尝试。不要尝试一次实施太多，否则它将很快变得不堪重负。慢慢找到学习方式，看看有什么用，什么没用。最终，您将需要找到最适合您的方法，并且随着时间的流逝可能会发生变化。不断探索，不断学习并增进理解。