Backlink Blindspots: The State of Robots.txt

Posted by rjonesx.

Here at Moz we have committed to making Link Explorer as similar to Google as possible, specifically in the way we crawl the web. I have discussed in previous articles some metrics we use to ascertain that performance, but today I wanted to spend a little bit of time talking about the impact of robots.txt and crawling the web.

Most of you are familiar with robots.txt as the method by which webmasters can direct Google and other bots to visit only certain pages on the site. Webmasters can be selective, allowing certain bots to visit some pages while denying other bots access to the same. This presents a problem for companies like Moz, Majestic, and Ahrefs: we try to crawl the web like Google, but certain websites deny access to our bots while allowing that access to Googlebot. So, why exactly does this matter?

Why does it matter?

As we crawl the web, if a bot encounters a robots.txt file, they're blocked from crawling specific content. We can see the links that point to the site, but we're blind regarding the content of the site itself. We can't see the outbound links from that site. This leads to an immediate deficiency in the link graph, at least in terms of being similar to Google (if Googlebot is not similarly blocked).

But that isn't the only issue. There is a cascading failure caused by bots being blocked by robots.txt in the form of crawl prioritization. As a bot crawls the web, it discovers links and has to prioritize which links to crawl next. Let's say Google finds 100 links and prioritizes the top 50 to crawl. However, a different bot finds those same 100 links, but is blocked by robots.txt from crawling 10 of the top 50 pages. Instead, they're forced to crawl around those, making them choose a different 50 pages to crawl. This different set of crawled pages will return, of course, a different set of links. In this next round of crawling, Google will not only have a different set they're allowed to crawl, the set itself will differ because they crawled different pages in the first place.

Long story short, much like the proverbial butterfly that flaps its wings eventually leading to a hurricane, small changes in robots.txt which prevent some bots and allow others ultimately leads to very different results compared to what Google actually sees.

So, how are we doing?

You know I wasn't going to leave you hanging. Let's do some research. Let's analyze the top 1,000,000 websites on the Internet according to Quantcast and determine which bots are blocked, how frequently, and what impact that might have.

Methodology

The methodology is fairly straightforward.

1. Download the Quantcast Top Million
2. Download the robots.txt if available from all top million sites
3. Parse the robots.txt to determine whether the home page and other pages are available
4. Collect link data related to blocked sites
5. Collect total pages on-site related to blocked sites.
6. Report the differences among crawlers.

Total sites blocked

The first and easiest metric to report is the number of sites which block individual crawlers (Moz, Majestic, Ahrefs) while allowing Google. Most site that block one of the major SEO crawlers block them all. They simply formulate robots.txt to allow major search engines while blocking other bot traffic. Lower is better.

Of the sites analyzed, 27,123 blocked MJ12Bot (Majestic), 32,982 blocked Ahrefs, and 25,427 blocked Moz. This means that among the major industry crawlers, Moz is the least likely to be turned away from a site that allows Googlebot. But what does this really mean?

Total RLDs blocked

As discussed previously, one big issue with disparate robots.txt entries is that it stops the flow of PageRank. If Google can see a site, they can pass link equity from referring domains through the site's outbound domains on to other sites. If a site is blocked by robots.txt, it's as though the outbound lanes of traffic on all the roads going into the site are blocked. By counting all the inbound lanes of traffic, we can get an idea of the total impact on the link graph. Lower is better.

According to our research, Majestic ran into dead ends on 17,787,118 referring domains, Ahrefs on 20,072,690 and Moz on 16,598,365. Once again, Moz's robots.txt profile was most similar to that of Google's. But referring domains isn't the only issue with which we should be concerned.

Total pages blocked

Most pages on the web only have internal links. Google isn't interested in creating a link graph — they're interested in creating a search engine. Thus, a bot designed to act like Google needs to be just as concerned about pages that only receive internal links as they are those that receive external links. Another metric we can measure is the total number of pages that are blocked by using Google's site: query to estimate the number of pages Google has access to that a different crawler does not. So, how do the competing industry crawlers perform? Lower is better.

Once again, Moz shines on this metric. It's not just that Moz is blocked by fewer sites— Moz is blocked by less important and smaller sites. Majestic misses the opportunity to crawl 675,381,982 pages, Ahrefs misses 732,871,714 and Moz misses 658,015,885. There's almost an 80 million-page difference between Ahrefs and Moz just in the top million sites on the web.

Unique sites blocked

Most of the robots.txt disallows facing Moz, Majestic, and Ahrefs are simply blanket blocks of all bots that don't represent major search engines. However, we can isolate the times when specific bots are named deliberately for exclusion while competitors remain. For example, how many times is Moz blocked while Ahrefs and Majestic are allowed? Which bot is singled out the most? Lower is better.

Ahrefs is singled out by 1201 sites, Majestic by 7152 and Moz by 904. It is understandable that Majestic has been singled out, given that they have been operating a very large link index for many years, a decade or more. It took Moz 10 years to accumulate 904 individual robots.txt blocks, and took Ahrefs 7 years to accumulate 1204. But let me give some examples of why this is important.

If you care about links from name.com, hypermart.net, or eclipse.org, you can't rely solely on Majestic.

If you care about links from popsugar.com, dict.cc, or bookcrossing.com, you can't rely solely on Moz.

If you care about links from dailymail.co.uk, patch.com, or getty.edu, you can't rely solely on Ahrefs.

And regardless of what you do or which provider you use, you can't links from yelp.com, who.int, or findarticles.com.

Conclusions

While Moz's crawler DotBot clearly enjoys the closest robots.txt profile to Google among the three major link indexes, there's still a lot of work to be done. We work very hard on crawler politeness to ensure that we're not a burden to webmasters, which allows us to crawl the web in a manner more like Google. We will continue to work more to improve our performance across the web and bring to you the best backlink index possible.

Thanks to Dejan SEO for the beautiful link graph used in the header image and Mapt for the initial image used in the diagrams.

Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don't have time to hunt down but want to read!

What Google's GDPR Compliance Efforts Mean for Your Data: Two Urgent Actions

Posted by willcritchlow

It should be quite obvious for anyone that knows me that I’m not a lawyer, and therefore that what follows is not legal advice. For anyone who doesn’t know me: I’m not a lawyer, I’m certainly not your lawyer, and what follows is definitely not legal advice.

With that out of the way, I wanted to give you some bits of information that might feed into your GDPR planning, as they come up more from the marketing side than the pure legal interpretation of your obligations and responsibilities under this new legislation. While most legal departments will be considering the direct impacts of the GDPR on their own operations, many might miss the impacts that other companies’ (namely, in this case, Google’s) compliance actions have on your data.

But I might be getting a bit ahead of myself: it’s quite possible that not all of you know what the GDPR is, and why or whether you should care. If you do know what it is, and you just want to get to my opinions, go ahead and skip down the page.

What is the GDPR?

The tweet-length version is that the GDPR (General Data Protection Regulation) is new EU legislation covering data protection and privacy for EU citizens, and it applies to all companies offering goods or services to people in the EU.

Even if you aren’t based in the EU, it applies to your company if you have customers who are, and it has teeth (fines of up to the greater of 4% of global revenue or EUR20m). It comes into force on May 25. You have probably heard about it through the myriad organizations who put you on their email list without asking and are now emailing you to “opt back in.”

In most companies, it will not fall to the marketing team to research everything that has to change and achieve compliance, though it is worth getting up to speed with at least the high-level outline and in particular its requirements around informed consent, which is:

"...any freely given, specific, informed, and unambiguous indication of the data subject's wishes by which he or she, by a statement or by a clear affirmative action, signifies agreement to the processing of personal data relating to him or her."

As always, when laws are made about new technology, there are many questions to be resolved, and indeed, jokes to be made:

Can you recommend a GDPR expert?
-yes
Can I have their email address?
-no
— Adam Cleevely (@ACleevely) May 2, 2018

But my post today isn’t about what you should do to get compliant — that’s specific to your circumstances — and a ton has been written about this already:

• I’ve found value in the content Ometria has produced on this front, e.g. 6 things e-commerce marketers should know about GDPR and their deeper GDPR guide (registration required)
• If you work in the area, this GDPR impact on social post from Buffer will get you up to speed there
• And for the paid marketers among us, this GDPR impact on AdWords article from WordStream is what you need

My intention is not to write a general guide, but rather to warn you about two specific things you should be doing with analytics (Google Analytics in particular) as a result of changes Google is making because of GDPR.

Unexpected consequences of GDPR

When you deal directly with a person in the EU, and they give you personally identifiable information (PII) about themselves, you are typically in what is called the "data controller" role. The GDPR also identifies another role, which it calls "data processor," which is any other company your company uses as a supplier and which handles that PII. When you use a product like Google Analytics on your website, Google is taking the role of data processor. While most of the restrictions of the GDPR apply to you as the controller, the processor must also comply, and it’s here that we see some potentially unintended (but possibly predictable) consequences of the legislation.

Google is unsurprisingly seeking to minimize their risk (I say it’s unsurprising because those GDPR fines could be as large as $4.4 billion based on last year’s revenue if they get it wrong). They are doing this firstly by pushing as much of the obligation onto you (the data controller) as possible, and secondly, by going further by default than the GDPR requires and being more aggressive than the regulation requires in shutting down accounts that infringe their terms (regardless of whether the infringement also infringes the GDPR).

This is entirely rational — with GA being in most cases a product offered for free, and the value coming to Google entirely in the aggregate, it makes perfect sense to limit their risks in ways that don’t degrade their value, and to just kick risky setups off the platform rather than taking on extreme financial risk for individual free accounts.

It’s not only Google, by the way. There are other suppliers doing similar things which will no doubt require similar actions, but I am focusing on Google here simply because GA is pervasive throughout the web marketing world. Some companies are even going as far as shutting down entirely for EU citizens (like unroll.me). See this Twitter thread of others.

Consequence 1: Default data retention settings for GA will delete your data

Starting on May 25, Google will be changing the default for data retention, meaning that if you don’t take action, certain data older than the cutoff will be automatically deleted.

You can read more about the details of the change on Krista Seiden’s personal blog (Krista works at Google, but this post is written in her personal capacity).

The reason I say that this isn’t strictly a GDPR thing is that it is related to changes Google is making on their end to ensure that they comply with their obligations as a data processor. It gives you tools you might need but isn’t strictly related to your GDPR compliance. There is no particular “right” answer to the question of how long you need to/should be/are allowed to keep this data stored in GA under the GDPR, but by my reading, given that it shouldn’t be PII anyway (see below) it isn’t really a GDPR question for most organizations. In particular, there is no particular reason to think that Google’s default is the correct/mandated/only setting you can choose under the GDPR.

Action: Review the promises being made by your legal team and your new privacy policy to understand the correct timeline setting for your org. In the absence of explicit promises to your users, my understanding is that you can retain any of this data you were allowed to capture in the first place unless you receive a deletion request against it. So while most orgs will have at least some changes to make to privacy policies at a minimum, most GA users can change back to retain this data indefinitely.

Consequence 2: Google is deleting GA accounts for capturing PII

It has long been against the Terms of Service to store any personally identifiable information (PII) in Google Analytics. Recently, though, it appears that Google has become far more diligent in checking for the presence of PII and robust in their handling of accounts found to contain any. Put more simply, Google will delete your account if they find PII.

It’s impossible to know for sure that this is GDPR-related, but being able if necessary to demonstrate to regulators that they are taking strict actions against anyone violating their PII-related terms is an obvious move for Google to reduce the risk they face as a Data Processor. It makes particular sense in an area where the vast majority of accounts are free accounts. Much like the previous point, and the reason I say that this is related to Google’s response to the GDPR coming into force, is that it would be perfectly possible to get your users’ permission to record their data in third-party services like GA, and fully comply with the regulations. Regardless of the permissions your users give you, Google’s GDPR-related crackdown (and heavier enforcement of the related terms that have been present for some time) means that it’s a new and greater risk than it was before.

Action: Audit your GA profile and implementation for PII risks:

• There are various ways you can search within GA itself to find data that could be personally identifying in places like page titles, URLs, custom data, etc. (see these two excellent guides)
• You can also audit your implementation by reviewing rules in tag manager and/or reviewing the code present on key pages. The most likely suspects are the places where people log in, take key actions on your site, give you additional personal information, or check out

Don’t take your EU law advice from big US tech companies

The internal effort and coordination required at Google to do their bit to comply even “just” as data processor is significant. Unfortunately, there are strong arguments that this kind of ostensibly user-friendly regulation which incurs outsize compliance burdens on smaller companies will cement the duopoly and dominance of Google and Facebook and enables them to pass the costs and burdens of compliance onto sectors that are already struggling.

Regardless of the intended or unintended consequences of the regulation, it seems clear to me that we shouldn’t be basing our own businesses’ (and our clients’) compliance on self-interested advice and actions from the tech giants. No matter how impressive their own compliance, I’ve been hugely underwhelmed by guidance content they’ve put out. See, for example, Google’s GDPR “checklist” — not exactly what I’d hope for:

So, while I’m not a lawyer, definitely not your lawyer, and this is not legal advice, if you haven’t already received any advice, I can say that you probably can’t just follow Google’s checklist to get compliant. But you should, as outlined above, take the specific actions you need to take to protect yourself and your business from their compliance activities.

Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don't have time to hunt down but want to read!

Brothers Osborne - Pushing Up Daisies (Love Alive) (Terrapin Care Station Sessions)

Music video by Brothers Osborne performing Pushing Up Daisies (Love Alive). © 2018 UMG Recordings, Inc. http://vevo.ly/uAenT0

Brothers Osborne - Pushing Up Daisies (Love Alive) (Terrapin Care Station Sessions)

Music video by Brothers Osborne performing Pushing Up Daisies (Love Alive). © 2018 UMG Recordings, Inc. http://vevo.ly/uAenT0

How to Write Meta Descriptions in a Constantly Changing World (AKA Google Giveth, Google Taketh Away)

Posted by Dr-Pete

Summary: As of mid-May 2018, Google has reverted back to shorter display snippets. Our data suggests these changes are widespread and that most meta descriptions are being cut off in the previous range of about 155–160 characters.

Back in December, Google made a significant shift in how they displayed search snippets, with our research showing many snippets over 300 characters. Over the weekend, they seem to have rolled back that change (Danny Sullivan partially confirmed this on Twitter on May 14). Besides the obvious question — What are the new limits? — it may leave you wondering how to cope when the rules keep changing. None of us have a crystal ball, but I'm going to attempt to answer both questions based on what we know today.

Lies, dirty lies, and statistics...

I pulled all available search snippets from the MozCast 10K (page-1 Google results for 10,000 keywords), since that's a data set we collect daily and that has a rich history. There were 89,383 display snippets across that data set on the morning of May 15.

I could tell you that, across the entire data set, the minimum length was 6 characters, the maximum was 386, and the mean was about 159. That's not very useful, for a couple of reasons. First, telling you to write meta descriptions between 6–386 characters isn't exactly helpful advice. Second, we're dealing with a lot of extremes. For example, here's a snippet on a search for "USMC":

Marine Corps Community Services may be a wonderful organization, but I'm sorry to report that their meta description is, in fact, "apple" (Google appends the period out of, I assume, desperation). Here's a snippet for a search on the department store "Younkers":

Putting aside their serious multi-brand confusion, I think we can all agree that "BER Meta TAG1" is not optimal. If these cases teach you anything, it's only about what not to do. What about on the opposite extreme? Here's a snippet with 386 characters, from a search for "non-compete agreement":

Notice the "Jump to Exceptions" and links at the beginning. Those have been added by Google, so it's tough to say what counts against the character count and what doesn't. Here's one without those add-ons that clocks in at 370 characters, from a search for "the Hunger Games books":

So, we know that longer snippets do still exist. Note, though, that both of these snippets come from Wikipedia, which is an exception to many SEO rules. Are these long descriptions only fringe cases? Looking at the mean (or even the median, in this case) doesn't really tell us.

The big picture, part 1

Sometimes, you have to let the data try to speak for itself, with a minimum of coaxing. Let's look at all of the snippets that were cut off (ending in "...") and remove video results (we know from previous research that these skew a bit shorter). This leaves 42,863 snippets (just under half of our data set). Here's a graph of all of the cut-off lengths, gathered into 25 character bins (0–25, 26–50, etc.):

This looks very different from our data back in December, and is clearly clustered in the 150–175 character range. We see a few Google display snippets cut off after the 300+ range, but those are dwarfed by the shorter cut-offs.

The big picture, part 2

Obviously, there's a lot happening in that 125–175 character range, so let's zoom in and look at just the middle portion of the frequency distribution, broken up into smaller, 5-character buckets:

We can see pretty clearly that the bulk of cut-offs are happening in the 145–165 character range. Before December, our previous guidelines for meta descriptions were to keep them below 155 characters, so it appears that Google has more-or-less reverted to the old rules.

Keep in mind that Google uses proportional fonts, so there is no exact character limit. Some people have hypothesized a pixel-width limit, like with title tags, but I've found that more difficult to pin down with multi-line snippets (the situation gets even weirder on mobile results). Practically, it's also difficult to write to a pixel limit. The data suggests that 155 characters is a reasonable approximation.

To the Wayback Machine... ?!

Should we just go back to a 155 character cut-off? If you've already written longer meta descriptions, should you scrap that work and start over? The simple truth is that none of us know what's going to happen next week. The way I see it, we have four viable options:

(1) Let Google handle it

Some sites don't have meta descriptions at all. Wikipedia happens to be one of them. Now, Google's understanding of Wikipedia's content is much deeper than most sites (thanks, in part, to Wikidata), but many sites do fare fine without the tag. If your choice is to either write bad, repetitive tags or leave them blank, then I'd say leave them blank and let Google sort it out.

(2) Let the ... fall where it may

You could just write to the length you think is ideal for any given page (within reason), and if the snippets get cut off, don't worry about it. Maybe the ellipsis (...) adds intrigue. I'm half-joking, but the reality is that a cut-off isn't the kiss of death. A good description should entice people to want to read more.

(3) Chop everything at 155 characters

You could go back and mercilessly hack all of your hard work back to 155 characters. I think this is generally going to be time badly spent and may result in even worse search snippets. If you want to rewrite shorter Meta Descriptions for your most important pages, that's perfectly reasonable, but keep in mind that some results are still showing longer snippets and this situation will continue to evolve.

(4) Write length-adaptive descriptions

Is it possible to write a description that works well at both lengths? I think it is, with some care and planning. I wouldn't necessarily recommend this for every single page, but maybe there is a way to have our cake and eat at least half of it, too...

The 150/150 approach

I've been a bit obsessed with the "inverted pyramid" style of writing lately. This is a journalistic style where you start with the lead or summary of your main point and then break that down into the details, data, and context. While this approach is well suited to the web, its origins come from layout limitations in print. You never knew when your editor would have to cut your article short to fit the available space, so the inverted pyramid style helped guarantee that the most important part would usually be spared.

What if we took this approach to meta descriptions? In other words, why not write a 150-character "lead" that summarizes the page, and then add 150 characters of useful but less essential detail (when adding that detail makes sense and provides value)? The 150/150 isn't a magic number — you could even do 100/100 or 100/200. The key is to make sure that the text before the cut can stand on its own.

Think of it a bit like an ad, with two separate lines of copy. Let's take this blog post:

Line 1 (145 chars.)

In December, we reported that Google increased search snippets to over 300 characters. Unfortunately, it looks like the rules have changed again.

Line 2 (122 chars.)

According to our new research (May 2018), the limit is back to 155-160 characters. How should SEOs adapt to these changes?

Line 1 has the short version of the story and hopefully lets searchers know they're heading down the right path. Line 2 dives into a few details and gives away just enough data (hopefully) to be intriguing. If Google uses the longer description, it should work nicely, but if they don't, we shouldn't be any worse for wear.

Should you even bother?

Is this worth the effort? I think writing effective descriptions that engage search visitors is still very important, in theory (and that this indirectly impacts even ranking), but you may find you can write perfectly well within a 155-character limit. We also have to face the reality that Google seems to be rewriting more and more descriptions. This is difficult to measure, as many rewrites are partial, but there's no guarantee that your meta description will be used as written.

Is there any way to tell when a longer snippet (>300 characters) will still be used? Some SEOs have hypothesized a link between longer snippets and featured snippets at the top of the page. In our overall data set, 13.3% of all SERPs had featured snippets. If we look at just SERPs with a maximum display snippet length of 160 characters (i.e. no result was longer than 160 characters), the featured snippet occurrence was 11.4%. If we look at SERPs with at least one display snippet over 300 characters, featured snippets occurred at a rate of 41.8%. While that second data set is fairly small, it is a striking difference. There does seem to be some connection between Google's ability to extract answers in the form of featured snippets and their ability or willingness to display longer search snippets. In many cases, though, these longer snippets are rewrites or taken directly from the page, so even then there's no guarantee that Google will use your longer meta description.

For now, it appears that the 155-character guideline is back in play. If you've already increased some of your meta descriptions, I don't think there's any reason to panic. It might make sense to rewrite overly-long descriptions on critical pages, especially if the cut-offs are leading to bad results. If you do choose to rewrite some of them, consider the 150/150 approach — at least then you'll be a bit more future-proofed.

Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don't have time to hunt down but want to read!

A Machine Learning Guide for Average Humans

Posted by alexis-sanders

Machine learning (ML) has grown consistently in worldwide prevalence. Its implications have stretched from small, seemingly inconsequential victories to groundbreaking discoveries. The SEO community is no exception. An understanding and intuition of machine learning can support our understanding of the challenges and solutions Google's engineers are facing, while also opening our minds to ML's broader implications.

The advantages of gaining an general understanding of machine learning include:

• Gaining empathy for engineers, who are ultimately trying to establish the best results for users
• Understanding what problems machines are solving for, their current capabilities and scientists' goals
• Understanding the competitive ecosystem and how companies are using machine learning to drive results
• Preparing oneself for for what many industry leaders call a major shift in our society (Andrew Ng refers to AI as a "new electricity")
• Understanding basic concepts that often appear within research (it's helped me with understanding certain concepts that appear within Google Brain's Research)
• Growing as an individual and expanding your horizons (you might really enjoy machine learning!)
• When code works and data is produced, it's a very fulfilling, empowering feeling (even if it's a very humble result)

I spent a year taking online courses, reading books, and learning about learning (...as a machine). This post is the fruit borne of that labor -- it covers 17 machine learning resources (including online courses, books, guides, conference presentations, etc.) comprising the most affordable and popular machine learning resources on the web (through the lens of a complete beginner). I've also added a summary of "If I were to start over again, how I would approach it."

This article isn't about credit or degrees. It's about regular Joes and Joannas with an interest in machine learning, and who want to spend their learning time efficiently. Most of these resources will consume over 50 hours of commitment. Ain't nobody got time for a painful waste of a work week (especially when this is probably completed during your personal time). The goal here is for you to find the resource that best suits your learning style. I genuinely hope you find this research useful, and I encourage comments on which materials prove most helpful (especially ones not included)! #HumanLearningMachineLearning

---

Executive summary:

Here's everything you need to know in a chart:

Machine Learning Resource	Time (hours)	Cost ($)	Year	Credibility	Code	Math	Enjoyability
Jason Maye's Machine Learning 101 slidedeck: 2 years of headbanging, so you don't have to	2	$0	'17
{ML} Recipes with Josh Gordon Playlist	2	$0	'16
Machine Learning Crash Course	15	$0	'18
OCDevel Machine Learning Guide Podcast	30	$0	'17-
Kaggle's Machine Learning Track (part 1)	6	$0	'17
Fast.ai (part 1)	70	$70*	'16
Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems	20	$25	'17
Udacity's Intro to Machine Learning (Kate/Sebastian)	60	$0	'15
Andrew Ng's Coursera Machine Learning	55	$0	'11
iPullRank Machine Learning Guide	3	$0	'17
Review Google PhD	2	$0	'17
Caltech Machine Learning on iTunes	27	$0	'12
Pattern Recognition & Machine Learning by Christopher Bishop	150	$75	'06				N/A
Machine Learning: Hands-on for Developers and Technical Professionals	15	$50	'15
Introduction to Machine Learning with Python: A Guide for Data Scientists	15	$25	'16
Udacity's Machine Learning by Georgia Tech	96	$0	'15
Machine Learning Stanford iTunes by Andrew Ng	25	$0	'08				N/A

*Free, but there is the cost of running an AWS EC2 instance (~$70 when I finished, but I did tinker a ton and made a Rick and Morty script generator, which I ran many epochs [rounds] of...)

---

Here's my suggested program:

1. Starting out (estimated 60 hours)

Start with shorter content targeting beginners. This will allow you to get the gist of what's going on with minimal time commitment.

• Commit three hours to Jason Maye's Machine Learning 101 slidedeck: 2 years of headbanging, so you don't have to.
• Commit two hours to watch Google's {ML} Recipes with Josh Gordon YouTube Playlist.
• Sign up for Sam DeBrule's Machine Learnings newsletter.
• Work through Google's Machine Learning Crash Course.
• Start listening to OCDevel's Machine Learning Guide Podcast (skip episodes 1, 3, 16, 21, and 26) in your car, working out, and/or when using hands and eyes for other activities.
• Commit two days to working through Kaggle's Machine Learning Track part 1.

2. Ready to commit (estimated 80 hours)

By this point, learners would understand their interest levels. Continue with content focused on applying relevant knowledge as fast as possible.

• Commit to Fast.ai 10 hours per week, for 7 weeks. If you have a friend/mentor that can help you work through AWS setup, definitely lean on any support in installation (it's 100% the worst part of ML).
• Acquire Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems, and read the first two chapters immediately. Then use this as supplemental to the Fast.ai course.

3. Broadening your horizons (estimated 115 hours)

If you've made it through the last section and are still hungry for more knowledge, move on to broadening your horizons. Read content focused on teaching the breadth of machine learning -- building an intuition for what the algorithms are trying to accomplish (whether visual or mathematically).

• Start watching videos and participating in Udacity's Intro to Machine Learning (by Sebastian Thrun and Katie Malone).
• Work through Andrew Ng's Coursera Machine Learning course.

Your next steps

By this point, you will already have AWS running instances, a mathematical foundation, and an overarching view of machine learning. This is your jumping-off point to determine what you want to do.

You should be able to determine your next step based on your interest, whether it's entering Kaggle competitions; doing Fast.ai part two; diving deep into the mathematics with Pattern Recognition & Machine Learning by Christopher Bishop; giving Andrew Ng's newer Deeplearning.ai course on Coursera; learning more about specific tech stacks (TensorFlow, Scikit-Learn, Keras, Pandas, Numpy, etc.); or applying machine learning to your own problems.

---

Why am I recommending these steps and resources?

I am not qualified to write an article on machine learning. I don't have a PhD. I took one statistics class in college, which marked the first moment I truly understood "fight or flight" reactions. And to top it off, my coding skills are lackluster (at their best, they're chunks of reverse-engineered code from Stack Overflow). Despite my many shortcomings, this piece had to be written by someone like me, an average person.

Statistically speaking, most of us are average (ah, the bell curve/Gaussian distribution always catches up to us). Since I'm not tied to any elitist sentiments, I can be real with you. Below contains a high-level summary of my reviews on all of the classes I took, along with a plan for how I would approach learning machine learning if I could start over. Click to expand each course for the full version with notes.

---

In-depth reviews of machine learning courses:

Starting out

Jason Maye's Machine Learning 101 slidedeck: 2 years of head-banging, so you don't have to ↓

Need to Know: A stellar high-level overview of machine learning fundamentals in an engaging and visually stimulating format.

Loved:

• Very user-friendly, engaging, and playful slidedeck.
• Has the potential to take some of the pain out of the process, through introducing core concepts.
• Breaks up content by beginner/need-to-know (green), and intermediate/less-useful noise (specifically for individuals starting out) (blue).
• Provides resources to dive deeper into machine learning.
• Provides some top people to follow in machine learning.

Disliked:

• That there is not more! Jason's creativity, visual-based teaching approach, and quirky sense of humor all support the absorption of the material.

Lecturer:

Jason Mayes:

• Senior Creative Technologist and Research Engineer at Google
• Masters in Computer Science from University of Bristols
• Personal Note: He's also kind on Twitter! :)

Links:

• Machine Learning 101 slide deck

Tips on Watching:

• Set aside 2-4 hours to work through the deck once.
• Since there is a wealth of knowledge, refer back as needed (or as a grounding source).
• Identify areas of interest and explore the resources provided.

{ML} Recipes with Josh Gordon ↓

Need to Know: This mini-series YouTube-hosted playlist covers the very fundamentals of machine learning with opportunities to complete exercises.

Loved:

• It is genuinely beginner-focused.
  • They make no assumption of any prior knowledge.
  • Gloss over potentially complex topics that may serve as noise.
  • Playlist ~2 hours
• Very high-quality filming, audio, and presentation, almost to the point where it had its own aesthetic.
• Covers some examples in scikit-learn and TensorFlow, which felt modern and practical.
• Josh Gordon was an engaging speaker.

Disliked:

• I could not get Dockers on Windows (suggested package manager). This wasn't a huge deal, since I already had my AWS setup by this point; however, a bit of a bummer since it made it impossible to follow certain steps exactly.
  • Issue: Every time I tried to download (over the course of two weeks), the .exe file would recursively start and keep spinning until either my memory ran out, computer crashed, or I shut my computer down. I sent this to Docker's Twitter account to no avail.

Lecturer:

Josh Gordon:

• Developer Advocate for at TensorFlow at Google
• Leads Machine Learning advocacy at Google
• Member of the Udacity AI & Data Industry Advisory Board
• Masters in Computer Science from Columbia University

Links:

• Hello World - Machine Learning Recipes #1 (YouTube)
• GitHub: Machine Learning Recipes with Josh Gordon

Tips on Watching:

• The playlist is short (only ~1.5 hours screen time). However, it can be a bit fast-paced at times (especially if you like mimicking the examples), so set aside 3-4 hours to play around with examples and allow time for installation, pausing, and following along.
• Take time to explore code labs.

Google's Machine Learning Crash Course with TensorFlow APIs ↓

Need to Know: A Google researcher-made crash course on machine learning that is interactive and offers its own built-in coding system!

Loved:

• Different formats of learning: high-quality video (with ability to adjust speed, closed captioning), readings, quizzes (with explanations), visuals (including whiteboarding), interactive components/ playgrounds, code lab exercises (run directly in your browser (no setup required!))
• Non-intimidating
  • One of my favorite quotes: "You don't need to understand the math to be able to take a look at the graphical interpretation."
  • Broken down into digestible sections
  • Introduces key terms

Disliked:

• N/A

Lecturers:

Multiple Google researchers participated in this course, including:

• Peter Norvig
  • Director of Research at Google Inc.
  • Previously he directed Google's core search algorithms group.
  • He is co-author of Artificial Intelligence: A Modern Approach
• D. Sculley
  • Senior Staff Software Engineer at Google
  • KDD award-winning papers
  • Works on massive-scale ML systems for online advertising
  • Was part of a research ML paper on optimizing chocolate chip cookies
  • According to his personal website, he prefers to go by "D."
• Cassandra Xia
  • Programmer, Software Engineer at Google
  • She has some really cool (and cute) projects based on learning statistics concepts interactively
• Maya Gupta
  • Leads Glassbox Machine Learning R&D team at Google
  • Associate Professor of Electrical Engineering at the University of Washington (2003-2012)
  • In 2007, Gupta received the PECASE award from President George Bush for her work in classifying uncertain (e.g. random) signals
  • Gupta also runs Artifact Puzzles, the second-largest US maker of wooden jigsaw puzzles
• Sally Goldman
  • Research Scientist at Google
  • Co-author of A Practical Guide to Data Structures and Algorithms Using Java
  • Numerous journals, classes taught at Washington University, and contributions to the ML community

Links:

• Machine Learning Crash Course

Tips on Doing:

• Actively work through playground and coding exercises

OCDevel's Machine Learning Guide Podcast ↓

Need to Know: This podcast focuses on the high-level fundamentals of machine learning, including basic intuition, algorithms, math, languages, and frameworks. It also includes references to learn more on each episode's topic.

Loved:

• Great for trips (when traveling a ton, it was an easy listen).
• The podcast makes machine learning fun with interesting and compelling analogies.
• Tyler is a big fan of Andrew Ng's Coursera course and reviews concepts in Coursera course very well, such that both pair together nicely.
• Covers the canonical resources for learning more on a particular topic.

Disliked:

• Certain courses were more theory-based; all are interesting, yet impractical.
• Due to limited funding the project is a bit slow to update and has less than 30 episodes.

Podcaster:

Tyler Renelle:

• Machine learning engineer focused on time series and reinforcement
• Background in full-stack JavaScript, 10 years web and mobile
• Creator of HabitRPG, an app that treats habits as an RPG game

Links:

• Machine Learning Guide podcast
• Machine Learning Guide podcast (iTunes)

Tips on Listening:

• Listen along your journey to help solidify understanding of topics.
• Skip episodes 1, 3, 16, 21, and 26 (unless their topics interest and inspire you!).

Kaggle Machine Learning Track (Lesson 1) ↓

Need to Know: A simple code lab that covers the very basics of machine learning with scikit-learn and Panda through the application of the examples onto another set of data.

Loved:

• A more active form of learning.
• An engaging code lab that encourages participants to apply knowledge.
• This track offers has a built-in Python notebook on Kaggle with all input files included. This removed any and all setup/installation issues.
  • Side note: It's a bit different than Jupyter notebook (e.g., have to click into a cell to add another cell).
• Each lesson is short, which made the entire lesson go by very fast.

Disliked:

• The writing in the first lesson didn't initially make it clear that one would need to apply the knowledge in the lesson to their workbook.
  • It wasn't a big deal, but when I started referencing files in the lesson, I had to dive into the files in my workbook to find they didn't exist, only to realize that the knowledge was supposed to be applied and not transcribed.

Lecturer:

Dan Becker:

• Data Scientist at Kaggle
• Undergrad in Computer Science, PhD in Econometrics
• Supervised data science consultant for six Fortune 100 companies
• Contributed to the Keras and Tensorflow libraries
• Finished 2nd (out of 1353 teams) in $3 million Heritage Health Prize data mining competition
• Speaks at deep learning workshops at events and conferences

Links:

• https://www.kaggle.com/learn/machine-learning

Tips on Doing:

• Read the exercises and apply to your dataset as you go.
• Try lesson 2, which covers more complex/abstract topics (note: this second took a bit longer to work through).

---

Ready to commit

Fast.ai (part 1 of 2) ↓

Need to Know: Hands-down the most engaging and active form of learning ML. The source I would most recommend for anyone (although the training plan does help to build up to this course). This course is about learning through coding. This is the only course that I started to truly see the practical mechanics start to come together. It involves applying the most practical solutions to the most common problems (while also building an intuition for those solutions).

Loved:

• Course Philosophy:
  • Active learning approach
    • "Go out into the world and understand underlying mechanics (of machine learning by doing)."
  • Counter-culture to the exclusivity of the machine learning field, focusing on inclusion.
    • "Let's do shit that matters to people as quickly as possible."
• Highly pragmatic approach with tools that are currently being used (Jupyter Notebooks, scikit-learn, Keras, AWS, etc.).
• Show an end-to-end process that you get to complete and play with in a development environment.
• Math is involved, but is not prohibitive. Excel files helped to consolidate information/interact with information in a different way, and Jeremy spends a lot of time recapping confusing concepts.
• Amazing set of learning resources that allow for all different styles of learning, including:
  • Video Lessons
  • Notes
  • Jupyter Notebooks
  • Assignments
  • Highly active forums
  • Resources on Stackoverflow
  • Readings/resources
    • Jeremy often references popular academic texts
  • Jeremy's TEDx talk in Brussels
• Jeremy really pushes one to do extra and put in the effort by teaching interesting problems and engaging one in solving them.
• It's a huge time commitment; however, it's worth it.
• All of the course's profits are donated.

Disliked:

• Overview covers their approach to learning (obviously I'm a fan!). If you're already drinking the Kool-aid, skip past.
• I struggled through the AWS setup (13-minute video) for about five hours (however, it felt so good when it was up and running!).
• Because of its practicality and concentration on solutions used today to solve popular problem types (image recognition, text generation, etc.), it lacks breadth of machine learning topics.

Lecturers:

Jeremy Howard:

• Distinguished Research Scientist at the University of San Francisco
• Faculty member at Singularity University
• Young Global Leader with the World Economic Forum
• Founder of Enlitic (the first company to apply deep learning to medicine)
• Former President and Chief Scientist of the data science platform Kaggle

Rachel Thomas:

• PhD in Math from Duke
• One of Forbes' "20 Incredible Women Advancing AI Research"
• Researcher-in-residence at the University of San Francisco Data Institute
• Teaches in the Masters in Data Science program

Links:

• http://course.fast.ai/start.html
• http://wiki.fast.ai/index.php/Main_Page
• https://github.com/fastai/courses/tree/master/deeplearning1/nbs

Tips on Doing:

• Set expectations with yourself that installation is going to probably take a few hours.
• Prepare to spend about ~70 hours for this course (it's worth it).
• Don't forget to shut off your AWS instance.
• Balance out machine learning knowledge with a course with more breadth.
• Consider giving part two of the Fast.ai program a shot!

Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems ↓

Need to Know: This book is an Amazon best seller for a reason. It covers a lot of ground quickly, empowers readers to walk through a machine learning problem by chapter two, and contains practical up-to-date machine learning skills.

Loved:

• Book contains an amazing introduction to machine learning that briskly provides an overarching quick view of the machine learning ecosystem.
• Chapter 2 immediately walks the reader through an end-to-end machine learning problem.
• Immediately afterwards, Aurélien pushes a user to attempt to apply this solution to another problem, which was very empowering.
• There are review questions at the end of each chapter to ensure on has grasped the content within the chapter and to push the reader to explore more.
• Once installation was completed, it was easy to follow and all code is available on GitHub.
• Chapters 11-14 were very tough reading; however, they were a great reference when working through Fast.ai.
• Contains some powerful analogies.
• Each chapter's introductions were very useful and put everything into context. This general-to-specifics learning was very useful.

Disliked:

• Installation was a common source of issues during the beginning of my journey; the text glided over this. I felt the frustration that most people experience from installation should have been addressed with more resources.

Writer:

Aurélien Géron:

• Led the YouTube video classification team from 2013 to 2016
• Currently a machine Learning consultant
• Founder and CTO of Wifirst and Polyconseil
• Published technical books (on C++, Wi-Fi, and Internet architectures)

Links:

• https://www.amazon.com/_/dp/1491962291?tag=oreilly20-20
• http://shop.oreilly.com/product/0636920052289.do
• https://github.com/ageron/handson-ml

Tips on Using:

• Get a friend with Python experience to help with installation.
• Read the introductions to each chapter thoroughly, read the chapter (pay careful attention to code), review the questions at the end (highlight any in-text answer), make a copy of Aurélien's GitHub and make sure everything works on your setup, re-type the notebooks, go to Kaggle and try on other datasets.

---

Broadening your horizons

Udacity: Intro to Machine Learning (Kate/Sebastian) ↓

Need to Know: A course that covers a range of machine learning topics, supports building of intuition via visualization and simple examples, offers coding challenges, and a certificate (upon completion of a final project). The biggest challenge with this course is bridging the gap between the hand-holding lectures and the coding exercises.

Loved:

• Focus on developing a visual intuition on what each model is trying to accomplish.
• This visual learning mathematics approach is very useful.
• Cover a vast variety and breadth of models and machine learning basics.
• In terms of presenting the concept, there was a lot of hand-holding (which I completely appreciated!).
• Many people have done this training, so their GitHub accounts can be used as reference for the mini-projects.
• Katie actively notes documentation and suggests where viewers can learn more/reference material.

Disliked:

• All of the conceptual hand-holding in the lessons is a stark contrast to the challenges of installation, coding exercises, and mini-projects.
  • This is the first course started and the limited instructions on setting up the environment and many failed attempts caused me to break down crying at least a handful of times.
• The mini-projects are intimidating.
• There is extra code added to support the viewers; however, it's done so with little acknowledgement as to what it's actually doing. This made learning a bit harder.

Lecturer:

Caitlin (Katie) Malone:

• Director of Data Science Research and Development at Civis Analytics
• Stanford PhD in Experimental Particle Physics
• Intern at Udacity in summer 2014
• Graduate Researcher at the SLAC National Accelerator Laboratory
  • https://www6.slac.stanford.edu/
• Podcaster with Ben Jaffe (currently Facebook UI Engineer and a music aficionado) on a machine learning podcast Linear Digressions (100+ episodes)

Sebastian Thrun:

• CEO of the Kitty Hawk Corporation
• Chairman and co-founder of Udacity
  • One of my favorite Sebastian quotes: "It occurred to me, I could be at Google and build a self-driving car, or I can teach 10,000 students how to build self-driving cars."
• Former Google VP
  • Founded Google X
  • Led development of the robotic vehicle Stanley
• Professor of Computer Science at Stanford University
• Formerly a professor at Carnegie Mellon University.

Links:

• https://www.udacity.com/course/intro-to-machine-learning--ud120
• Udacity also offers a next step, the Machine Learning Engineer Nanodegree, which will set one back about $1K.

Tips on Watching:

• Get a friend to help you set up your environment.
• Print mini-project instructions to check off each step.

Andrew Ng's Coursera Machine Learning Course ↓

Need to Know: The Andrew Ng Coursera course is the most referenced online machine learning course. It covers a broad set of fundamental, evergreen topics with a strong focus in building mathematical intuition behind machine learning models. Also, one can submit assignments and earn a grade for free. If you want to earn a certificate, one can subscribe or apply for financial aid.

Loved:

• This course has a high level of credibility.
• Introduces all necessary machine learning terminology and jargon.
• Contains a very classic machine learning education approach with a high level of math focus.
• Quizzes interspersed in courses and after each lesson support understanding and overall learning.
• The sessions for the course are flexible, the option to switch into a different section is always available.

Disliked:

• The mathematic notation was hard to process at times.
• The content felt a bit dated and non-pragmatic. For example, the main concentration was MATLAB and Octave versus more modern languages and resources.
• Video quality was less than average and could use a refresh.

Lecturer:

Andrew Ng:

• Adjunct Professor, Stanford University (focusing on AI, Machine Learning, and Deep Learning)
• Co-founder of Coursera
• Former head of Baidu AI Group
• Founder and previous head of Google Brain (deep learning) project
• Former Director of the Stanford AI Lab
• Chairman of the board of Woebot (a machine learning bot that focuses on Cognitive Behavior Therapy)

Links:

• https://www.coursera.org/learn/machine-learning/
• Andrew Ng recently launched a new course (August 2017) called DeepLearning.ai, a ~15 week course containing five mini-courses ($49 USD per month to continue learning after trial period of 7 days ends).
  • Course: https://www.coursera.org/specializations/deep-learning
    • Course 1: Neural Networks and Deep Learning
    • Course 2: Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization
    • Course 3: Structuring Machine Learning Projects
    • Course 4: Convolutional Neural Networks
    • Course 5: Sequence Models

Tips on Watching:

• Be disciplined with setting aside timing (even if it's only 15 minutes a day) to help power through some of the more boring concepts.
• Don't do this course first, because it's intimidating, requires a large time commitment, and isn't a very energizing experience.

---

Additional machine learning opportunities

iPullRank Machine Learning Guide ↓

Need to Know: A machine learning e-book targeted at marketers.

Loved:

• Targeted at marketers and applied to organic search.
• Covers a variety of machine learning topics.
• Some good examples, including real-world blunders.
• Gives some practical tools for non-data scientists (including: MonkeyLearn and Orange)
  • I found Orange to be a lot of fun. It struggled with larger datasets; however, it has a very visual interface that was more user-friendly and offers potential to show some pretty compelling stories.
    • Example: World Happiness Dataset by:
      • X-axis: Happiness Score
      • Y-axis: Economy
      • Color: Health

Disliked:

• Potential to break up content more with relevant imagery -- the content was very dense.

Writers:

iPullRank Team (including Mike King):

• Mike King has a few slide decks on the basics of machine learnings and AI
• iPullRank has a few data scientists on staff

Links:

• http://ipullrank.com/machine-learning-guide/

Tips on Reading:

• Read chapters 1-6 and the rest depending upon personal interest.

Review Google PhD ↓

Need to Know: A two-hour presentation from Google's 2017 IO conference that walks through getting 99% accuracy on the MNIST dataset (a famous dataset containing a bunch of handwritten numbers, which the machine must learn to identify the numbers).

Loved:

• This talk struck me as very modern, covering the cutting edge.
• Found this to be very complementary to Fast.ai, as it covered similar topics (e.g. ReLu, CNNs, RNNs, etc.)
• Amazing visuals that help to put everything into context.

Disliked:

• The presentation is only a short conference solution and not a comprehensive view of machine learning.
• Also, a passive form of learning.

Presenter:

Martin Görner:

• Developer Relations, Google (since 2011)
• Started Mobipocket, a startup that later became the software part of the Amazon Kindle and its mobile variants

Links:

• Part 1 - https://www.youtube.com/watch?v=u4alGiomYP4
• Part 2 - https://www.youtube.com/watch?v=fTUwdXUFfI8

Tips on Watching:

• Google any concepts you're unfamiliar with.
• Take your time with this one; 2 hours of screen time doesn't count all of the Googling and processing time for this one.

Caltech Machine Learning iTunes ↓

Need to Know: If math is your thing, this course does a stellar job of building the mathematic intuition behind many machine learning models. Dr. Abu-Mostafa is a raconteur, includes useful visualizations, relevant real-world examples, and compelling analogies.

Loved:

• First and foremost, this is a real Caltech course, meaning it's not a watered-down version and contains fundamental concepts that are vital to understanding the mechanics of machine learning.
• On iTunes, audio downloads are available, which can be useful for on-the-go learning.
• Dr. Abu-Mostafa is a skilled speaker, making the 27 hours spent listening much easier!
• Dr. Abu-Mostafa offers up some strong real-world examples and analogies which makes the content more relatable.
  • As an example, he asks students: "Why do I give you practice exams and not just give you the final exam?" as an illustration of why a testing set is useful. If he were to just give students the final, they would just memorize the answers (i.e., they would overfit to the data) and not genuinely learn the material. The final is a test to show how much students learn.
• The last 1/2 hour of the class is always a Q&A, where students can ask questions. Their questions were useful to understanding the topic more in-depth.
• The video and audio quality was strong throughout. There were a few times when I couldn't understand a question in the Q&A, but overall very strong.
• This course is designed to build mathematical intuition of what's going on under the hood of specific machine learning models.
  • Caution: Dr. Abu-Mostafa uses mathematical notation, but it's different from Andrew Ng's (e.g., theta = w).
• The final lecture was the most useful, as it pulled a lot of the conceptual puzzle pieces together. The course on neural networks was a close second!

Disliked:

• Although it contains mostly evergreen content, being released in 2012, it could use a refresh.
• Very passive form of learning, as it wasn't immediately actionable.

Lecturer:

Dr. Yaser S. Abu-Mostafa:

• Professor of Electrical Engineering and Computer Science at the California Institute of Technology
• Chairman of Machine Learning Consultants LLC
• Serves on a number of scientific advisory boards
• Has served as a technical consultant on machine learning for several companies (including Citibank).
• Multiple articles in Scientific American

Links:

• https://work.caltech.edu/telecourse.html
• https://itunes.apple.com/us/course/machine-learning/id515364596

Tips on Watching:

• Consider listening to the last lesson first, as it pulls together the course overall conceptually. The map of the course, below, was particularly useful to organizing the information taught in the courses.

Image source: http://work.caltech.edu/slides/slides18.pdf

"Pattern Recognition & Machine Learning" by Christopher Bishop ↓

Need to Know: This is a very popular college-level machine learning textbook. I've heard it likened to a bible for machine learning. However, after spending a month trying to tackle the first few chapters, I gave up. It was too much math and pre-requisites to tackle (even with a multitude of Google sessions).

Loved:

• The text of choice for many major universities, so if you can make it through this text and understand all of the concepts, you're probably in a very good position.
• I appreciated the history aside sections, where Bishop talked about influential people and their career accomplishments in statistics and machine learning.
• Despite being a highly mathematically text, the textbook actually has some pretty visually intuitive imagery.

Disliked:

• I couldn't make it through the text, which was a bit frustrating. The statistics and mathematical notation (which is probably very benign for a student in this topic) were too much for me.
• The sunk cost was pretty high here (~$75).

Writer:

Christopher Bishop:

• Laboratory Director at Microsoft Research Cambridge
• Professor of Computer Science at the University of Edinburgh
• Fellow of Darwin College, Cambridge
• PhD in Theoretical Physics from the University of Edinburgh

Links:

• https://www.amazon.com/Pattern-Recognition-Learning-Information-Statistics/dp/0387310738/ref=sr_1_2?ie=UTF8&qid=1516839475&sr=8-2&keywords=Pattern+Recognition+%26+Machine+Learning

Tips on Reading:

• Don't start your machine learning journey with this book.
• Get a friend in statistics to walk you through anything complicated (my plan is to get a mentor in statistics).
• Consider taking a (free) online statistics course (Khan Academy and Udacity both have some great content on statistics, calculus, math, and data analysis).

Machine Learning: Hands-on for Developers and Technical Professionals ↓

Need to Know: A fun, non-intimidating end-to-end launching pad/whistle stop for machine learning in action.

Loved:

• Talks about practical issues that many other sources didn't really address (e.g. data-cleansing).
• Covered the basics of machine learning in a non-intimidating way.
  • Offers abridged, consolidated versions of the content.
  • Added fun anecdotes that makes it easier to read.
  • Overall the writer has a great sense of humor.
  • Writer talks to the reader as if they're a real human being (i.e., doesn't expect you to go out and do proofs; acknowledges the challenge of certain concepts).
• Covers a wide variety of topics.
• Because it was well-written, I flew through the book (even though it's about ~300 pages).

Disliked:

• N/A

Writer:

Jason Bell:

• Technical architect, lecturer, and startup consultant
• Data Engineer at MastodonC
• Former section editor for Java Developer's Journal
• Former writer on IBM DeveloperWorks

Links:

• https://www.amazon.com/Machine-Learning-Hands-Developers-Professionals/dp/1118889061
• https://www.wiley.com/en-us/Machine+Learning%3A+Hands+On+for+Developers+and+Technical+Professionals-p-9781118889060
• Jason's Blog: https://dataissexy.wordpress.com/

Tips on Reading:

• Download and explore Weka's interface beforehand.
• Give some of the exercises a shot.

Introduction to Machine Learning with Python: A Guide for Data Scientists ↓

Need to Know: This was a was a well-written piece on machine learning, making it a quick read.

Loved:

• Quick, smooth read.
• Easy-to-follow code examples.
• The first few chapters served as a stellar introduction to the basics of machine learning.
• Contain subtle jokes that add a bit of fun.
• Tip to use the Python package manager Anaconda with Jupyter Notebooks was helpful.

Disliked:

• Once again, installation was a challenge.
• The "mglearn" utility library threw me for a loop. I had to reread the first few chapters before I figured out it was support for the book.
• Although I liked the book, I didn't love it. Overall it just missed the "empowering" mark.

Writers:

Andreas C. Müller:

• PhD in Computer Science
• Lecturer at the Data Science Institute at Columbia University
• Worked at the NYU Center for Data Science on open source and open science
• Former Machine Learning Scientist at Amazon
• Speaks often on Machine Learning and scikit-learn (a popular machine learning library)
• And he makes some pretty incredibly useful graphics, such as this scikit-learn cheat sheet:

Image source: http://peekaboo-vision.blogspot.com/2013/01/machin...

Sarah Guido:

• Former senior data scientist at Mashable
• Lead data scientist at Bitly
• 2018 SciPy Conference Data Science track co-chair

Links:

• https://www.amazon.com/Introduction-Machine-Learning-Python-Scientists/dp/1449369413/ref=sr_1_7?s=books&ie=UTF8&qid=1516734322&sr=1-7&keywords=python+machine+learning
• http://shop.oreilly.com/product/0636920030515.do

Tips on Reading:

• Type out code examples.
• Beware of the "mglearn" utility library.

Udacity: Machine Learning by Georgia Tech ↓

Need to Know: A mix between an online learning experience and a university machine learning teaching approach. The lecturers are fun, but the course still fell a bit short in terms of active learning.

Loved:

• This class is offered as CS7641 at Georgia Tech, where it is a part of the Online Masters Degree. Although taking this course here will not earn credit towards the OMS degree, it's still a non-watered-down college teaching philosophy approach.
• Covers a wide variety of topics, many of which reminded me of the Caltech course (including: VC Dimension versus Bayesian, Occam's razor, etc.)
• Discusses Markov Decision Chains, which is something that didn't really come up in many other introductory machine learning course, but they are referenced within Google patents.
• The lecturers have a great dynamic, are wicked smart, and displayed a great sense of (nerd) humor, which make the topics less intimidating.
• The course has quizzes, which give the course a slight amount of interaction.

Disliked:

• Some videos were very long, which made the content a bit harder to digest.
• The course overall was very time consuming.
• Despite the quizzes, the course was a very passive form of learning with no assignments and little coding.
• Many videos started with a bunch of content already written out. Having the content written out was probably a big time-saver, but it was also a bit jarring for a viewer to see so much information all at once, while also trying to listen.
• It's vital to pay very close attention to notation, which compounds in complexity quickly.
• Tablet version didn't function flawlessly: some was missing content (which I had to mark down and review on a desktop), the app would crash randomly on the tablet, and sometimes the audio wouldn't start.
• There were no subtitles available on tablet, which I found not only to be a major accessibility blunder, but also made it harder for me to process (since I'm not an audio learner).

Lecturer:

Michael Littman:

• Professor of Computer Science at Brown University.
• Was granted a patent for one of the earliest systems for Cross-language information retrieval
• Perhaps the most interesting man in the world:
  • Been in two TEDx talks
    • How I Learned to Stop Worrying and Be Realistic About AI
    • A Cooperative Path to Artificial Intelligence
  • During his time at Duke, he worked on an automated crossword solver (PROVERB)
  • Has a Family Quartet
  • He has appeared in a TurboTax commercial

Charles Isbell:

• Professor and Executive Associate Dean at School of Interactive Computing at Georgia Tech
• Focus on statistical machine learning and "interactive" artificial intelligence.

Links:

• https://www.udacity.com/course/machine-learning--ud262

Tips on Watching:

• Pick specific topics of interest and focusing on those lessons.

Andrew Ng's Stanford's Machine Learning iTunes ↓

Need to Know: A non-watered-down Stanford course. It's outdated (filmed in 2008), video/audio are a bit poor, and most links online now point towards the Coursera course. Although the idea of watching a Stanford course was energizing for the first few courses, it became dreadfully boring. I made it to course six before calling it.

Loved:

• Designed for students, so you know you're not missing out on anything.
• This course provides a deeper study into the mathematical and theoretical foundation behind machine learning to the point that the students could create their own machine learning algorithms. This isn't necessarily very practical for the everyday machine learning user.
• Has some powerful real-world examples (although they're outdated).
• There is something about the kinesthetic nature of watching someone write information out. The blackboard writing helped me to process certain ideas.

Disliked:

• Video and audio quality were pain to watch.
  • Many questions asked by students were hard to hear.
• On-screen visuals range from hard to impossible to see.
• Found myself counting minutes.
• Dr. Ng mentions TA classes, supplementary learning, but these are not available online.
• Sometimes the video showed students, which I felt was invasive.

Lecturer:

Andrew Ng (see above)

Links:

• https://itunes.apple.com/us/course/machine-learning/id495053006
• https://www.youtube.com/watch?v=UzxYlbK2c7E

Tips on Watching:

• Only watch if you're looking to gain a deeper understanding of the math presented in the Coursera course.
• Skip the first half of the first lecture, since it's mostly class logistics.

Additional Resources

• Fast.ai (part 2) - free access to materials, cost for AWS EC2 instance
• Deeplearning.ai - $50/month
• Udacity Machine Learning Engineer Nanodegree - $1K
• https://machinelearningmastery.com/

---

Motivations and inspiration

If you're wondering why I spent a year doing this, then I'm with you. I'm genuinely not sure why I set my sights on this project, much less why I followed through with it. I saw Mike King give a session on Machine Learning. I was caught off guard, since I knew nothing on the topic. It gave me a pesky, insatiable curiosity itch. It started with one course and then spiraled out of control. Eventually it transformed into an idea: a review guide on the most affordable and popular machine learning resources on the web (through the lens of a complete beginner). Hopefully you found it useful, or at least somewhat interesting. Be sure to share your thoughts or questions in the comments!

Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don't have time to hunt down but want to read!