up

Mooc Work

Machine Learning

June 9th, 2014

This coursera mooc was offered by Stanford University, taught by Andrew Ng.

I've taken a few moocs at this point, and I have to say, this was the best one I've taken so far. Not only the best, but of a much higher teaching quality, like several levels higher. Keep in mind (I found out), Andrew Ng is not only a leading authority on machine learning and artificial intelligence, he's actually a cofounder of the coursera platform itself. He has a vested interest to make his course of a high quality, but if you take this course you'll find he's also a good teacher regardless—it shows in his videos.

As for the course itself: For my own personal gain, and maybe you're the same, I found this to be a much better way to learn the tools of statistics than many statistics courses I've taken. The focus is different, but if you want to get use to the toolset, sometimes it helps to find an alternate motivational approach first. You can always return to the standard statistical treatment after.

As for the course breakdown: Weeks 1-7 deal with what's called supervised learning while the remaining weeks 8-10 focus on unsupervised learning. The special care taken to build intuition for the theoretical tools we use, as well as the insights offered regarding best practices such as how to best use a data set for training are highlights of this course. A lot of courses would rush through the material, but here we even have a week (week 6) to decompress and allow everything we've learned in the previous weeks to sink in. When I finished this mooc, compared with any other, I felt like I had actually been taught from a teacher I respect and like.

Finally, for our assignments we get to use octave (or matlab) as our programming language. I was still pretty new to octave, but after coming out of this course I have a lot more practice with it, and I feel pretty comfortable with its basics and what it offers overall as a language.

If you're wondering whether this mooc is worth it, I highly recommend it. Thanks.

Week 1 - Linear Regression

June 16th, 2014

If I were to summarize this week, I would say (being the first week) it introduces terminology you probably won't understand until later—I didn't, but that's okay, at least being exposed to them helps to know what's coming. Regardless, what did help motivate me this week were the examples, they stand out: Housing prices, (supervised) cancer classifier, the (unsupervised) cocktail party problem and its one line of code algorithm were all very intriguing.

Aside from the introduction, we actually got right into it with linear regression. We discussed ideas and terminology including hypothesis, parameters, cost function, goal (minimize cost function), training set. We also got into gradient descent and the learning rate. I really liked how Andrew Ng points out the most common ways this model could go wrong, including slow convergence, overshooting, failing to converge, not to mention plain old divergence. We also compared the local minimum with the global minimum.

As for personal comments for this week: This mooc is serious! It's a 10 week course and week 1 is 3 hours of video for the lecture.

Okay, so it turns out we're using the octave programming language. From what I can tell, from what Andrew Ng himself has mentioned about it, it's a good prototyping language for machine learning. The way to put it is developer's time is most expensive, so a prototyping language is ideal. If for example you end up needing to implement your prototype in an industry application, you can always translate it after-the-fact into a more efficient and/or optimized language later. There are so many best practices to learn! It's good! :)

I don't think I'll write so much each week, but there was a lot to take in and think about and comment on this week in particular. What I especially enjoyed was the following: "By the time you finish this class, you'll know how to apply the most advanced machine learning algorithms to such problems as anti-spam, image recognition, clustering, building recommender systems, and many other problems. You'll also know how to select the right algorithm for the right job, as well as become expert at 'debugging' and figuring out how to improve a learning algorithm's performance."

Finally, I am excited to say I learned my first machine learning algorithm: Linear regression, implemented with the square error function and gradient decent. Also, it surprised me to learn that machine learning itself uses multivariable calculus. It's nice to apply a bunch of math I had learned in the past that I otherwise never get the chance to use.

Week 2 - Multivariable Linear Regression

June 23rd, 2014

This week can be summarized as: Multivariable linear regression, multivariable gradient descent. In a lot of ways, it's an extension of last week.

In particular though, we looked at best practices such as feature scaling (normalizing features to be on the same scale; to be comparable), mean normalization (centre features so they have nearly zero mean), choice of learning rate (scouting out at different scales). Maybe I'm reading too much into it, but what's interesting about Andrew Ng's best practices is they are in-and-of-themselves strategies and tactics for machine learning, haha. That's great!

We ended with a comparison of the normal equation solution (analytic) to finding our local minimum with the gradient descent (approximation) approach.

In terms of personal comments this week: One of the first practical examples of machine learning we've seen is to write some code to scrape housing market data and use that to provide an estimate of the selling price of an arbitrary house. In real life it's more complicated than that—I don't even know the landscape of where to scrape quality data (retail listings?) and also there are still negotiations I believe, but at the very least this provides you with a near objective measure and anchor for which to negotiate (a starting point). Regardless, that's pretty interesting. Mostly, for as enthused as I am to learn this toolset, I'm trying to be critical of it as well.

Other thoughts: I'm trying to build general stories and narratives about machine learning. It's still early on so I don't yet know what to expect, but one thing that already stands out is to question the nature of optimization which is the other half of engineering (as far as I can tell)—the first being modularization.

I'm kind of jumping around I admit, but I'm having a lot of random thoughts as I'm trying to get a feel for all of this: I will add, I truthfully don't think I would be able to keep up with machine learning without my previous Humanities and literary theory training. This might be a provocative statement, so let me explain: Literary theories allow you to be critical and break down the landscape of a communications infrastructure pretty quickly as a skillset. Any formal math oriented theory still has a language as its interface, so being able to recognize concepts and get a feel for where they fit really helps when learning an entirely new discipline such as machine learning. This is especially true for as I come from a math background, not an engineering background. I'm able to recognize the engineer's way as different, but I don't yet have enough experience with it to know how. My humanities training helps.

As an example, I've been thinking about the nature of a cost function: With regression, you need a measure of error, so first you specify standards (usually intuitive) for which any possible cost function should satisfy (design specification). This implies that cost functions are semiotic spaces (because there's more than one way to represent context [context here being error]).

If that's the case, then of all possible cost functions, the one you choose (design implementation)—you more or less choose one that is most user-friendly of all the ones you could choose. Even then, due to complexity theory, you likely will be able to cluster a few known cost functions (semiotic spaces) none best in the general sense—the choice among these will be narrowed down finally by the actual context itself (or possibly the media space, which is in theory never a best practice, but sometimes you have to work with the real world resource constraints given to you).

Well anyway, for better or worse that's my rant this week, if it even makes sense, ahah. Though if I'm ending on a joke I might as well go all out: Oh, by the way, the "learning" in machine learning is equivalent to being able to find your way downhill. So it turns out us humans are really only as smart as gravity. Zing!

Week 3 - Logistic Regression

June 30th, 2014

This week we changed topics toward logistic regression, which Andrew Ng points out is actually a type of classification. It's only called regression as an historical inheritance of name.

As for specifics: We got into the sigmoid function, logistic function, decision boundary (choosing a convex cost function), as well as applying gradient descent to the learning module. We were also introduce to some optimization algorithms other than gradient descent: conjugate gradient, BFGS, L-BFGS. I will say I'm disappointed the explanation of these other optimization algorithms was outside the scope of this mooc, but at least they were introduced. As far as criticisms of a course go, that's not too bad.

We also looked at multiclass classification, which is to say: One-vs-all (compare different classifiers and choose the most likely one). We ended with regularization, which is the smoothening of overfitting (vs feature reduction), where we penalize features. This seems like a hack, again I come from a pure math background and we'd never do something like this. I admit I found it pretty innovative and interesting anyway.

As for personal thoughts this week: Economics was my minor in my undergraduate degree, and although I didn't end up taking any econometrics course, I did hear about them. Now with machine learning—emphasis on "cost" functions (which we are mentioned in economics)—I'm starting to piece together what I would've learned if I had taken econometrics, and a lot of other economics courses are starting to make more sense too. So that's cool.

There's always more to learn, in so many different ways. Hooray!

Week 4 - Neural Networks: Representation

July 5th, 2014

This week we again change direction, moving to neural networks. It's a big topic so most of this week has been introduction actually. We started with the non-linear hypotheses. Tied in with that we stepped back from machine learning directly and discussed the one learning algorithm in neuroscience. Several examples were given in how the brain can adapt to different input signals. I especially liked the example where people learned to see with their tongues. That was cool!

We started into forward propagation, but just as an overview of neural nets in general. I have to admit I was very excited to reach this part of the course. The name neural networks just sounds cool and sci-fi. I'm disappointed to learn we don't get into the mechanics of it until next week. So sad, so impatient. Waiting is the worst!

Week 5 - Neural Networks: Learning

July 12th, 2014

This week we continued with neural networks, but got into the actual math of it. As it turns out the main idea is that it is an extension of logistic regression, which makes sense as it is a classifier. The complication comes from backpropagation, with subtleties of potential bias which are solved by random initialization.

I'm very happy to have learned this. I feel like if I can do this much, maybe I really can do machine learning in general.

Week 6 - Diagnostics (Best Practices)

July 19th, 2014

This week was actually a really big help. I didn't think it would be at first, as we weren't moving on to new material, it felt like we were losing traction. Instead, we consolidated everything we've learned so far by looking at best practices. In particular we were introduced to the ideas of diagnostics, test sets, bias, variance, precision/recall. This is more along the lines of statistics learning to distinguish things like false positives. What stood out the most was the strategy taken in using a given data set for training, summarized as:

  1. 70% training set
  2. 30% test set

When we are selecting from several hypothesis models, we extend this strategy as follows:

  1. 60% training set
  2. 20% cross validation set
  3. 20% test set

What this means is you take 60% of your data set (randomized), train until your error is sufficiently small. As you have more than one hypthesis model, you train each, and then use the cross-validation subset 30% to test which model still has a sufficiently small model. What's more, for a given model, the information provided by looking at both can help you determine if you're underfitting (bias) or overfitting (variance). You're underfitting if both your training and cross validation errors are large, and you're overfitting if your training error is small but your cross-validation error is large. Finally, once you are satisfied with the right model, you can see if the test set error is also satisfactory.

As for the remainder of this week, we looked at some real world applications. I'll add our instructor just told us regarding spam filters that researchers actually try to honeypot fake addresses as an easy way of collecting real world data for better training sets, lol. That's cool.

Week 7 - Support Vector Machines

July 26th, 2014

This week we discussed support vector machines with gaussian kernels. Think of it as an alternative learning technique to everything we've learned so far in supervised learning.

Upon first glance I don't have a full appreciation for SVMs I admit. To be fair we've spent far more time with linear and logistic regression, and then we just rushed through this new tool in a week. My guess is the point was to introduce us to the fact that alternatives do exist, that if we are to really master machine learning we need to know more about the landscape of tools available to us. Also, alternatives exist for a reason, which usually means there are optimization trade offs when it comes to the contexts these algorithms are deployed in.

I will have to let this week sink in more I think. In any case, we are finally on to unsupervised machine learning this coming week. Clustering in particular sounds awesome! I am very impatient for next week already...

Week 8 - Clustering

August 2nd, 2014

Hooray! It's the start of unsupervised learning! I should be catching up on my other moocs before continuing with this one, but we've finally reached clustering! How can I not?

This week blew my mind! We learned about clustering using K-means, as well as dimensionality reduction with principal component analysis (PCA).

K-Means! It's like the most straightforward intuitive clustering algorithm. If someone asked me to come up with a clustering algorithm without knowing any beforehand, it's the one I would've come up with. It's what I expected, but it's still cool. And worth going over of course as what's being taught here is polished with best practices.

I absolutely loved learning about dimensionality reduction! It was explained both intuitively and in detail. What's more, I had known about this technique while exploring biology and genetics, but I didn't know the name or the math behind it. Now I do, and it's less complicated than they made it out to be. Also, or lastly, it even shifted my perspective. Years ago when I had been introduced to linear regression through things like least squares approximation, I had considered how graphically the vertical projections weren't the best approximation, so I had wondered what was. Not only does PCA answer this question, it gives me a new appreciation for linear regression as well.

Week 9 - Anomaly Detection

August 9th, 2014

We can summarize this week as learning about anomaly detection, using the gaussian distribution (and its multivariate version). As well we went into recommender systems, with content-based recommendations, and into collaborative filtering recommender systems.

I hadn't realized anomaly detection was a type of unsupervised machine learning. Very useful for security :) I took it for granted because it's obvious in a lot of ways, although this is the first time I've been exposed to multivariate gaussian distributions. It's the sort of thing I think I'd learn better with use, but at least for now I know the name and know where to review the basics if I ever need to delve deeper.

Recommender systems I found really interesting. I recommend them! Ahah, so that was my bad joke for the day. Actually my favorite joke so far is: This week in my machine learning mooc I learned that behind every recommender system secretly is a human who just recommends stuff to people :)

Jokes aside, I'm actually grateful and glad to have learned about recommender systems. I have heard about Netflix holding a machine learning competition in regards to improving their recommender systems, so it's nice to see what kind of math is actually involved.

Week 10 - Stochastic Learning

August 16th, 2014

Wow. So it's the final set of videos. Has it been 10 weeks already? Feels strange for it to be winding down.

I would summarize this week as: Big Data Learning. It's for when the datasets are too large for our existing algorithms. The learning algorithms we equip our learning models with don't scale so well with large matrices it seems. In response to this we took more heuristic approaches, notable stochastic gradient descent, parallel batch gradient descent (using map reduce). I liked using map reduce actually, because of the functional programming I've been studying lately. Aside from its interesting theoretical considerations, I've always wondered where in industry it's applied.

I really appreciated the diagnostics for this week, where we learned about ceiling analysis (pipeline bottlenecks). I've said it before, but coming from a pure math background we don't spend much time with the engineer's way of thinking, such things as diagnostics.

Okay, so if anything disappointed me this week, it's the use of sliding windows. Like, sliding windows? Really? Is that the best you can do? Ahah. Anyway, I'll end with hopefully a more encouraging machine learning thought:

This gets more into my own personal line of research (and specialized use of language), but if semantics as objects are exactly media space constructs as modularization templates (partial specifications of data structures more or less), then does the brain learn free-form by first trying to mimic such non-deterministic media constructs by what it sees around it ? (local environment) To which it would then apply natural selection to see which constructs are reusable and which aren't?

Also: What if we use machine learning to structure comment sections. Then when looking at them people can choose filters and look at the types of comments they're interested in? Should I start my own Indigenous newspaper with innovative media spaces for people to interact?

Well, that's it! I hope you enjoyed my review of this course. Thanks!