This is the second installment of a 3-part series on machine learning. If you want to read the first part, the link is below. The outline of the 3 installments is:

1. Machine Learning Introduction
2. Various implementations of machine learning
3. Impact to business of machine learning computers

The last installment will be published in the next few weeks. I will update this article with a link to it.

In the last post, we outlined the basic underpinnings of Machine Learning in the context of Big Data. Here we shall examine some implementations of Machine Learning, a discipline that stemmed from research into Artificial Intelligence (AI), especially concerning algorithms that learn from the data autonomously. In this post, let’s look at some AI machines programmed to perform machine learning, and their fascinating results and occasionally surprising side-effects. Although the focus is on machine learning, the mechanics themselves belong to a broader context of AI.

Deep Blue

On February 10, 1996, a computer called Deep Blue beat a human at a game of chess. What was historic about this was the fact that the human was the reigning world chess champion, Gary Kasparov. Although, eventually Kasparov went on the beat the computer, it was the beginning of the end of human dominance in this extremely strategy-based board game.

Then in May 1997, the same computer after several upgrades played Kasparov again. This time the computer beat Kasparov 2-1 in a 6-game tournament. If there was any debate that a computer cannot beat the best human in a game of chess, this put an end to it. Or did it?

Deep blue is just as fascinating in what it is not, as much as in what it is. While, technically, Deep Blue’s deep knowledge of chess came from learning prior games, its super-power truly lies in its massive processing capacity. Deep Blue is a massively parallel system (technical term for a lot of processors, running in parallel), with 30 processors plus 480 chess chips. In simple English, it’s a beast. It can process nearly 300 million positions/moves per second.

That raises a question: Is Deep Blue really a learning machine? Firstly, according to the rules of the tournament, Deep Blue’s developers were allowed to alter/upgrade the software between games. This means that Deep Blue’s engineers were learning the game and teaching it to the computer, rather than the computer doing it for itself. But IBM chalks it up to upgrading the system’s learning mechanism. Fair enough.

Compare deep blue’s functioning to how a human thinks. A human gets better and better at playing chess as they play more and against better opponents. Even if IBM engineers were right in that their inter-game intervention was only upgrading, the question still is: Does deep blue understand the game of chess and its strategies the same way a human learns. I am not referring to the wetware [http://en.wikipedia.org/wiki/Wetware_%28brain%29] vs hardware/software argument. Humans obviously are not processing more than 200 million moves per second: far from it. Deep blue, on the other hand, does have this capability and uses it quite effectively. This clearly points to the difference in human insight and the brute force a computer. Deep Blue may or may not have had the same insight Kasparov had, but at the end of the day, Deep Blue won.

 

Source: Encyclopaedia Brittanica

Watson

For our second example of an artificially intelligent computer, we go back to IBM.

In 2011, Watson played Jeopardy! (a popular TV quiz show) with Brad Rutter and Ken Jennings, former winners of the game show. Brad Rutter never lost Jeopardy to a human opponent. Ken Jennings, of course, holds the record for the longest winning streak in Jeopady history, 73 in total. Together, these two men had walked away with more than $5 million dollars, when they played against Watson. In a 3 day tournament, Rutter and Jennings went head-to-head against Watson, only to see Watson trounce them in the end.

While crunching numbers is something computers are very good at, language processing has been a very challenging problem in computer science. IBM set out to take this challenge up, when they designed and built a computer named Watson. More specifically, Watson was exclusively designed to answer trivia questions asked in natural language.

Watson’s first task is to extract meaning from a vast database of documents, ranging from encyclopedias, news articles and literary works – millions of them. Keep in mind that this is not the same as Google’s search engine indexing the Internet. Document searching (and page ranking) is to take a keyword as query and return a list of documents that are relevant to the query. While a search engine knows the words in the indexed documents, Watson is supposed to understand the contents of the documents.

Understanding Human language is a difficult task for a computer, which thrives in a world of discrete numerical constructs. Real language is full of implicit and ambiguous references, whose meaning can only be extracted within the context of the conversation.  Take for instance this phrase: “I like tea, but Java is what gets me going.” An average English speaker has no problem understanding that Java is a reference to coffee. We will be hard-pressed to find people who would mistake this java for either the computer programming language or one of the main islands in Indonesia. For a computer, on the other hand, this is a very difficult task – determining meaning by context. A compute trying to understand a phrase like “Here I am sitting in my hotel room in Java, sipping on my Java, and coding away in Java” would probably lead to a blue-screen-of-death [http://en.wikipedia.org/wiki/Blue_Screen_of_Death].

Let’s take look at how Watson discovers the right answer to a given question. Just like Deep Blue, Watson is a massively parallel computer: one built with a cluster of 90 IBM Power 750 servers, each one powered by an 8-core processor. All added up, Watson had 2,880 POWER7 processor cores and 16 tBs of RAM. Watson ran on SUSE Linux Enterprise Server OS using Apache Hadoop’s distributed computing framework.

On this infrastructure, Watson ran IBM’s DeepQA software, a software designed to run on massively parallel infrastructure. The mechanism with which this computer turned trivia questions is as impressive as the hardware and software it ran on.

Watson firstly, accumulated a vast amount of knowledge by storing and indexing millions of documents. When Watson received a question, it would first perform a task called “Question decomposition,” where it would break down and parse the question/clue into keywords and phrasal fragments. It would then run several natural language analysis algorithms on the decomposed question phrases, all in parallel. The more algorithms that return with the same answer, the more confident Watson becomes of that answer.  These two steps were “hypothesis generation” and “evidence scoring.” Based on this scoring, Watson would then rank the hypotheses. Together, score and rank, helped Watson determines the “confidence” it had in the answer. When the confidence is high enough, Watson would buzz the answer.

Jennings and several experts in the field of neuroscience and artificial intelligence are convinced that this is very similar to how a human brain works, under the same circumstances.

Answering Trivia might sound trivial, but the implications of Watson are far-reaching. From then on, Watson went on to do some serious work around the world. You can read more about it here

http://www.forbes.com/sites/bruceupbin/2013/02/08/ibms-watson-gets-its-first-piece-of-business-in-healthcare/

The final scores of the Jeopardy! game was:

• Rutter: $21,600
• Jennings: $24,000
• Watson: $77,147

Ken Jennings, in his final jeopardy question, wrote underneath his answer: “I for one welcome our new computer overloads.”

 

Source: New York Times

 

Other examples

While doing research for this article, I came across several fascinating examples of machine learning and artificial intelligence. In the interest of keeping this article short, I have not included them. If you’re interested in hearing more fascinating stories like these, please leave a comment at the end of this post. If there is enough interest, I will put them together in a separate post.

 

 

We are seeing a technological revolution, one that began at the turn of the century. In the past, though, technology replaced manual labor and work that involved a high degree of calculation. But recently we are seeing great progress made in domains which were previously thought to be exclusively human domain – thinking, learning, and strategizing. Increasingly this new computing paradigm seems to be encroaching and infringing upon our prized abilities. From the 19^th century till the end of the 20^th century we saw robotics replacing jobs that required manual labor. But the current trend is of replacing mid- and high-skilled workers. The computer that beat us at chess and trivia pave a path for the computers that will replace business strategists, scientists, doctors. Creativity will soon be replaced by algorithms for creativity.

There are social implications that we need to be concerned about, apart from the economic benefits of these. In the next post (the 3^rd installment of this series), we will examine exactly these :  the economic and social implications of the new wave of learning machines.

This is the first installment of a 3-part series on machine learning. The outline of the 3 installments is:

1. Machine Learning Introduction
2. Various implementations of machine learning
3. Impact to business of machine learning computers

The remaining installments will be published in the next few weeks. I will update this article with links to them.

Can you build a computer that can think for you, and may be run your business? Computers have been an integral part of any business enterprise for a long time now, insomuch that everything we do in business has computers’ finger prints all over. But the primary function of computers has been that of aiding the humans in their running of the business, not actually running it for us. But a brave new generation of computers is on its way that plan on turning this paradigm on its head.

Imagine computers that think, that understand how your business is run, recognize what works and what does not, correct issues as they go along, and most importantly, ones that do this with virtually no aid from you. These computers almost do not exist today, but will be business-as-usual someday.  What does it entail to build such computers? How much of this is hype and how much reality? What are the classes of problems these computers are expected to solve? A branch of study in Artificial Intelligence (AI) called Machine Learning has been trying to answer these questions.

What is machine learning?

A typical computer program is a series of instructions executed on a set of data. The code is supposed to read the data, manipulate and transform it. Machine learning, on the other hand is not about transforming data, instead about recognizing patterns in the data, discovering deep insights in the structure of the data, all on its own. In fact, it is even more than that: It is about a computer that understands data and gets smarter and smarter.

Machine leaning is about creating algorithms that start with a blank slate, and builds up its knowledge as they process and analyze data. The bigger the data, the smarter they get. The do so in a variety of ways.

Take for instance, a hypothetical chess-playing learning machine. This machine first learns the basic rules of chess by watching people playing chess – rook moves horizontally and vertically, bishop moves diagonally, etc.  Then the computer learns the goal of the game – to win. To extend this example further, the computer learns how to achieve the goal. It learns from the games strategies. While the computer is never coded to perform any specific list of chess plays, it is designed to learn them from the input data – studying previous games and reassessing the games it plays against its opponents.

Hypoethetically, one copy of this software is given amateur chess player, and another one to, say, Gary Kasparov (former world chess champion), and both are allowed to play solely against their masters. After a certain period of time, if these two copies are made to go head to head, the latter copy will trounce the former. This is because the latter learned from a superior dataset, one that was accumulated playing against Kasparov.

To Teach or not to Teach

Broadly speaking, there are two categories of learning algorithms: Supervised and Unsupervised. If the input data also known as training-set is clearly labeled – the computer knows what it is looking at – then it is supervised learning. However, if your data is a big clutter of bits, and the computer starts off without knowing what it is looking at, but it learns as it goes along, that would be unsupervised.

For example, a face-recognition learning system processes several images of faces and non-faces from a training-set that is clearly labeled – faces in each image are marked and labeled. During the learning process, the computer isolates the unique features of faces that are not to be found on non-faces. Subsequently, the computer is now equipped with the “knowledge” of how to distinguish between faces and non-faces. The more images of faces and non-faces it processes, the stronger its knowledge is.

On the other hand, in the construction of a computer that classifies human emotions. The training-set contains no labels for facial expressions and how they correspond to human emotions. The system is fed series of video snippets with human faces containing various emotions and their subsequent actions. Your computer is supposed to see the subtle changes in the facial expression and group them into various categories and associate them with the subtle difference in the subsequent actions of these actors.

Given that these two examples of very simple, there may be an overlap in their classification. But the essential idea is this: Supervised is when the computer clearly starts off knowing what it needs to do, and goes on to becoming really good at doing that. On the other hand, in unsupervised learning, the computer typically has no idea what it is looking at or what it’s supposed to find, and then goes on to discover hidden pattern and deep structures in data.

Although different functionalities dictate which learning is more suitable for the specific purposes, when it comes to the context of big data, unsupervised learning algorithms are expected to be heavily used in the near future. Unsupervised learning is well-suited in systems of data which contains deep hierarchical and/or causal relationships between observations and/or latent variables.

Learning algorithms may not clearly fall along this dichotomy. Most algorithms have a combination of the two. Within the same system, some aspects of learning may be supervised, while others may be unsupervised.

Business Intelligence vs. Learning Machines

There are some striking similarities between the data mining components of Business Intelligence suites that are used for pattern recognition, and the actual machine learning implementations. While BI data mining is a set of tools and techniques used by humans to aid in pattern recognition and eventually make better decisions, machine learning is performed primarily to the advantage of the machines themselves — in order to perform better by reorganizing itself. There is a significant overlap in the various techniques used in these two domains of data analysis.

So how does a computer learn?

Another way of classifying learning machines is the expected output. Let us see the kinds of output we expect from these computers.

Regression Analysis: The excepted output of this computer is to find hidden relationships between two or more variables. For e.g., is there any relationship between the weather outside and my sales data.

Classification: The expected output on this learning system is to take a large chunk of data and classify them according to preset categories. For e.g., what criteria can I use to categorize my employees as top performers, team players and laggers.

Cluster Analysis: Similar to classification, cluster analysis takes a series of similar objects and classifies them. The difference between this and the previous method is that clustering has not preset categories. Objects scatter over the data-space, and computer identifies clusters in them. For e.g., identifying customer clusters for targeted marketing

Computers can learn to run your business for you

While the science of machine learning has been flourishing in the scientific and mathematical circles, the business community has been slow to adopt the trends. With the exception of financial institutions and some sales and marketing campaigns, thinking and learning machines have not made much headway. With the ubiquity and popularity of big data infrastructure, such as Hadoop, it is easy to see that the near future hold exciting trends in machine learning in the business.

Big data enables businesses to adopt machine learning technologies. The potential for machine learning as a field of study and its business applications is unlimited. There are so many problems we know, but do not know how to solve. There is a bigger list of problems that we do not even know exists, let alone know how to solve them. If we ever hope to discover these problems and solve them effectively, learning machines are our fiends.

 

Big data Analytics is the new buzz in town. It is every big data analysts’ dream to analyze data, big or small, and get insights into things that were previously unseen by others. And now we have the tools — software and hardware — to analyze large amounts of data in a way that leads to meaningful and sometimes actionable analytics.

However, data is only as valuable as the analysis one performs on it, and Analytics is only as insightful as the person doing the analysis. The main function of a Data Analyst is to derive actionable analytics from the swath of data available. Sometimes, the same data can be analyzed by different analysts reaching different conclusions – quite possibly contradictory conclusions. There are many reasons that might lead to this contradiction. I thought it would be a good idea to list some cautions on any exercise of data analysis, big or small.

 

Even a broken clock is right twice a day

As stated earlier, the analytics derived from data is only as insightful as the analyst performing it. As an analyst looking for patterns in data, you will encounter several apparent hits. While it is exhilarating to see patterns emerge from data analysis, it is just as important to look further to see if the emergent pattern is indeed a real-world pattern or just a statistical anomaly.

More often than not, emergent patterns are just that — patterns that appear in small segments of a larger data set. If patterns appear sporadically or sparsely, it is most likely a false positive. Large data sets inherently contain statistical anomalies. In fact, the bigger the data, the higher the probability that improbable (and consequently inaccurate) patterns are hidden in smaller subsets of the larger dataset.

Even a broken clock is right twice a day, and if you happen to check the time at those right times, you will wind up thinking that the clock was working.

 

Climate is not weather

I am always mildly amused by people who say things like “It’s so cold today; so much for global warming”. While I get that it is just a joke, the undertone in the statement clearly points to a lack of understanding in the difference between weather and climate, or more generally, perceiving apparent patterns in really random or unrelated events.

Science spokesperson and astrophysicist, Neil deGrasse Tyson drops some knowledge in this video below, a segment from his TV Show, Cosmos.

Courtesy: National Geographic

Keep your eye on the man, not the dog!

The point to take home is that, at smaller scales, the data might look either haphazard or manifest a specific pattern. However, applying a broader context may reveal another pattern, one that is quite different.

In the chart below (courtesy NASA Goddard Institute for Space Studies), the lighter line represents annual global temperatures, and the darker thicker line represent the 5 year running mean. If you focus on the thinner line, you will notice that patterns are haphazard. There are three years between 2000 and 2010 where the temperature is actually cooler than 1998. However, the thicker line which averages the 5 years around the data point shows a clear rise in temperature over the past couple of decades. Yet again, the thicker line also shows the temperatures flattening out in the last decade. It’s easy to see why one would just observe the last decade and claim the global temperatures have flattened. But upon closer inspection in a broader context, there have been similarly sized flattening happening in the 80s and 90s. The bigger picture from 1960 to 2010 is one of clear and undeniable growth.

Source: NASA Goddard Institute for Space Studies (http://www.giss.nasa.gov/)

Applying a broader context is an essential part of data analysis. Keep your eye on the man, not the dog!

 

Correlation does not imply causation

This phrase is very often used in statistics, almost to the point that it is common sense in the data analysis community. Despite that, it is often an overlooked occurrence.

A correlation between two variables, no matter how strong they appear, need not necessarily mean that one has a causal relationship with the other. In fact, it does not necessarily even mean that one has any relationship to the other.

As shown in this comical chart below, there seems to be correlation between Internet Explorer usage and number of murders in the U.S. But it does not make any sense in the real world.

Courtesy: Geek.com (http://www.geek.com/microsoft/does-internet-explorers-falling-market-share-mirror-the-drop-in-us-homicides-1537095/)

I am not sure if the data behind this chart is accurate, and I have no reason to believe one way or the other. However, I have no reason to believe that these two phenomena are even related, let alone one causes the other. Surely, it’s absurd to state that as people start using IE less they murder less, or the reverse, that as murder rates fall so does IE usage. However, the data seems to support these assertions.

While this example is patently absurd, there are many cases, where it might seem likely that there is a relationship. It is best to keep this catchphrase in the back of your minds: Correlation does not imply causation.

An extreme version of this catchphrase goes to the extent of saying that Causation can never be determined by analytical means, and should be left to the realm of philosophy. Correlation is good enough, as long as the corresponding hypothesis can be validated by its power of predictability. It essentially means that if a correlation is noted, and it is repeatable and predictable, then it is not required to prove the causality aspect of the correlation.

While it is important to understand that correlation does not imply causation, a strong correlation is sufficient to make effective business decisions.  True causality is certainly a curiosity, but one that is quite unnecessary in the context of business intelligence.

A picture is worth a thousand numbers, or a couple of terabytes

While computers are very powerful number crunchers, we humans are far superior visual analyzers, at least as yet. Computers can analyze visual data, but they do so, by and large, by converting the graphical data into numerical equivalents.

We certainly should use the power of computational data analysis. At times though, there is nothing better than the good old staring-at-a-chart data analysis. This is not because of any inherent problem with computational methods. Instead, it is because of the simple fact that computers do what the data analyst asks it to do. The problem is when the data analyst does not know what to ask for.

Computational (statistical) tools like Standard Deviation and Linear Regression, while powerful in their own right, are not suitable to spot blips or spikes in otherwise smooth data, unless one knows to look for it. Plot it on a chart, and the human eye will catch it in a blink. Again, I am not arguing that looking at charts is the best way of looking for blips. If you know you’re looking for a blip, it is really easy to have the computer look for it for you. But, in case you didn’t know what irregularities exist in your data, it’s better to plot it on a chart, and let the human eye/brain look for them.

There are many tools that let you plot your data. Microsoft Excel is a good place to start.

 

Follow the data; don’t make the data follow you

Almost every exercise in data analysis starts with a hypothesis. Although it would be ideal to perform research with a blank slate and come to conclusions after the analysis, it is almost inevitable that you will start with a hypothesis. In fact, in statistical inference, one starts with a null hypothesis.

The problem is in setting up scientific experiments or performing data analysis with the intention of proving a hypothesis. This is where confirmation bias kicks in. The level of attachment you have to your hypothesis determines how many data points you observe that prove you’re right. This is reminiscent of an adage President Obama used during his 2008 Presidential campaign: “If all you have is a hammer, every problem looks like a nail.”

There is a saying in the scientific community, and I paraphrase: if you have a hypothesis and you want to know if it’s true or false, then do not set up your experiments to prove that it is true. Instead try to prove that it is false. And if you can’t, then it may be true.

Also, always remember: You can never prove that a non-trivial hypothesis is true. You can either prove that it is false or that it is not false.