This is the third and last installment of a 3-part series on machine learning. If you want to read the first two parts, follow the links 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

In the last two posts, we explored the idea of Machine Learning and its application. This post will be about how learning machines, and computers in general, impact and influence businesses and vice versa.

Humans and machine have had symbiotic relationships. While humans shape machines, the irony is that machines have shaped humans just as much as, if not more. I don’t mean it literally, but machines may be seen to have influenced human progress in the political, social and most importantly economic systems to their current state of the art. Science and technology has been the center-stage of philosophical discourse in the Western hemisphere for more than three centuries. And computers have been the engine behind the economic success of the past few decades.

While it may be argued that the era of machines began either with the start of the Industrial age, or later at the Machine age, in the context of this post it really didn’t begin until the 1890 US Census. Herman Hollerith’s punched cards and the tabulating machine essentially cut down the census time from 8 years to 1 year.

“[Hollerith’s] system made it possible for one Census Bureau employee to compute each day the data on thousands of people, keypunching information that had been captured by tens of thousands of census takers.”

Library of Congress

http://memory.loc.gov/cgi-bin/query/r?ammem/mcc:@field(DOCID+@lit(mcc/023))

If a rudimentary computer by today’s standards can shave off 7 years, then imagine the time-savings today’s computers can provide. Of course, this tremendous positive impact is not without its negatives. This automated system measurably replaced thousands of Census Bureau employees overnight.

Since then, there has been a steady employment of computing machines replacing humans. After the end of the Second World War, ENIAC and UNIVAC paved the path for electronic computers to dominate business computation.

The business of running business has been transformed by newer and more powerful electronic computers. Again, as in the US Census case, while the economic sector has reaped the benefits the labor market saw the most negative disruption in this new world.

In the first wave of these computers, the world saw the low-skilled jobs vanish. Jobs that required repetitive tasks, especially ones that involve computations on large sets of numeric data, were slowly transitioned to automation. People holding jobs that can be described as routine and codifiable were the first to see the door. While in the early part of the 20^th century, the job market was dominated by low-skilled workers, a constant decline in this class of jobs can be directly attributable to computers.

However, another trend that has caught the eye of the economists and policy makers is called the Polarization of the job market. Wikipedia describes it as “when middle-class jobs—requiring a moderate level of skills, like autoworkers’ jobs—appear to disappear relative to those at the bottom, requiring few skills, and those at the top, requiring greater skill levels.” As this chart depicts (by Harry Holzer and Robert Lerman [http://www.urban.org/UploadedPDF/411633_forgottenjobs.pdf]), while there has been a modest (and insignificant) growth in low-skilled jobs, there has been a constant decline in mid-level jobs – from 55% of all jobs in 1986 to 48% in 2006. What is notable is that the loss in mid-level skills is mostly offset by growth in high-skilled jobs.

http://www.urban.org/UploadedPDF/411633_forgottenjobs.pdf

It is clear that technology and automation has caused a good deal of increase in wealth and prosperity. However, economists such as Joseph Stiglitz and Thomas Picketty have eloquently and persuasively argued that the problem is that the gain in wealth is concentrated in a relatively small number of participants in the economic system. Drawing from the experiences from the past, a policy dedicated to public and private investment in education and specialized skills training to the impacted masses is advised.

Having stated that the threat to low-skilled and mid-skilled labor categories is imminent, it seems to me that there is complacency on the high-skilled labor category. The general consensus is that high-skilled jobs will continue its strong growth. While I don’t argue that this consensus is wrong, one should, however, employ caution in prophesies such as these. Computers are increasingly doing tasks that had once been deemed impossible to be automated. Let’s take a look at some examples, where smart machines are doing high-skilled jobs.

Robojournalism

A surprising amount of what we read in newspapers and journals are actually written by computers with no human aid. Robojournalism was brought to people’s attention by “Quakebot”, a program written by Ken Schwencke. Quakebot wrote the first news article about the Mar-2014 earthquake off the Pacific Coast. This was not the first time a computer was employed to write news posts. But since then, robojournalism has gained momentum. Soon after, Associated Press announced that “the majority of U.S. corporate earnings stories for our business news report will eventually be produced using automation technology.” AP and Yahoo use a program called Wordsmith, developed by Automated Insights.

A particularly interesting case study is that of a Chicago-based company called Narrative Science They developed a piece of software called “Quill”, and similar to the earlier examples, this is a Natural Language Generator (NLG) platform that can process data, mostly in numerical format, and convert it into perfectly written narratives to be consumed by humans. Narrative Science started off by commercializing a research project at Northwestern University. It was first adopted to be used by sports channels and websites to report headlines for baseball games. It did so just by looking at the numbers. Take a look at this text that Quill generated:

Tuesday was a great day for W. Roberts, as the junior pitcher threw a perfect game to carry Virginia to a 2-0 victory over George Washington at Davenport Field.

Twenty-seven Colonials came to the plate and the Virginia pitcher vanquished them all, pitching a perfect game. He struck out 10 batters while recording his momentous feat. Roberts got Ryan Thomas to ground out for the final out of the game.

http://deadspin.com/5787397/we-heard-from-the-robot-and-it-wrote-a-better-story-about-that-perfect-game

This is an excerpt straight from Quill – no human literary intervention. This is just by parsing the box score of the game.

Narrative Science has taken this to the market to have their software write narratives for financial reports, etc.

Automatic Statistician

Big Data has generated a renewed interest in data analysis, especially applying statistical concepts to data in order to derive insights and meaning from large swaths of raw data. People calling themselves data scientists are popping up everywhere. Their main job is to take a deep look at the data (perform statistical analysis and modeling on it) and identify patterns and insights in it. These patterns and insights are valuable in predictive analytics as well as operations research. Automatic Statistician is purportedly in the business of automating this process of discovery.

Automatic Statistician is the brainchild of Zoubin Ghahramani and his research team at the Department of Engineering – University of Cambridge. They set out to make a computer do what a Data Scientist is paid to do – make sense of data. Automatic Statistician gained recognition outside the academic circles when Google awarded the team the Google Focused Research Award. At the writing of this post, Automatic statistician is still in its early stages. Yet it has shown strong potential in applying statistical and machine learning methods on raw data, and automatically discovering trends in data; completely unsupervised.

Below is the screenshot of the automatic analysis and report by Automatic Statistician, when fed the raw data of unemployment figures

http://www.automaticstatistician.com/abcdoutput/11-unemployment.pdf

There are scores of interesting examples such as these. Machine Learning has been gaining momentum over the past decade. To repeat Ken Jennings’ sentiment: I, for one, welcome our new computer overlords.

Conclusion

Technological innovation has been reshaping the labor market for a long time now. One can date it back to the industrial revolution. For all the buzz the phrase “Big Data” has created in the recent past, I believe the advancement in AI, Robotics and Machine Learning, applied to big data is one such wave – one that will change the way we are used to do things. Many jobs from today will not exist in the near future, and many jobs in the near future are completely unknown right now. Phone operators didn’t exist for most of 18^th century and neither did rocket scientists until the early 20^th. Web-developers and computer network analysts didn’t exist for most of the 20^th century.  Big Data, AI and Machine Learning will lead to jobs that we just cannot imagine at this time.

Sure, this disruptive technology is going to negatively impact the labor market, but there is more to gain. Technology is a net job-creator. However, to mitigate the short term negative impact, a strong role from both the government and private sector is prescribed by policy experts and economists.

Computers in business are not just about making machines do the drudgery that we do not want to do. Today’s computers are much smarter – they can simulate thinking and reasoning that was previously thought of as a purely human endeavor. Today’s computers help us strategize, perform market analysis, build modes, and explore new opportunities. Tomorrow’s thinking machines will not just be helping us, but modeling these themselves, in an unsupervised way.

 

 

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.