Machine Learning

Mapping out the technology industry of today requires understanding the trajectory of machine learning. Machine learning is a hot topic, garnering billions of dollars in investments from tech giants Google, Facebook, and Cisco with its recent 1.4 billion dollar investment into the Internet of Things; however, Machine Learning Consultant Charles Martin suggests the fascination with machine learning isn’t as abrupt as it appears.

About 10 years ago, Martin invented a machine learning SEO (search engine optimization) algorithm for eHow that achieved the first one billion dollar IPO since Google. Demand Media, which launched eHow, held a patent allowing them to predict what people wanted to search for on Google by analyzing the data of the questions that were asked.

“There’s usually one dominant one [answer] that people are looking for, but they enter it into Google in different ways and so getting search relevance correct is really hard,” Martin explained. eHow also operated with a recommender system in place, a second piece of machine learning that allowed it to determine the intent of a user’s search. “We were able to essentially reverse engineer search and figure out what exactly people were searching for. If you know what people are searching for, you can front-run it, create content for it, and run ads.”

These search engines and recommender systems have huge implications for businesses in every industry, as they enable companies to quickly increase site traffic, increase revenue, and make more accurate recommendations for customers. Tools like the search engine function by including a feedback loop that enables it to continue learning beyond what it has been programmed to learn. For example, if Google provides four answers to a question asked by a user, but the fourth site provided is never clicked, the engine will adapt so that this site will no longer be provided as a potential option.

“When you’re using a mobile device or if you’re using it from your voice like voice search, or Siri, you really don’t have the ability to enter in complex queries. So you need to be able to enter in some sort of information, it needs to learn something about what you’ve done, it needs to learn about what other people who are similar to you have done, and it needs to learn when to provide information to you that’s personalized and when to provide a variety of information so it can collect feedback for the users,” Martin said. “So this is a very complicated problem; that’s where machine learning is not statistics. You really have to deal with these feedback loops and these biases that come up.”

Martin has worked with big media companies, shopping retailers, and dating sites that all want to cash in on the 30 percent revenue increase that a good recommender system can provide. But even without an increase in revenue, machine learning enabled search engines are becoming an expectation of the online user experience.

“I would even say that the consumer experience has become so good that businesses are now trying to replicate it, either internally or for their customers,“ Martin said. “People are seeing this occur in their every-day activity, whether they’re using Siri on their phone, whether they’re using Facebook, and they’re just recognizing: ‘why can’t we do this as a company.’ ”

Recent releases of AI open-source software libraries, such as Google’s TensorFlow, make machine learning a much more accessible reality. Other software developments in Python, R, and Hadoop allow machine learning consultants like Martin to approach a client and have them enter in their data, which generates a prototype as soon as six weeks later; 10 years ago, Martin would have had to code everything from scratch. Just five years ago, Martin describes working at Aardvark, which created a natural learning processor that was then acquired by Google for $50 million, and custom coding C++ on top of Ruby on Rails.

“Now we’re seeing the development of open source machine learning tools, even simple tools, but they’re becoming more and more accessible into the enterprise so that you can build products,” Martin said. “I wouldn’t say they work end-to-end; it’s much more where software was in 1995.”

Machine learning has a long way to go before it is easily within most companies’ reach. The software that machine learning is built upon is still fairly green, making integration into companies and creating scalable products difficult. Further, the technology has not yet developed user-friendly elements that support long-term changes, roll-backs, or updates.

“The problem is that machine learning is fundamentally different in how you build systems because you’re not building from architecture,” Martin explained. “For someone like me to come in and be able to do work, the difference between it taking me three years, six months, three weeks or even an hour to do something is really the infrastructure support I have on the tooling.”

Until machine learning software becomes standardized, it will largely remain in the hands of big companies with highly-specialized talent and boutique consultancy firms, which are high-cost and low-risk.

“The world is changing, the intelligence inside a company is becoming more and more important and you have to have leadership,” Martin stated. “You have to have people in the organization that really understand this technology.”

While machine learning products are becoming staples of a modern industry, the lack of a standard software platform from which to build and the complexity of current technology has prevented any many players from rising amidst the tumult. High-touch consultants will corner the market, as the developments over the next five years attempt to establish a secondary market of consultants who can work off of a stable software platform.

“We’re going to see big changes coming, new technologies and restructuring of industries. There’s going to be a lot of disruption,” Martin projected. “We don’t know what it’s going to be.”

 

About the Author:

Dan Faggella is a graduate of UPENN’s Master of Applied Positive Psychology program, as well as a national martial arts champion. His work focuses heavily on emerging technology and startup businesses (TechEmergence.com), and the pressing issues and opportunities with augmenting consciousness. His articles and interviews with philosophers / experts can be found at SentientPotential.com

During the last fifteen years, a strange parallel economy has covertly developed to the point where it envelops almost all internet users, including you. No money changes hands in this immense network, but it produces enormous transactional benefits nonetheless.

In this economy, you labor daily as a trainer, teaching software robots how to perform tasks. In return, the bots then take over much of those tasks for you. You trade your daily labor in exchange for the value produced by the work of your powerful and ubiquitous robot apprentices.

The most successful products of this epoch – applications like Gmail, YouTube, Amazon’s store, Facebook, Google Maps, and Spotify – have learned a great deal about their users’ likes, dislikes, and similarities. Applying Machine Learning technology, they use that data to better present what consumers will want to see and hide what they do not.

These applications did not get so powerful through traditional programming, but through self-learning. Instead of a team of coders defining the steps to be followed using a computing language, Machine Learning starts with a set of observable data – such as items that users bought – and learns to infer the patterns within the data.

So how does one get the data set to train a Machine Learning system?

Sometimes a Data Scientist can mine the data out of existing records collected for some other purpose. This is one of the reasons that companies now like to collect every bit of data they can about you. It is much easier to re-use existing data than to collect it from scratch. But frequently to obtain a well-organized set of questions and users’ responses, the team must gather new data.

One way to collect orderly data is to pay humans to answer questions. Using a system like the Amazon Mechanical Turk, you can define a question and get many thousands of answers from workers, paying a few cents for an answer. This approach is often used for problems that people solve well, such as image recognition. A Mechanical Turk worker might, for example, classify images as landscapes or indoor photos, or draw a circle around faces in photos. This well-organized set of images and identified areas is ideal for accelerating the training of a Machine Learning program.

It is even cheaper to get the humans to answer the questions for no money at all, by providing them some utility value. This is where you come in.

You have probably used a CAPTCHA, an application that requires that you identify a number or a small piece of text in order to prove that you are human. In doing so, you are doing useful work for somebody. Google initially trained its street number recognizers for Google Street View on data sets it built by putting photos of doorway areas into its CAPTCHA system.

Another way to get free data sets is by turning data collection into a game. Development teams are great at those kinds of problems, as software and UX designers often love to make (and play) games. They can quickly turn a Data Science problem into a slick quiz with a polished user experience.

You have probably taken a quiz like this on Facebook or another website. Applications that allow users to learn about themselves – or purport to do so – are very popular. However, the real goal of the quiz may have nothing to do with its ostensible purpose. For example, a quiz that purports to give you personality insight might well be measuring the subtle difference in response speed when you reply on questions containing one group of words vs. another. It might also correlate that information with metadata you allow it to access in your social profile, like your gender, age, or political affiliation. When companies speak about “converting clicks to value,” this is what they are talking about.

In a stealthy economic transition, most of us have acquired a new secondary role as a machine trainer. But while this work produces useful value, you can’t use it to pay for groceries. And here we come to the cusp of a looming economic crisis.

Up until now, the internet economy of smart agents has been subsidized by the traditional economy, in which employers pay workers paychecks in a structured manner. But as the role of machine training grows in importance, automation technologies are, through efficiencies, eliminating jobs. Automation creates new “traditional” wage-paying jobs, but not as many as it eliminates.

In the last such transition – the industrial revolution – farmers moved into factory jobs. Now, the industrial workers are moving stealthily into knowledge jobs such as machine training. But unlike during the industrial revolution, these jobs are not a direct replacement for the old ones. At present, working as a machine trainer is a second, usually unpaid job. It does provide value, such as more efficient email processing and better autonomous agents, but you can’t feed your family by helping to train a recommender system.

Discussions of the AI revolution often focus on the permanent elimination of entire job classes, such as drivers being replaced by self-driving cars. Proponents of the “abundance” view believe that new jobs will arise in previously unforeseen areas to replace the old ones. The problem we have at this moment is that the new jobs, like machine trainer, are arriving – but they are not replacing the earnings of the old ones.

As the internet economy continues to subsume the “real” economy, this automation crisis is coming to a head. In a follow-on article, we will discuss the evolving AI Economy and directions in which it might develop.

 

About the Author:

David Rostcheck is a consulting data scientist helping companies tackle challenging problems and develop advanced technology. He can be reached at drostcheck [at] leopardllc.com.