Machine Learning and Our (Insatiable) Penchant for Making Things Smarter

We have been learning by experience for millennia and these days, with computer power growing more and more, we have the opportunity to teach our machines how to take cues from the environment and be able to “see”, “hear”, maybe even “read” for us.

Enter machine learning.

To and Fro a Definition of Machine Learning

There are as many definitions of machine learning as there are engineers busy with designing and training systems by feeding them specific data. Frequently, in research and business, machine learning definitions include terms like artificial intelligence, neural networks, supervised and unsupervised learning, pattern recognition, cognitive computing, etc. Somewhat aside from these specialized words lies a definition that does a great job of explaining machine learning without oversimplifying:

The effort to build machines better.

Ref. Machine Learning: Making Sense of a Messy World

Another definition, which provides the much-needed simplicity at the end of complexity is the one by Laura Tolosi, Ontotext’s enthusiastic data scientist who always searches to improve semantic technologies with the latest machine learning tools.

Machine learning, she explained, is the way we, humans, try to teach computers to reason and act more like we do. If we see a tiger, we would understand that this means a threat and further act upon this understanding. We will do this because, throughout our evolving as species, we have been developing skills to interpret the environment and take action based on the information we have received from our surroundings. For a machine to do the same (process and act upon any given data), we need to teach it with lots and lots of examples about the characteristic features of the things in the environment it operates within and further let it extract patterns from the data it was fed with.

The Penchant for Making Things Smarter and Smarter

Or Why We Want Machines to Learn?

The ultimate goal of the research and exploration related to machine learning is that one day smart systems will be augmenting or even automating tasks once only humans were capable of undertaking. And building machines that learn from example and from observing the environment they operate in has always been at the back of the minds of those who want to develop better machines.

At the beginning of the 1950s, a small robot made a big difference, impressing the cybernetics geeks. The little machine solved mazes and was called Shannon’s rat, named after Claude Shannon – the American mathematician, electrical engineer and cryptographer known as “the father of information theory”.

This rat wheeled out a cabinet with a five-by-five grid on its top panel. […] Under the hood lay an array of electrical relays, about seventy of them, interconnected, switching on and off to form the robot’s memory. […] When the machine was turned off, the relays essentially forgot everything they new, so that they […] started afresh, with no knowledge of the maze. […] The machine made each “decision” based on its previous “knowledge” […] according to a strategy Shannon had designed. It wandered about the space by trial and error, turning down blind alleys and bumping into walls. Finally, as they all watched, the rat found the goal, a bell rang, a light bulb flashed on, and the motors stopped. Then Shannon put the rat back at the starting point for a new run. This time it went directly to the goal without making any wrong turns or hitting any partitions. It had learned.

Cit. The Information, James Gleick, p. 249-250

More than half a century after Shannon’s rat (and an Atlas later), machine’s networks are close to neural and the “learning” happens seemingly at the speed of light, without any limitations. Add to that the increasing amount of data available for training and the ever-growing computer processing power. The result: teaching machines to reason based on what they have “learned” is poised to disrupt each and every industry and sector.

The Need for Making Texts Smarter and Smarter

Or Why We Want  To Teach Machines to Read

One of the disruptions already happening is in the field of text analysis.

As complex the matter of understanding and reading a text is, with the amount of digital texts across the Internet and on Intranets rapidly increasing, we need to find smarter ways to manage texts on a scale.

Today, high-quality text processing is unthinkable without machine learning. Certainly, our algorithms still have a hard time understanding texts the way we human readers do, but they have advanced to a point where we have a kind of cognitive reading machines. Click To Tweet These machines are able to help us classify texts, enrich them with semantic metadata and extract mentions of people, things, locations, events, etc.

The reason we would teach a machine learning algorithm to “read” is that its abilities can be further leveraged for all kinds of purposes: from semantic search for documents retrieval, through automated tagging and dynamic presentation, to automated topical clustering and questions answering.

Feed for Thought

“I am young and still learning,” will inform you Stanford’s chatbot Woebot – a therapy chatbot for depression and anxiety. Or as the chatbot describes itself: “your charming robot friend who is ready to listen, 24/7”.

Woebot is not the typical bot who, while interacting with you, is using machine learning algorithms to learn from the data it gathers (your age, occupation, preferences, the words you use, etc.). But the chatbot’s first words are important as they clearly show that any system’s training comes from the data it is fed with and the way it is designed to automatically process and manipulate these data.

That is to say, machines do not “learn” and “know” the way we do. As Rodney Brooks notes, in his clear and concise write up about machine learning:

Machine Learning⁠ is not magic. Neither AI programs, nor robots, wander around in the world ready to learn about whatever there is around them. Every successful application of ML is hard won by researchers or engineers carefully analyzing the problem that is at hand.

Cit. Rodney Brooks: [FoR&AI] Machine Learning Explained

As a matter of fact, Alan Turing once told Shannon:

I am not interested in developing a powerful brain. All I am after is just a mundane brain.

Cit. The Information, James Gleick, p. 205

Now, with the development of technologies that can perform tasks just like a “mundane brain” would do, machines are no longer only used for simple jobs such as finding the nearest location, assessing credit risk or getting the spam out of your inbox. Machine learning algorithms are now part of many and complex assemblages capable of “grading essays and diagnosing diseases” (check out Anthony Goldbloom’s talk: “The job we’ll lose to machines and the ones we won_t”).

Trained well (or when the right algorithms and systems have been created), machines could be of immense help with various tasks where fast and error-free computation over big amounts of data are required.

To get back to Shannon’s rat, the maze ahead of today’s learning machines is boundless and the paths it contains – innumerable. So are the opportunities before machine learning and the solutions it can bring to meeting computational challenges.

Want to learn more about text analysis?

White Paper: Text Analytics for Enterprise Use