Real-Life Artificial Intelligence


Computers that learn may sound like science fiction, but the technology exists! If you’ve ever used the “titles you might like” feature of an entertainment streaming service or an email client that filtered spam messages, you’ve already enjoyed the results of machine learning. Machine learning occurs when computers use new data to develop new algorithms and improve upon existing ones, without being specifically programmed to do so. The email client, for example, will analyze details about what type of emails you mark as spam, and try to filter out similar messages in the future.

If chess computer Deep Blue comes to mind, there’s an important distinction. While Deep Blue was pre-programmed with chess rules and strategies to implement, a computer with machine learning capabilities could adapt, and generate new methods for interacting with data. A modern rival for Deep Blue could play chess with minimal instruction and develop strategies by trial and error, the way a person would.

Machine Learning in the News

The Tesla Model S Autopilot is making headlines as a smart car that can not only drive itself, but can also learn to become a better driver. The cars share information about road conditions they have encountered, resulting in collective “fleet learning”. It would be difficult – and maybe impossible – for Tesla to program cars with protocols for every possible driving scenario; but if the fleet learning model is successful, manual updates will not be needed as frequently. In the future, self-driving cars may actually become safer than human-driven ones.

Google-acquired DeepMind created a program that has learned to play 49 video games – and beat human testers at over half of them! Though the games’ rules are not pre-programmed, the system can analyze gameplay patterns to learn which strategies earn a higher score. The DeepMind VP of engineering says this is the first time a self-learning system has managed “to complete [complex]tasks that are very challenging to humans.”

Also from Google Land, the new Inbox by Gmail app goes beyond filtering mail. A new feature called Smart Reply evaluates messages you receive to suggest possible responses. For example, faced with a work email requesting a document, Smart Reply might suggest answers like “I’m working on it” or “I just sent it.” Inbox also offers several other neat features that blur the line between email and to-do list, including the ability to “snooze” message notifications, set reminders, and mark threads as “done.”

What Machines Can (And Can’t Yet) Do

A computer processing data can be much more accurate and more efficient than a person – or team of people – doing the same work. Computers are good at finding patterns across large quantities of data, and are able to quickly analyze data for specific features. As IBM exec Guruduth Banavar puts it, “When given a goal, the work to get to the goal can be done by [machines]without supervision, but the goal itself… has to be provided to the system… [Machines are not] built for self-direction.”

In short, even complex systems that can learn to do intricate tasks need help making judgement calls.
Just look at IBM’s Watson, a powerful tool for answering complex questions, extracting pertinent information from documents, and analyzing patterns from large amounts of difficult-to-process unstructured data. In 2011, Watson’s creators added Urban Dictionary into his vocabulary to make him sound more human, and accidentally taught him to swear. Still, data scientists are hopeful that machine learning might help them to find insights in dark data, the portion of big data that current processing methods aren’t effective at mining for insights. This includes any data that companies routinely collect, but don’t use. IBM suggests that as much as 80% of all information in existence is untapped dark data.

Using common-sense reasoning, future search engines might be able to interpret images and videos that today’s engines can’t make sense of, and serve up improved search results. Currently, computers have trouble understanding analogies and relationships that children easily grasp. For example, a child would understand that a photo of a woman on a bicycle could represent the verb “to ride,” or that a woman pictured with a baby is also a mother. If strides in cognitive computing – which attempts to create computer systems that are structured more like a human brain – can close this gap, applications for artificial intelligence could range from healthcare to computer security, and nearly everything between.

Why Big Data Matters

How Lumiata wants to scale medicine with machine learning and APIs


About Author


Chelsea studied English and Sociology at the University of Georgia. Her interests change daily and span from tech to searching for the perfect cookie recipe.

Comments are closed.