What Machine Intelligence Cannot Do (Yet)

Part 3 of "Separating Hype from Reality," Chapter 2 of The Machine Intelligence Primer

Despite at-times astounding advances in MI, it is still quite far from approaching parity with human intelligence.

Intelligent machines suffer from a fundamental inability to perceive or make use of context—the circumstances or information that surround and give meaning to the tasks they perform. While humans rely on prior knowledge and situational awareness gleaned from multiple sources of information to take the right actions at the right times, machines see only the task directly in front of them. 

ANNs can have anywhere from dozens to thousands of layers of neurons. Advances in digital data availability and computing power in the past 5 years have made it possible to make ANNs perform better by massively increasing the number of layers in them. This is where the “deep” in DNN and deep learning comes from, and it’s what makes deep learning so special and exciting.¹ DNNS have proven to be remarkably accurate and efficient at wide-ranging tasks—particularly those concerned with categorization, such as image identification and speech recognition. 

If you’d like to visualize how a neural network works, we recommend you visit Google’s TensorFlow playground, which allows you to see how neural networks are trained and provides more background on how these amazing algorithms work.

Why is it that MI can perform certain tasks with great competence while lacking basic comprehension? Today’s MI is generated using machine learning algorithms designed to address specific problem spaces and trained to recognize patterns in those problem spaces with thousands upon thousands of carefully labeled example data relevant to that specific problem space (supervised machine learning). A given machine learning algorithm’s understanding of the world is therefore limited to whatever use case it was designed to address and whatever data was used to train it. An algorithm trained to recognize cats in pictures cannot recognize zebras unless it’s been shown thousands of tagged photos from the plains of the Serengeti. Moreover, an algorithm trained to recognize cats in pictures cannot also be trained to translate the word cat from English to Spanish—different use case, new algorithm needed. With each machine learning algorithm operating in relative isolation, it is quite difficult for machines to stitch together context. And, without context, it can be hard for machines to make good decisions in ambiguous or uncertain environments. 

Driverless cars are an excellent example of an emerging MI technology that attempts to address this problem. Autonomous vehicles rely on dozens or hundreds of sensors and cameras, each with their own carefully designed algorithms configured to work together to perceive and act on stimuli in the environment: a red light, a car in the next lane, a deer in the road. This works well enough for highway driving, where the roadway scenarios a car encounters are relatively predictable. Driving on city roads, however, is an exponentially more difficult and varied problem. While we can perceive whether a pedestrian is preparing to jaywalk, a driverless car would need dozens of sophisticated algo- rithms trained on data concerning people’s body language, gestures, facial expressions, speed of movement, and even the region the person is in (there are more jaywalkers in New York City than Charleston). MI’s issues grasping context is the reason driverless cars for highways are already here, while car companies forecast fully autonomous driving in cities to be 5, 10, or even 20 years away.^{1, 2}

The complexity and volume of the math that underlies machine learning algorithms makes them largely inscrutable—even to the researchers who create them. This is partic- ularly true of deep learning—currently the most popular and commercially ubiquitous form of machine learning—where each algorithm comprises many interdependent abstractions that cannot be disentangled or analyzed in isolation. When a deep learning algorithm produces an output incongruent with what its creator, or the organization that commissioned it, expected or intended, there is often no way to know why or how it happened. This has led many to refer to machine learning algorithms as “algorithmic black boxes.” 

Some MI researchers question why this matters: Humans make decisions all the time that are opaque, even to themselves. And humans, some argue, are inherently more fallible and prone to bias than cold, unfeeling computers. How do we know, for example, that a hiring manager’s decisions aren’t based on prejudice? Or that an executive’s “gut instinct” isn’t misguided? We can’t split our brains down the middle to understand our impulses and inclinations, so why should it matter that the same be true of the intelligent machines we create? 

There are a few important reasons to care about—and work to overcome—the inscru- tability of MI. First, MI enables tremendous scale: Machines can process information and make decisions on a much larger scale than humans. Their decisions and actions, therefore, can have much more far-reaching consequences than those of one or a few humans. A hiring manager might in isolation discount a few qualified candidates, but a recruiting algorithm used across a company could erroneously shut off many hundreds of candidates from job interviews. This ties into the second problem with algorithmic black boxes. If MI is perceived to have made a mistake—treated a certain group unfairly or told a driverless vehicle to take a wrong and dangerous turn—it becomes quite difficult to assign blame. 

MI’s reliance on supervised machine learning—and the tens to hundreds of thou- sands of examples or more of tagged data it requires—is one of its chief limitations today. While a child can learn to understand her mother’s tone of voice almost instinctively at a young age, machines require massive reams of labeled data to extract the same meaning from sound. This requires time and resources many organizations and individuals don’t have. 

Demand for labeled data is so high that some have suggested labeling data might become the new blue collar job of the 21st century (not a job most of us would want!). Organizations have turned to interns, junior staff, and even crowd-sourced labor platforms, like Amazon’s Mechanical Turk and Crowdflower, just to prepare their data for machine learning algorithms. This is an expensive solution that precludes smaller, less well-resourced compa- nies from taking advantage of emerging MI applications.^{3, 4}

Advancing unsupervised learning will unlock and make more accessible wide applications of MI.