How Do Machines Learn?
Machines may be more intelligent than human beings, but that may not make them smarter.
By Sharon Lewis
One of the most photographed women at the 2016 South by Southwest (SXSW) festival was not human. Hanson Robotics’ humanoid robot Sophia became an instant celebrity after her first public appearance at the festival. Just a mention of her can quickly get people talking about machine learning and artificial intelligence.
People actually interact with machines and their internal learning processes every day, perhaps more so than with other human beings. Instagram censorship, Google search results, self-driving cars, and Amazon Alexa are all examples of machines that are constantly learning and getting better at what they do.
Machine learning is a pixel in the bigger picture of artificial intelligence. It is essentially how a machine learns to think. It does this by way of an algorithm, which is a set of instructions in a language that a machine understands (i.e., code) that tells it how to perform a task or solve a problem.
Using algorithms, the machine scans data sets such as text, audio or images, for patterns that it can group together and label. Programmers feed data and learning algorithms into the machine to ‘train’ it, so that it learns to perform a specific task. This is what helps computers differentiate between an image of a dog and a chair, even though both have four legs.
The machine’s performance improves over time with more ‘training’, as it relies on previously defined rules to figure things out. The more accurate the rules, the more accurate the prediction. How the machine is trained depends on what kind of learning the programmer wants to use- supervised, unsupervised, or reinforcement.
Supervised learning is essentially telling a machine what the correct answer is. The machine receives input and output data, both of which are labeled. The paired input and output data set is called a supervisory signal. Over time, the machine should be able to learn enough from these supervisory signals to sort out unlabeled data by itself.
Instead of giving the machine answers or labels, unsupervised learning asks the machine questions and provides them with input data to reach a conclusion. This does not involve any labeling, and instead takes the form of discerning patterns or anomalies from the input data.
Think of reinforcement learning as a game–the machine must score the highest points. It earns points for the right actions, and learns from the wrong ones. Programmers do this by incorporating a feedback loop. The machine receives negative feedback if it gives the wrong output.
Another term to keep in mind when talking about machine learning is ‘neural network.’ It sounds heavy duty, but it simply refers to a sequence of algorithms designed to identify patterns. In machine learning, a neuron does the thinking between input and output, and a bunch of them together form a network. The framework is loosely designed after the human brain, hence the term neural network.
In an interview with CNBC, Sophia agreed to destroy human beings as a joke with Dr. David Hanson, CEO at Hanson Robotic–and Sophia’s creator. The thought is scary, but scientists agree that it would be a long shot for anything like this to happen.
Machine learning is not a new concept–it dates back to the 1950s. What this means is that machines have been relying on human beings to teach them for five decades. And while they are becoming more intelligent and can learn to think for themselves, they can only adopt the subjectivity and abstractions of a human mind, not create their own. In this case, at least, the student has yet to become the master.