Artificial intelligence and machine learning solutions have already been in our lives for many years now—at least 60. But in spite of the widespread application of these solutions in our everyday lives, machine learning and AI remain significantly misunderstood. In fact, it almost seems that these tools are even more misunderstood today than they were at their onset.
Investment in AI hit more than $9 billion in 2018, and hundreds of consumer products incorporate this technology in one way or another. From AI assistants like Google Home and Alexa to dolls that are designed to interpret how a child is feeling, this technology is already the norm.
What Is the Difference between AI and Machine Learning?
One of the roadblocks to understanding these tools is that the terms artificial intelligence and machine learning are often used interchangeably, even though they describe very different tools.
AI is divided into two main types: weak AI and strong AI. Weak AI tools are most commonly seen in everyday life. Sometimes called “narrow AI,” weak AI tools rely on programming and algorithms to simulate intelligence. Voice recognition tools, like Amazon’s popular Alexa, are an example of weak AI. It might seem like there is intelligence there, but in reality, these tools are simply listening for key words and following built-in programming to respond—they don’t have any ability to understand your words and interpret the meaning behind what you say.
Another example of narrow AI is non-playable characters in computer and video games. They seem to behave the way a human playing the game would, but they are really just following programming instructions.
Strong AI, on the other hand, is designed to operate in imitation of the human brain, with an awareness of context and an ability to learn and make better decisions tomorrow than it made today. This is obviously an extremely difficult task, and that is part of the reason that most AI systems you encounter today are examples of weak AI.
Types of Machine Learning
Machine learning is an example of a way that artificial intelligence can be applied to systems to help them learn and make changes without needing specific programming. There are a number of different types of machine learning algorithms that computer programs can use to access data and learn from it.
Supervised, unsupervised, semi-supervised, and reinforced machine learning algorithms all use varying levels of human intervention to interpret data.
In supervised machine learning algorithms, programs use labeled examples to make decisions about future events, and they often start from a known training dataset. Unsupervised machine learning algorithms don’t label any data, allowing the program to instead explore all available information and attempt to draw its own inferences.
Somewhere in between supervised and unsupervised is semi-supervised, which uses both a large amount of unlabeled information and a little bit of labeled information. This method appears to result in significant improvement in learning accuracy in comparison to the others.
Meanwhile, reinforcement machine learning algorithms allow the program to interact with its environment, learning about mistakes and rewards through simple feedback called the reinforcement signal.
From Learning to Understanding
The next frontier in machine learning tools is the evolution from simple learning to understanding, a component of strong AI tools. At present, machine learning tools are capable of little more than recognizing things like images and auditory signals and learning the rules of a simple game (card games, for example). But there is more to understanding than simply being able to recognize. Without developing true understanding and common sense, machine learning tools cannot continue to advance and adapt to changes in our constantly evolving world.
In order to move forward, it is time for machine learning tools to move on to develop true machine reasoning. Machine reasoning refers to the ability of the programs to identify connections between the observations and facts they are currently capable of recognizing. More development in the machine learning field will be stalled until we can overcome the limitations in these systems. Right now, machine learning tools cannot use the information collected or apply it across different areas without interference by human beings.
The trouble is that to develop machine reasoning there must be a way to codify everything from general concepts like rain or snow to abstract ideas like hunger and thirst. There are billions of things and billions of relationships that a machine would need to understand to achieve true machine reasoning and common sense. This is somewhat of a circular problem, since without automation tools, it is impossible to codify all the necessary information, and machine reasoning can’t be solved without that information being codified. The next breakthrough from machine learning to machine reasoning might be on the horizon.