If you’ve ever used Amazon’s Alexa, Apple’s Face ID or interacted with a chatbot, you’ve interacted with artificial intelligence technology. AI technology actually hasn’t been around that long. In fact, it’s only in the past decade has artificial intelligence evolved into technology like facial recognition and autonomous cars.
“AI has become more tangible and accessible,” said Dave Rogenmoser, CEO of Austin-based AI writing company Jasper. “It’s less based in theory and academia and now it can be applied in ways that it can help people.”
Artificial intelligence research is one of the most exciting fields in tech. There are a lot of ongoing discoveries and developments, most of which are divided into four categories: reactive machines, limited memory, theory of mind, and self-aware AI. These types reveal more of a storyline than a taxonomy, one that can tell us how far AI has come, where it’s going and what the future holds. From reactive machines to artificial superintelligence, here’s what we can expect from AI.
More on AI Innovation 54 Artificial Intelligence (AI) Companies Delivering on Innovation
Related4 Types of Machine Learning You Should Know
When categorizing different types of artificial intelligence, the story of the technology’s evolution unfolds. The genesis of AI began with the development of reactive machines, the most fundamental type of AI. Reactive machines are just that — reactionary. They can respond to immediate requests and tasks, but they aren’t capable of storing memory or learning from past experiences.
“They cannot improve their functionality through experience, and can only respond to a limited combination of inputs.”
“Reactive Machines perform basic operations,” said Rogenmoser. “They cannot improve their functionality through experience, and can only respond to a limited combination of inputs.”
In practice, reactive machines can read and respond to external stimuli in real time. This makes them useful for performing basic autonomous functions, such as filtering spam from your email inbox or recommending movies based on your most recent Netflix searches.
Most famously, IBM’s reactive AI machine Deep Blue was able to read real-time cues in order to beat Russian chess grandmaster Garry Kasparov in a 1997 chess match. But beyond that, reactive AI can’t build upon previous knowledge or perform more complex tasks. In order to apply AI in more advanced scenarios, developments in data storage and memory management needed to occur.
The next step in AI’s evolution is developing a capacity for storing knowledge. But it would be nearly three decades before that breakthrough was reached, said Rafael Tena, senior AI researcher at Austin-based insurance company Acrisure Technology Group.
“All present-day AI systems are trained by large volumes of training data that they store in their memory to form a reference model for solving future problems.”
“There was a huge amount of progress in the 80s,” said Tena. But that eventually slowed. “There were small incremental changes,” he said. “Until deep learning came around.”
In 2012, the field of AI made major progress. New innovations from Google and Image Net made it possible for artificial intelligence to store past data and make predictions using it. This type of AI is referred to as limited memory AI, because it can build its own limited knowledge base and use that knowledge to improve over time. Today, the limited memory model represents the majority of AI applications.
“Nearly all existing applications that we know of come under this category of AI,” said Rogenmoser. “All present-day AI systems are trained by large volumes of training data that they store in their memory to form a reference model for solving future problems.”
Limited memory AI can be applied in a broad range of scenarios, from smaller scale applications such as chatbots, to self-driving cars and other advanced use cases. One example Rogenmoser cited was image recognition technology, which uses a massive database of photos to build memory and practice labeling in order to improve visual accessibility online.
“When an image is scanned by such an AI, it uses the training images as references to understand the contents of the image presented to it, and based on its ‘learning experience’ it labels new images with increasing accuracy,” he said.
In terms of AI’s progress, limited memory technology is the furthest we’ve come — but it’s not the final destination. Limited memory machines can learn from past experiences and store knowledge, but they can’t pick up on subtle environmental changes or emotional cues.
“Current models have a one-way relationship,” said Rogenmoser. “AI [tools] like Alexa and Siri don’t react with any emotional support when you yell at them.”
The concept of AI that can perceive and pick up on the emotions of others hasn’t been fully realized yet. This concept is referred to as “theory of mind,” a term borrowed from psychology that describes humans’ ability to read the emotions of others and predict future actions based on that information. As of right now, Rogenmoser said theory of mind in AI is just that: a theory.
“Machines may work better than us 90 percent of the time, but that last ten percent, what you would describe as common sense, is really hard to get to.”
“There is still a lot of research and development needed for AI to truly understand human needs,” he said. “[It] will have to get to a point where it can perceive humans as individuals whose minds can be shaped by multiple factors, and will have to adjust its behavior based on the perceived emotion as it interacts with people.”
Limited memory AI can accomplish a lot, but it can’t reach the same level as human intelligence yet. To illustrate: a self-driving car may perform better than a human driver the majority of the time because it won’t make the same human errors. But if you, as a driver, know that your neighbor’s kid tends to play close to the street after school, you’ll know instinctively to slow down while passing that neighbor’s driveway — something an AI vehicle equipped with basic limited memory wouldn’t be able to do. This example, provided by Tena, is just one of many that showcase the current limits of AI, and how a successful theory of mind application would revolutionize the technology.
“Chatbots, for instance, are very advanced, but they can still make errors,” Tena said. “Machines may work better than us 90 percent of the time, but that last 10 percent, what you would describe as common sense, is really hard to get to.”
Theory of mind could bring plenty of positive changes to the tech world, but it also poses its own risks. Since emotional cues are so nuanced, it would take a long time for AI machines to perfect reading them, and could potentially make big errors while in the learning stage. Some people also fear that once technologies are able to respond to emotional signals as well as situational ones, the result could mean automation of some jobs. But no need to worry just yet — Rogenmoser said that this hypothetical future, however, is still very far off.
“Right now, this intelligence is science fiction,” he said. “We’re not even close to developing this type of AI, so no one is getting their job stolen by AI.”
The stage beyond theory of mind, when artificial intelligence develops self awareness, is referred to as the AI point of singularity. It’s thought that once that point is reached, AI machines will be beyond our control, because they’ll not only be able to sense the feelings of others, but will have a sense of self as well.
“People both strive to create this type of AI and fear the consequences of its creation, worrying that this type of AI could steal our jobs or take over our world,” Rogenmoser said. “If this type of AI is successfully created, no one knows what the impact will be.”
Robot armies, mechanical overlords, sentient humanoids — there’s no divorcing public perception of AI from its sci-fi reputation. But with so many leaps left between limited memory and theory of mind, Rogenmoser said the real science fiction, where robots reach a near-human level of intelligence, is thankfully still very far off. “This type of AI only exists in movies and television — think Westworld,” he said.
“If this type of AI is successfully created, no one knows what the impact will be.”
Steps are being taken by researchers and engineers to develop rudimentary versions of self-aware AI. Perhaps one of the most famous of these is Sophia, a robot developed by Hong Kong-based robotics company Hanson Robotics. While not technically self aware, Sophia’s advanced application of current AI technologies provides a glimpse of AI’s potentially self-aware future. It’s a future of promise as well as danger — and there’s debate about whether it’s ethical to build sentient AI at all. But for now, Rogenmoser said we don’t need to worry about AI conquering the world.
“AI is going to become much better at solving real use cases, but I want to express that I don’t think this [means] the end of humans and the end of work,” he said. “We will continue to see AI pop up in useful ways to amplify the great work that people are already doing.”
More on AI Innovation Are You Sure You Can Trust That AI?
AI can be broken down into three umbrella groups: Artificial narrow intelligence, artificial general intelligence, and artificial superintelligence.
The first group, artificial narrow intelligence, is made up of AI tools designed to carry out very specific actions or commands. For instance, natural language processing AI is a type of narrow intelligence because it can recognize and respond to voice commands, but cannot perform other tasks beyond that. It’s designed to serve one function and cannot independently learn skills beyond its design.
The next category, artificial general intelligence, describes AI that can perform a wider range of actions. It can learn and think similarly to humans. Still a work in progress, the goal of designing artificial general intelligence is to be able to create machines that are capable of performing multifunctional tasks and act as lifelike intelligent assistants to humans in everyday life.
The last category, artificial superintelligence, is the stuff of science fiction. It’s theorized that once AI has reached the general intelligence level, it will soon learn at such a fast rate that its knowledge and capabilities will become stronger than that even of humankind. But at this point, it’s all speculation.
“Artificial super intelligence will become by far the most capable forms of intelligence on earth,” said Rogenmoser. “It will have the intelligence of human beings and will be exceedingly better at everything that we do.”
Leave a Reply
You must be logged in to post a comment.