How Does Current AI Stack Up Against Human Intelligence?
Minds in machines: comparing biological and synthetic intelligence.
Posted November 12, 2018
We can gauge the current state of artificial intelligence by comparing it with what is currently known about human intelligence. Nine of the most impressive current examples of machine intelligence include IBM Watson, driverless cars being produced by Google and other companies, applications of deep learning such as the DeepMind programs for Go and chess, speech recognition programs such as Alexa, Siri, and Dragon NaturallySpeaking, Google Translate, recommender systems used by Netflix, Amazon, and other companies, face recognition used by Facebook, robots produced by Boston Dynamics, and poker players developed at the University of Alberta and Carnegie Mellon University.
Human intelligence results from 12 typical features that are characteristic of human intelligence even though they may not be necessary or sufficient conditions. These features are perceiving, problem-solving, learning, reasoning, abstracting, planning, deciding, understanding, feeling, acting, creating, and communicating.
Drawing on my forthcoming book, Brain-Mind, these 12 features operate in human minds by virtue of eight mental mechanisms that include imagery, concepts, rules, analogy, emotions, actions based on intentions, language, and consciousness. Together, the 12 features and eight mechanisms provide a total of 20 benchmarks for evaluating current artificial intelligence.
Each of the machine intelligence exemplars reproduces specific intelligence features. For example, IBM Watson successfully carries out problem-solving, learning, reasoning, planning, deciding, creating, and communicating. But it seems very weak on perceiving, abstracting, understanding, feeling, and acting. Driverless cars are already adept at perceiving, problem-solving, learning, deciding, planning, acting, and communicating. But they do little reasoning, abstracting, understanding, feeling, or creating.
Overall, current AI systems are notably ineffective with respect to abstracting, understanding, and feeling. Moreover, although they carry out some kinds of problem-solving and learning, there are other kinds that are within human capabilities but not yet operating in computers.
The human ability to carry out the various features of intelligence is the result of the mind possessing numerous mental mechanisms that result from neural mechanisms, but not all of these are present in current AI systems. Although some examples of machine intelligence can take sensory input, none of them currently are capable of manipulating complex images such as visual and auditory representations that are an important part of human creativity. Some AI systems such as Watson work with purely syntactic concepts that are ungrounded in semantics that connect with the world, whereas other AI systems such as driverless cars that do have world-connected semantics are not adept at connecting concepts to each other. Many AI systems work with rules, but the rules usually operate with ungrounded symbols. Recommender systems do a very simple kind of analogy but cannot deal with complex relations, such as causality, that are part of human analogy use. There are some AI systems such as driverless cars and robots that carry out actions, but action is missing in many AI systems. Some AI systems, such as IBM Watson’s Debater and Google Translate can do powerful operations on language, but still fall short of human abilities. No current AI system operates with emotions and consciousness. So current AI is clearly missing some of the mechanisms that are important for human intelligence.
Most strikingly, current AI systems are lacking in generality. All the programs are special-purpose, applying only to specific applications such as driving or playing games. In contrast, humans are general-purpose problem solvers and learners, able to operate in many domains such as building, cooking, socializing, teaching, driving, playing, and so on.
In sum, current AI is very impressive in approximating to some aspects of human problem-solving and learning. But it has notable gaps in imagery, abstraction, relational analogy, emotions, consciousness, full-blown language, and general creativity. So I think that AI has a long way to go before human intelligence is surpassed. Super-intelligence and the singularity are far away.
The incredible explosion in the power of artificial intelligence is evident in daily headlines proclaiming big breakthroughs. What are the remaining differences between machine and human intelligence? Could we simulate a brain on current computer hardware if we could write the software? What are the latest advancements in the world's largest brain model? Participate in the discussion about what AI has done and how far it has yet to go while discovering new technologies that might allow it to get there.
See the video here.