by Roman Taraban, Ph.D., Texas Tech University
In the movie Arrival, the character Louise Banks is portrayed as a linguist who can decipher an alien language. For much of this task, Louise and colleagues are doing pattern matching, trying to establish a correspondence between English and the alien language. A critical piece of the plot is in the interpretation given to the translation of the alien message “offer weapon.” Louise’s competitors interpret this as “use weapon” and propose to attack and destroy the aliens. Alternatively, Louise considers whether there might be an alternative interpretation for weapon, like “tool” or “technology.” From a metacognitive perspective, we might describe the competitors as thinking at a cognitive level, interpreting the phrase literally and acting accordingly. Louise, we could say, acted metacognitively, questioning whether the literal cognitive process was sufficient. Throughout the movie, Louise questions the sufficiency of her cognitive resources for the task at hand, and figures out how to overcome her limitations. In the end, metacognition saves the day.
Normally, we think of metacognition as a value-added add-on to everyday thinking. The metacognitive individual in a sense transcends his or her own limitations. The person recognizes limitations of memory storage and speed of processing, and the value of external memory, spaced practice, planning, and so on. With this recognition of limitations of memory and processing comes a search for and discovery of strategies for managing cognition. This “higher-order” processing is metacognitive, and in the movie Arrival, Louise Banks is our metacognitive hero.
Although we are inclined to attribute metacognition to bright individuals, like Louise Banks, can we dismiss the possibility that metacognition can exist in “dumb” machines – dumb in the sense that they do not have human-like understanding? Intelligent machines, like computers, process patterns mechanically. Does a computer need to experience metacognition like a human in order for the process to be regarded as metacognitive? Is a jolt of adrenalin a necessary part of the recognition process signaling to us that we should monitor and check our calculations? The present blog is not about some distance aliens, but about a smart machine that is being widely used in many different applications today. The machine is IBM’s Watson.
There are clearly some areas in which today’s computers do not need to be metacognitive. Humans can only hold 7 + 2 chunks of information in short-term memory. An intelligent system like IBM’s Watson https://www.ibm.com/watson/developercloud/nl-classifier.html has 15 terabytes of cache memory and processes 80 teraflops per second, so neither short-term memory nor speed of processing are issues. Metacognitive processes for recognizing and preserving short-term memory would seem to be pointless, as would many of the metacognitive resource-management strategies that humans depend on. Would IBM Watson need to grab a pencil and jot a phone number onto a scrap of paper? Not likely
There may be other ways, though, that machines could exhibit metacognitive behaviors. For instance, a machine like IBM Watson might know that humans are limited in how much information they can process in a unit of time. As a metacognitive strategy, Watson might control and monitor the rate at which he verbalizes in conversation. Watson might change his linguistic register when conversing with a young child https://www.youtube.com/watch?v=vqjndtS8jQU . Watson could attempt to choose an appropriate theme with specific speakers, like Bob Dylan. In a commercial with Dylan, Watson wisely chose to talk about Dylan https://www.youtube.com/watch?v=oMBUk-57FGU. Watson apparently can monitor and modulate its own behavior depending on the context, goal, and particular interaction.
What about monitoring its own resources? If we gave Watson a set of test questions, it is not likely that Watson would reason about them metacognitively like a human. For example, Watson would not acknowledge that word problems are more difficult than straight calculations, so would attack the calculations first. However, it is not difficult to imagine situations in which Watson could reason metacognitively about his own processing and plan, control, and monitor those processes. For instance, recognizing that in the context of a crisis certain information is more critical, Watson could modify the order in which information is compiled and provided to, say, paramedics at the scene of a disaster. This would involve prioritizing specific information, queueing it up in a specific order, delivering it, and monitoring its effective transfer to the paramedics.
The irony, perhaps, is that Watson is not exhibiting “transcendent” behavior, which is how we might view metacognition in humans. Instead, Watson is simply carrying out mechanical computations, which, in a sense, are like any other computations that Watson carries out. The existence of machines like Watson should prompt us to ask whether our metacognitive ruminations may also simply be computational processes. Perhaps the sense of metacognition involving “higher-order” thinking, the self-pride we take in thinking about thinking, is an add-on, an epiphenomenon on which actual metacognitive processing in no way depends. In any case, the question of whether computers can be designed to use metacognitive strategies, to plan and modulate behaviors depending on circumstances, and to monitor the success of their efforts, may deserve a positive “yes.”