Intelligence Limits on Information Technology
Computer intelligence is subject to myriad constraints dealing with the categorization and comprehension of the information they are fed, process, and derive answers and solutions from. Over time, computers have grown more complex in terms of the information they process, how they process it, and how the results intersect in meaningful ways with the humans that rely on them. In the article The Stupidity of Computers David Auerbach details how this intersection evolved, from the earliest attempts to optimize search engine results to the most recent forays into artificial intelligence research, and what the implications may be for social media and evolving technology.
The essence of the article boils down to "garbage in, garbage out." In essence, if humans can't categorize or quantify information parameters in the design and input stage correctly, then the outputs from their work will not deliver the results they are looking for. I found the material surrounding ontology and metadata of particular interest. Computers are unable to organize metadata or categorize ontologically without human intervention and input. Auerback asserts that on a platform like Amazon, a virtual sales floor with easily assigned, contextually limited, and easily understood categorization, objects have a logical organizational taxonomy without the need for a computer to try and piece together any underlying relational context. A chair or a book only have a couple of places where they can be categorized in the system and these can be easily understood as long as their inputs are correctly assigned. Outside a rigid system like this things get messy quickly, as Auerback explains using Wikipedia as an example. Wikipedia is the internet's go-to for quick and easy answers on almost any subject. But it's a process built, monitored, and categorized by humans, and the categorization possibilities are not as clear-cut as they are in Amazon. Many of the associations created are incomplete, or in error, and all of them are outside the understanding of machine intelligence as it is today. There are too many contextual leaps for a computer to be able to grasp how Wikipedia is organized, how these connections have been made, and why they are meaningful to reach other.
Jason Pontin in his WIRED article Greedy, Brittle, Opaque, and Shallow: the Downsides to Deep Learning discusses the very real limits of machine leaning and comprehension, stating "[d]eep learning’s advances are the product of pattern recognition: neural networks memorize classes of things and more-or-less reliably know when they encounter them again. But almost all the interesting problems in cognition aren’t classification problems at all." In consulting with experts on the topic, Pontin explains that artificial intelligence require massive amounts of information (greedy), that they are unable to transfer their understanding to a new set of parameters without essentially crashing (brittle), their coding is difficult to comprehend (opaque), and they have little to no understanding of a larger contextual world outside the parameters of their data set (shallow). Given these rather serious limitations, it's hard to see how machine intelligence can begin to draw close to a human model of independent intelligence.
Student Emily Zhao in her article The Limits of Artificial Intelligence Today was able to interview Yann LeChun, Facebook's Artificial Intelligence Research Lab's chief scientist, to get a more specific answer of why machine intelligence was struggling at making these connections. According to LeChun, “[humans] rely, at a minimum, on four interconnected capabilities to successfully navigate the world: (1) to perceive and categorize things around us; (2) to contextualize those things for understanding and learning; (3) to be able to make predictions based on past experience and present circumstances and (4) to make plans based on all of the above.” Machine intelligence isn't able to grasp the contextualization, to link together the cause-and-effect that humans use to make decisions. Like Auerbach, Zhao also uses Amazon as an example of what machine intelligence is good at- the ability to make single cause-and-effect links in logic. Moving beyond that single-link complexity, to ask a machine to consider multiple links when determining a course of action, is beyond our current capabilities.
Computers are capable of processing a tremendous amount of information much faster than the human brain. But it lacks the contextual information, the ability to draw associations between things that it has not be taught to be related, that would be required for a truly intelligent computer.
The essence of the article boils down to "garbage in, garbage out." In essence, if humans can't categorize or quantify information parameters in the design and input stage correctly, then the outputs from their work will not deliver the results they are looking for. I found the material surrounding ontology and metadata of particular interest. Computers are unable to organize metadata or categorize ontologically without human intervention and input. Auerback asserts that on a platform like Amazon, a virtual sales floor with easily assigned, contextually limited, and easily understood categorization, objects have a logical organizational taxonomy without the need for a computer to try and piece together any underlying relational context. A chair or a book only have a couple of places where they can be categorized in the system and these can be easily understood as long as their inputs are correctly assigned. Outside a rigid system like this things get messy quickly, as Auerback explains using Wikipedia as an example. Wikipedia is the internet's go-to for quick and easy answers on almost any subject. But it's a process built, monitored, and categorized by humans, and the categorization possibilities are not as clear-cut as they are in Amazon. Many of the associations created are incomplete, or in error, and all of them are outside the understanding of machine intelligence as it is today. There are too many contextual leaps for a computer to be able to grasp how Wikipedia is organized, how these connections have been made, and why they are meaningful to reach other.
Jason Pontin in his WIRED article Greedy, Brittle, Opaque, and Shallow: the Downsides to Deep Learning discusses the very real limits of machine leaning and comprehension, stating "[d]eep learning’s advances are the product of pattern recognition: neural networks memorize classes of things and more-or-less reliably know when they encounter them again. But almost all the interesting problems in cognition aren’t classification problems at all." In consulting with experts on the topic, Pontin explains that artificial intelligence require massive amounts of information (greedy), that they are unable to transfer their understanding to a new set of parameters without essentially crashing (brittle), their coding is difficult to comprehend (opaque), and they have little to no understanding of a larger contextual world outside the parameters of their data set (shallow). Given these rather serious limitations, it's hard to see how machine intelligence can begin to draw close to a human model of independent intelligence.
Student Emily Zhao in her article The Limits of Artificial Intelligence Today was able to interview Yann LeChun, Facebook's Artificial Intelligence Research Lab's chief scientist, to get a more specific answer of why machine intelligence was struggling at making these connections. According to LeChun, “[humans] rely, at a minimum, on four interconnected capabilities to successfully navigate the world: (1) to perceive and categorize things around us; (2) to contextualize those things for understanding and learning; (3) to be able to make predictions based on past experience and present circumstances and (4) to make plans based on all of the above.” Machine intelligence isn't able to grasp the contextualization, to link together the cause-and-effect that humans use to make decisions. Like Auerbach, Zhao also uses Amazon as an example of what machine intelligence is good at- the ability to make single cause-and-effect links in logic. Moving beyond that single-link complexity, to ask a machine to consider multiple links when determining a course of action, is beyond our current capabilities.
Computers are capable of processing a tremendous amount of information much faster than the human brain. But it lacks the contextual information, the ability to draw associations between things that it has not be taught to be related, that would be required for a truly intelligent computer.
Comments
Post a Comment