What if machines could think like us — comprehending social cues, visual prompts and spoken words just like a human would? For Computer Science and Artificial Intelligence Laboratory (CSAIL) Professor Patrick Winston, the Ford Professor of Artificial Intelligence and Computer Science and leader of the Genesis Group at CSAIL, uncovering the true nature of human intelligence is the next grand challenge.
To solve the puzzle of how humans think, Winston is employing classic engineering methodology to build systems that think and comprehend as people do using computational methods.
Motivated by a desire to advance artificial intelligence and create systems that operate in a manner consistent with high-level human thinking, Winston feels there is a substantial difference between machines that actually display human-like intelligence and those that possess superb computational powers such as IBM’s Watson system. Despite Watson’s Jeopardy! success, blunders such as answering “Richard Nixon” on a question asking for a first lady point to Watson’s lack of human-like intelligence, according to Winston.
For Winston, understanding what makes us different leads to questioning our uniquely symbolic nature, our ability to build descriptions using an inner language, and especially our ability to construct and tell stories, from fairy tales to case studies. By outfitting machines with language-enabled characteristics — the ability to direct our perceptual apparatus to solve problems, to describe events, and to teach through the sharing of stories — Winston feels scientists will be able to develop systems that not only can sift through vast amounts of information, but also can deploy human-like precedent-based judgment to help solve complex problems and deal with complex situations.
Winston and his students focus on creating systems that use previously acquired common sense knowledge and knowledge of plot patterns when tasked with story-understanding problems, just like humans do. During this long-term project many systems, with increasing levels of human-like intelligence, have been developed. The most recent incarnation, dubbed Genesis, embraces not only story understanding, but also perception-based question answering to shed light on human behavior.
“My approach to the scientific question of what we humans do differently is to develop computational solutions to behavioral-based problems,” Winston says. “We do all we can to ensure that our solutions are consistent with the strengths, weaknesses, and overall characteristics of human thinking. Then we do exploratory development to see if our solutions work. When we run into a brick wall, we stop, we study, we revisit the results in other fields, and we start over.”
The current Genesis model relies on CSAIL Principal Research Scientist Boris Katz’s START natural language processing systems to comprehend language and a vision system developed by CSAIL Principal Research Scientist Sajit Rao to comprehend visual information. Along with Rao, the Genesis Group is tackling DARPA’s Mind’s Eye Program, a project dedicated to developing visual intelligence capabilities in unmanned systems. Their current challenge is to develop a computer system that can successfully recognize, comprehend and converse about 48 different verbs.
A central component to the success of the Genesis system is a story-understanding program Winston and his students have developed that tries to understand everything from familiar play plots to descriptions of the alleged Russian cyber attack on Estonia in 2007. In order to evaluate different stories, the system is first provided with common sense knowledge in English, then inferred knowledge and finally higher level knowledge, referred to as the plot pattern level, where the system comprehends intangibles such as revenge and love. From these differing levels of information, the system is able to draw analogies between various stories it has encountered.
Winston and his students also place importance on how cultural bias can impact an individual’s perception of a story, thus when operating the story understanding apparatus they often display plot pattern level conclusions through an interface that exhibits the thoughts of one persona on the right and another on the left.
“We’re pretty excited about it because we see it as a first step toward understanding how we can understand other cultures better,” Winston explains. “It’s a kind of amplification of human intelligence by things that won’t be as smart as people for a very, very long time, but which, nevertheless, can sometimes see things through the fog of conflict and urgency that would not otherwise be obvious.”
Winston believes that eventually, as machines become increasingly intelligent, Genesis-like systems will be used in education to advance "personalized learning," tutoring each student based on a sophisticated model of what the student already knows and with an equally sophisticated model of the student’s style of learning, using both to help the student acquire knowledge in the best manner possible.
Descendants of Genesis may someday enable high-level analysis tools that will help decision makers in crisis situations by providing cultural context and historical grounding, thereby anticipating the unintended consequences of various policy options.
“We have problems in the world that look like they’ll be with us indefinitely and if we can imagine that intelligent systems will be one means to keep us out of trouble, then it makes sense to start on it now because if we don’t start on it we’ll never get it,” Winston says. “Of course the real reason we do all of this is because we humans are curious types and we want to understand how everything works, especially ourselves.”
For more information on Winston’s work, visit http://groups.csail.mit.edu/genesis/.
To solve the puzzle of how humans think, Winston is employing classic engineering methodology to build systems that think and comprehend as people do using computational methods.
Motivated by a desire to advance artificial intelligence and create systems that operate in a manner consistent with high-level human thinking, Winston feels there is a substantial difference between machines that actually display human-like intelligence and those that possess superb computational powers such as IBM’s Watson system. Despite Watson’s Jeopardy! success, blunders such as answering “Richard Nixon” on a question asking for a first lady point to Watson’s lack of human-like intelligence, according to Winston.
For Winston, understanding what makes us different leads to questioning our uniquely symbolic nature, our ability to build descriptions using an inner language, and especially our ability to construct and tell stories, from fairy tales to case studies. By outfitting machines with language-enabled characteristics — the ability to direct our perceptual apparatus to solve problems, to describe events, and to teach through the sharing of stories — Winston feels scientists will be able to develop systems that not only can sift through vast amounts of information, but also can deploy human-like precedent-based judgment to help solve complex problems and deal with complex situations.
Winston and his students focus on creating systems that use previously acquired common sense knowledge and knowledge of plot patterns when tasked with story-understanding problems, just like humans do. During this long-term project many systems, with increasing levels of human-like intelligence, have been developed. The most recent incarnation, dubbed Genesis, embraces not only story understanding, but also perception-based question answering to shed light on human behavior.
“My approach to the scientific question of what we humans do differently is to develop computational solutions to behavioral-based problems,” Winston says. “We do all we can to ensure that our solutions are consistent with the strengths, weaknesses, and overall characteristics of human thinking. Then we do exploratory development to see if our solutions work. When we run into a brick wall, we stop, we study, we revisit the results in other fields, and we start over.”
The current Genesis model relies on CSAIL Principal Research Scientist Boris Katz’s START natural language processing systems to comprehend language and a vision system developed by CSAIL Principal Research Scientist Sajit Rao to comprehend visual information. Along with Rao, the Genesis Group is tackling DARPA’s Mind’s Eye Program, a project dedicated to developing visual intelligence capabilities in unmanned systems. Their current challenge is to develop a computer system that can successfully recognize, comprehend and converse about 48 different verbs.
A central component to the success of the Genesis system is a story-understanding program Winston and his students have developed that tries to understand everything from familiar play plots to descriptions of the alleged Russian cyber attack on Estonia in 2007. In order to evaluate different stories, the system is first provided with common sense knowledge in English, then inferred knowledge and finally higher level knowledge, referred to as the plot pattern level, where the system comprehends intangibles such as revenge and love. From these differing levels of information, the system is able to draw analogies between various stories it has encountered.
Winston and his students also place importance on how cultural bias can impact an individual’s perception of a story, thus when operating the story understanding apparatus they often display plot pattern level conclusions through an interface that exhibits the thoughts of one persona on the right and another on the left.
“We’re pretty excited about it because we see it as a first step toward understanding how we can understand other cultures better,” Winston explains. “It’s a kind of amplification of human intelligence by things that won’t be as smart as people for a very, very long time, but which, nevertheless, can sometimes see things through the fog of conflict and urgency that would not otherwise be obvious.”
Winston believes that eventually, as machines become increasingly intelligent, Genesis-like systems will be used in education to advance "personalized learning," tutoring each student based on a sophisticated model of what the student already knows and with an equally sophisticated model of the student’s style of learning, using both to help the student acquire knowledge in the best manner possible.
Descendants of Genesis may someday enable high-level analysis tools that will help decision makers in crisis situations by providing cultural context and historical grounding, thereby anticipating the unintended consequences of various policy options.
“We have problems in the world that look like they’ll be with us indefinitely and if we can imagine that intelligent systems will be one means to keep us out of trouble, then it makes sense to start on it now because if we don’t start on it we’ll never get it,” Winston says. “Of course the real reason we do all of this is because we humans are curious types and we want to understand how everything works, especially ourselves.”
For more information on Winston’s work, visit http://groups.csail.mit.edu/genesis/.
