A century of research on human brains and minds makes three things clear. First, human cognition can be understood only if it is studied at multiple levels, from neurons to concepts to computations (Marr Reference Marr1982/2010). Second, human and animal brains/minds are highly similar. Indeed, most of what we have discovered about our own capacities comes from multileveled studies of other animals (e.g., Hubel & Wiesel, Reference Hubel and Wiesel1959), suggesting that an understanding of human cognition is achievable. Third, the project of understanding human cognition has a long way to go. We have learned a lot about what we know when and what our brains are made of, but not how or why we know, think, and learn as we do.
Research on cognition in human infancy provides a case in point. Infants represent key geometric properties of the navigable layout, track objects as solid, continuously movable bodies, and endow others with goals and causal powers in ways that are highly similar to those of other inexperienced animals (e.g., Spelke & Lee Reference Spelke and Lee2012). These abilities are not shaped by encounters with the postnatal environment: precocial and controlled-reared animals exhibit them the first time they move through a navigable space (e.g., Chiandetti et al. Reference Chiandetti, Spelke and Vallortigara2014; Wills et al. Reference Wills, Cacucci, Burgess and O'Keefe2010), track an object over occlusion (e.g., Regolin et al. Reference Regolin, Vallortigara and Zanforlin1995), or encounter another animal (e.g., Mascalzoni et al. Reference Mascalzoni, Regolin and Vallortigara2010). Moreover, the basic ways that infants and animals understand space, objects, and goal-directed action remain central to our intuitive thinking as adults (e.g., Doeller & Burgess Reference Doeller and Burgess2008) and to the brain systems that support it (e.g., Doeller et al. Reference Doeller, King and Burgess2008; Reference Doeller, Barry and Burgess2010).
None of these findings should be surprising or controversial. Human cognitive and neural architecture is unlikely to differ radically from that of other animals, because evolution proceeds by modifying what is already present. Abilities to represent space, objects, and other agents are unlikely to be entirely learned, because most animals need to get some problems right the first time they arise, including finding their way home, distinguishing a supporting surface from a teetering rock, avoiding predators, and staying with their group. And innate knowledge is unlikely to be overturned by later learning, because core knowledge captures fundamental properties of space, objects, and agency and because learning depends on prior knowledge.
Despite these findings, we do not know how human knowledge originates and grows, and a wealth of approaches to this question are rightly being pursued. One class of models, however, cannot plausibly explain the first steps of learning and development in any animal: deep learning systems whose internal structure is determined by analyzing massive amounts of data beyond any human scale. Human learning is fast and effective, in part, because it builds on cognitive and neural systems by which we understand the world throughout our lives. That's one reason why the effort described by Lake et al., implemented in computational models and tested against the judgments of human adults, is important to the grand challenge of achieving a deep understanding of human intelligence. We think the biggest advances from this work are still to come, through research that crafts and tests such models in systems that begin with human core knowledge and then learn, as young children do, to map their surroundings, develop a taxonomy of object kinds, and reason about others' mental states.
Computational models of infant thinking and learning may foster efforts to build smart machines that are not only better at reasoning, but also better for us. Because human infants are the best learners on the planet and instantiate human cognition in its simplest natural state, a computational model of infants' thinking and learning could guide the construction of machines that are more intelligent than any existing ones. But equally importantly, and sometimes left out of the conversation, a better understanding of our own minds is critical to building information systems for human benefit. Whether or not such systems are designed to learn and think as we do, such an understanding will help engineers build machines that will best foster our own thinking and learning.
To take just one example from current technology that is already ubiquitous, consider mobile, GPS-guided navigation systems. These systems can choose the most efficient route to a destination, based on information not accessible to the user, and allowing users to get around in novel environments. A person with a GPS-enabled cell phone never needs to know where he or she is or how the environment is structured. Is such a device good for us? A wealth of research indicates that the systems that guide active, independent navigation in humans and animals from the moment that they first begin to locomote independently are broadly involved in learning and memory (e.g., Squire Reference Squire1992). Moreover, the brain activity observed during active navigation diminishes when the same trajectory is followed passively (O'Keefe & Nadel Reference O'Keefe and Nadel1978). How are these systems affected by the use of devices that do our navigating for us? If the cognitive and brain sciences could answer such questions in advance, researchers could design intelligent devices that both eliminate unnecessary cognitive burdens and provide needed cognitive exercise. Without a deeper understanding of how and why we learn and remember what we do, however, the designers of current technologies are working in the dark, even when they design devices to aid navigation, one of our best-understood cognitive functions (e.g., O'Keefe Reference O'Keefel2014; Moser et al. Reference Moser, Kropff and Moser2008).
Working in the dark posed less of a problem in past centuries. When each new tool that humans invented had limited function, and its use spread slowly, tools could be evaluated and modified by trial and error, without benefit of scientific insights into the workings of our minds. Today's tools, however, have multiple functions, whose workings are opaque to the end user. Technological progress is dizzyingly rapid, and even small advances bring sweeping, worldwide changes to people's lives. To design future machines for human benefit, researchers in all of the information technologies need to be able to foresee the effects that their inventions will have on us. And as Lake et al. observe, such foresight comes only with understanding. A great promise of human-inspired artificial intelligence, beyond building smarter machines, is to join neuroscience and cognitive psychology in meeting the grand challenge of understanding the nature and development of human minds.
A century of research on human brains and minds makes three things clear. First, human cognition can be understood only if it is studied at multiple levels, from neurons to concepts to computations (Marr Reference Marr1982/2010). Second, human and animal brains/minds are highly similar. Indeed, most of what we have discovered about our own capacities comes from multileveled studies of other animals (e.g., Hubel & Wiesel, Reference Hubel and Wiesel1959), suggesting that an understanding of human cognition is achievable. Third, the project of understanding human cognition has a long way to go. We have learned a lot about what we know when and what our brains are made of, but not how or why we know, think, and learn as we do.
Research on cognition in human infancy provides a case in point. Infants represent key geometric properties of the navigable layout, track objects as solid, continuously movable bodies, and endow others with goals and causal powers in ways that are highly similar to those of other inexperienced animals (e.g., Spelke & Lee Reference Spelke and Lee2012). These abilities are not shaped by encounters with the postnatal environment: precocial and controlled-reared animals exhibit them the first time they move through a navigable space (e.g., Chiandetti et al. Reference Chiandetti, Spelke and Vallortigara2014; Wills et al. Reference Wills, Cacucci, Burgess and O'Keefe2010), track an object over occlusion (e.g., Regolin et al. Reference Regolin, Vallortigara and Zanforlin1995), or encounter another animal (e.g., Mascalzoni et al. Reference Mascalzoni, Regolin and Vallortigara2010). Moreover, the basic ways that infants and animals understand space, objects, and goal-directed action remain central to our intuitive thinking as adults (e.g., Doeller & Burgess Reference Doeller and Burgess2008) and to the brain systems that support it (e.g., Doeller et al. Reference Doeller, King and Burgess2008; Reference Doeller, Barry and Burgess2010).
None of these findings should be surprising or controversial. Human cognitive and neural architecture is unlikely to differ radically from that of other animals, because evolution proceeds by modifying what is already present. Abilities to represent space, objects, and other agents are unlikely to be entirely learned, because most animals need to get some problems right the first time they arise, including finding their way home, distinguishing a supporting surface from a teetering rock, avoiding predators, and staying with their group. And innate knowledge is unlikely to be overturned by later learning, because core knowledge captures fundamental properties of space, objects, and agency and because learning depends on prior knowledge.
Despite these findings, we do not know how human knowledge originates and grows, and a wealth of approaches to this question are rightly being pursued. One class of models, however, cannot plausibly explain the first steps of learning and development in any animal: deep learning systems whose internal structure is determined by analyzing massive amounts of data beyond any human scale. Human learning is fast and effective, in part, because it builds on cognitive and neural systems by which we understand the world throughout our lives. That's one reason why the effort described by Lake et al., implemented in computational models and tested against the judgments of human adults, is important to the grand challenge of achieving a deep understanding of human intelligence. We think the biggest advances from this work are still to come, through research that crafts and tests such models in systems that begin with human core knowledge and then learn, as young children do, to map their surroundings, develop a taxonomy of object kinds, and reason about others' mental states.
Computational models of infant thinking and learning may foster efforts to build smart machines that are not only better at reasoning, but also better for us. Because human infants are the best learners on the planet and instantiate human cognition in its simplest natural state, a computational model of infants' thinking and learning could guide the construction of machines that are more intelligent than any existing ones. But equally importantly, and sometimes left out of the conversation, a better understanding of our own minds is critical to building information systems for human benefit. Whether or not such systems are designed to learn and think as we do, such an understanding will help engineers build machines that will best foster our own thinking and learning.
To take just one example from current technology that is already ubiquitous, consider mobile, GPS-guided navigation systems. These systems can choose the most efficient route to a destination, based on information not accessible to the user, and allowing users to get around in novel environments. A person with a GPS-enabled cell phone never needs to know where he or she is or how the environment is structured. Is such a device good for us? A wealth of research indicates that the systems that guide active, independent navigation in humans and animals from the moment that they first begin to locomote independently are broadly involved in learning and memory (e.g., Squire Reference Squire1992). Moreover, the brain activity observed during active navigation diminishes when the same trajectory is followed passively (O'Keefe & Nadel Reference O'Keefe and Nadel1978). How are these systems affected by the use of devices that do our navigating for us? If the cognitive and brain sciences could answer such questions in advance, researchers could design intelligent devices that both eliminate unnecessary cognitive burdens and provide needed cognitive exercise. Without a deeper understanding of how and why we learn and remember what we do, however, the designers of current technologies are working in the dark, even when they design devices to aid navigation, one of our best-understood cognitive functions (e.g., O'Keefe Reference O'Keefel2014; Moser et al. Reference Moser, Kropff and Moser2008).
Working in the dark posed less of a problem in past centuries. When each new tool that humans invented had limited function, and its use spread slowly, tools could be evaluated and modified by trial and error, without benefit of scientific insights into the workings of our minds. Today's tools, however, have multiple functions, whose workings are opaque to the end user. Technological progress is dizzyingly rapid, and even small advances bring sweeping, worldwide changes to people's lives. To design future machines for human benefit, researchers in all of the information technologies need to be able to foresee the effects that their inventions will have on us. And as Lake et al. observe, such foresight comes only with understanding. A great promise of human-inspired artificial intelligence, beyond building smarter machines, is to join neuroscience and cognitive psychology in meeting the grand challenge of understanding the nature and development of human minds.