Artificial intelligence systems have beaten humans at chess, poker, Jeopardy, Go, and countless other games. But machines still falter when it comes to understanding some basic rules about the physical world. Building a machine-learning system based on how babies’ brain works could be a step towards making machine learning systems more efficient thinkers -- like humans, write, Susan Hespos and Brendan Dalton.
Computers have come a long way. From punch-card behemoths to hand-held voice-activated smartphones, advances in miniaturisation and computing power have supported the development of Artificial Intelligence (AI) from smart marketing algorithms, incredible image recognition capabilities, operating within the global financial market, efficient search engines and achievements like beating humans at games, considered to represent the apogee of human intelligence like chess or Go. Despite these achievements, AI is falling short.
In 1950 Alan Turing threw down a gauntlet when he said, “Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child’s?” In the seven decades that have passed since this challenge was issued, we have yet to build an artificial intelligence model that can rival the infant cognition of a typically developing 1-year-old.
One place where artificial intelligence has failed is in replicating an infant’s understanding of physics. We are not referring to the difficult physics questions about the nature of the big bang or black holes, but rather simple physics concepts like knowing that an unsupported object falls, knowing that a hidden object continues to exist and expecting that a ball rolling down a ramp will ricochet off a wall at the bottom of the ramp. This asks us to call into question when it was that learnt how objects behave and interact. Did anyone explicitly teach us these notions?
Research in cognitive development shows that humans have expectations about common-sense aspects of physics in the first months of life.
Research in cognitive development shows that humans have expectations about these common-sense aspects of physics in the first months of life. To us, visual perception seems effortless. This is because we have a dedicated low-level neural machinery for processing visual information that occurs unconsciously and automatically without draining our mental energy or conscious attention. This knowledge is evident without anyone explicitly teaching us. Pause and ponder for a moment that underneath all the things that vary across humans, there exists a set of perceptual and conceptual capacities common to everyone. These capacities include expectations that objects have permanence, in that they do not blink in and out of existence (e.g., your lost keys still exist even though you can’t find them), two solid objects do not pass through one another (e.g., the ball will bounce off, not pass through, the wall).
Humans aren’t the only smart ones when it comes to physical concepts. Expectations about where an object is and how it moves are apparent to other species as well. Rhesus macaques expect an object to stop when it meets a wall and not pass through it. Humans and chickens have similar expectations about partially hidden objects. To humans and many other animals, knowledge about how objects behave and interact seems ubiquitous, but how we acquire these expectations remains a mystery.
What we seem to do unconsciously and automatically takes a massive amount of computing power to reconstruct.
One gains more respect for this ubiquitous human capacity when computers try to model these expectations. What we seem to do unconsciously and automatically takes a massive amount of computing power to reconstruct. Traditionally, AI models start with a blank slate and are trained on data with many different examples akin to gaining experience in the real world. For example, an AI model can process videos of unsupported objects falling and balls rolling down ramps and into obstacles in its path. The model detects patterns across the examples and constructs predictions about where objects will go next. A recent paper by Luis Piloto and colleagues made a direct comparison between a deep-learning model that started with a blank slate versus one that was based on the physical concepts inspired by perceptual and conceptual capacities that are evident in infants. The findings indicate that visual animations can account for some common-sense physics learning, but not enough to account for what we see in infants. The infant-cognition inspired model made more accurate predictions, generalized the expectations to new animations, and learned faster than the traditional model. In other words, the AI models require some ‘built-in knowledge’ about physical concepts to match what a typically developing baby can do.
It turns out that embodied experience leads to critical insights about the world.
These results lend new insight to the age-old question of what capacities are present from birth (therefore a product of evolution) and what capacities are acquired through experience. There are several examples of ‘built-in knowledge’ across the animal kingdom (e.g., a songbird’s ability to learn a song characteristic of its species, an ant’s ability to wander a terrain in search of food and then make a direct path back to the nest, and a human’s ability to acquire language). These examples appear to be universal across each species, show little variation, and are never explicitly taught. The capacities may emerge early, but they tend to be primitive and become more elaborate and refined over time.
A second reason why AI falls short compared to babies and other animals could be how tactile and kinaesthetic experience play a role in understanding the common-sense aspects of physics. It turns out that embodied experience leads to critical insights about the world. For example, in the first months of life, babies do not have very good control over their limbs. It takes the majority of the first year to walk or crawl and their ability to pick up objects is initially very limited. Consequently, this provides an interesting opportunity to test what infants know from watching the world compared to when they can watch and manipulate it. A series of clever experiments by Amy Needham and colleagues got 3-month-old infants to wear Velcro mittens and play with Velcro covered toys. In essence, the infants could pick up the toys before they developed the manual dexterity to achieve these goals on their own. After training trials with Velcro mittens, there were measurable advances in the infants’ thinking. These findings demonstrate how tactile experience changes infants’ expectations about the world. In the case of Velcro mittens, it allows us to see that behind the variable swipes at objects, infants have goals and intentions that drive their behaviours.
What does this mean for the study of artificial intelligence? It suggests that as engineers create more sophisticated robotics, they might want to look at what developmental science can contribute to building better AI systems that simulate the human mind.
There is a growing acceptance that artificial intelligence falls short of actual intelligence, especially in developing the kind of general intelligence seen in humans and animals.
A third and final difference to consider is the types of problems where computer models succeed and fail. The most successful examples of artificial intelligence tend to focus on process. In contrast, most theories on humans’ thinking agree in positing that both the nature (the innate capacities) and nurture (the part learned from experience) matter. According to Lisa Miracchi, most AI avoids modelling consciousness, thought and knowledge.
Infants are, no doubt, cuddly bundles of joy. In addition, we champion another important attribute that infants have – they are stunningly smart! They are smart from an early age, and it is difficult to understand how the ‘knowledge’ they have could be constructed merely from experience in the world. Infants in the first months of life have expectations about how the world works. They expect unsupported objects to fall, and that objects don’t just blink in and out of existence. These concepts are foundational for us adults – we regard them as common-sense.
Turing reasoned that a computer could be made to think like an adult if we simulate a child version of the mind and provide it with the appropriate experiences. What Turing was suggesting was that we teach a computer common sense. Common sense is usually defined as sound judgement in practical matters. But that which is practical to an intelligent human being is anathema to an AI. There is a growing acceptance that artificial intelligence falls short of actual intelligence, especially in developing the kind of general intelligence seen in humans and animals. However, models might improve by leaps and bounds if they rise to Turing’s challenge and model how infants develop. Until it then, perhaps we should think twice before we allow AI to drive our car?