In A Brief History of Intelligence: , artificial intelligence-developer Max Bennett asks, why don’t we have Rosey the Robot yet? We have robots that can do back flips and AI that can pass difficult exams and write entire essays, so why isn’t there one that can just do the dishes?
The answer is lengthy, because according to Bennett, we first need to understand the intelligence of our evolutionary forebears. Brief History of Intelligence makes this story accessible by looking at five key evolutionary breakthroughs that led to human intelligence. Despite the complex topic, Bennett’s excitement at his own findings jump off the page.
The first breakthrough described in Brief History of Intelligence, believe it or not, is the ability to steer. Our invertebrate ancestors of 600 million years ago, like the coral polyps, anemones, and jellies of today, had radial symmetry and didn’t steer toward or away from things. Bennett points out that nevertheless, the modern examples of these animals have neurons that operate remarkably similar to those of modern humans.
Around the same time, the ancestor of modern nematode worms appeared. The nematode was very different, because it had bilateral symmetry. Brief History of Intelligence makes the argument that this is why nematodes developed brains. “All these sensory inputs voting for steering in different directions *had to be integrated together* in a single place to make a single decision; you can go in only one direction at a time,” Bennett says. He also points out that the Roomba’s steering and learning abilities were explicitly modeled after this simple bilaterian template.
The second breakthrough, 500 million years ago, Bennett compares to reinforcement learning, a type of machine learning. Early invertebrates had neurons and neurotransmitters (e.g. serotonin and dopamine) similar to modern humans, but didn’t have the same brain structures, and the first vertebrates show versions of some of the very same brain structures found in humans today, like the basal ganglia.
Brief History of Intelligence says that research in AI helped solve problems in neuroscience, and vice versa. Neurogammon, a backgammon software program from the mid-1990s, implemented a machine version of temporal difference reinforcement learning which allowed it to challenge even the best human backgammon players. Some neuroscientists read an academic paper about Neurogammon and came to realize that this type of learning explained the behavior of subjects in their studies of dopamine’s role in learning.
Brief History of Intelligence‘s third breakthrough, 200 million years ago, is simulation, which includes imagination and learning by imagining. Mental simulation was seen in the earliest mammals thanks to a new brain structure, the neocortex. Bennett surmises that, “simulating actions is astronomically more computationally expensive and time-consuming than the reinforcement-learning mechanisms in the cortex-basal-ganglia system.”
Reinforcement learning enabled fast, but inflexible decisions based on “direct associations between current state and the best actions.” That means it’s model-free. Mammals retain this ability, but the neocortex allowed our lineage to additionally learn through model-based reinforcement, which is based on “a model of how actions affect the world and uses this to simulate different actions before choosing.”
Model-based reinforcement learning has proven difficult to implement in machine learning. “In most real-world situations,” Bennett says, “it is impossible to search through all possible options.” Exactly how do humans perform this search? We don’t know, but Bennett looks toward AlphaZero for clues.
AlphaZero, a computer program that plays chess, “used search not to logically consider all future possibilities (something that is impossible in most situations) but to simply verify and expand on the hunches that an actor-critic system [model-free learning] was already producing. … this approach, in principle, may have parallels to how mammals navigate the search problem.” However, an important difference is that AlphaZero applies the same search strategy to every move it makes. Bennett thinks it’s unlikely that mammals have a significantly better search algorithm, but we have more search strategies to choose from.
The fourth breakthrough in Brief History of Intelligence, 15 million years ago, is mentalizing, which appears in the first primates. The granular prefrontal cortex, a new brain structure, seems to allow us to simulate ourselves. Furthermore, mirror neurons in the premotor cortex of primates fire when the primates watch other primates do something, suggesting a kind of mimicry through simulation. This allows primates to acquire novel skills, not limiting them to known ones.
The fifth breakthrough, 100,000 years ago, is language, which is seen only in humans. While some animals can learn simplistic naming, like chimps that can use limited forms of sign language, only humans are known to create and use elaborate languages with complex grammar. Thanks to language, humans can collect and store knowledge, and pass it down for future generations to learn. This is something no other animal can do, and it underlies the rapid pace of human technological progress.
In the final section of Brief History of Intelligence, Bennett muses on humanity’s ability to cooperate via common myths, which he thinks can scale infinitely. Looking back at the five breakthroughs, it’s pretty easy to see why we don’t have a Rosey the Robot today, even though it looks closer than ever. Looking forward to a more distant future, we don’t know what the sixth breakthrough will be, or when it will come.
AIPT Science is co-presented by AIPT and the New York City Skeptics.
You must be logged in to post a comment.