How The Mind Works

At the beginning of Chapter 2, Pinker takes us through a thought exercise: If we met an alien, “what would it have to do to make us think it was intelligent?” (61). Pinker summarizes the criteria for intelligence as acting using rules based on some relation to reality or logical inference, pursuing a goal or desire, and using the rational rules to pursue the goal in different ways depending on the obstacles in place. Not meeting these rules indicates that the being is not intelligent. Magnets, for example, are obeying physical laws in moving toward one another, but they can’t be said to “want” to be touching, and they can only move in a direct line. If an obstacle appears, they remain on either side of the obstacle, as close as they can physically get; They do not attempt other means to get closer to each other.

These rules bring up the question of why people choose the goals they choose. Pinker points out that we still use common sense and intuitive psychology, our internal sense of how our mind and other people’s minds work, to understand why people behave the way they do. We do not look to models or algorithms. Many of our best theories of how the mind works still involve some leap of common sense—a step that isn’t backed by math or molecular movements but instead by what we intuitively know about how the system works.

The mind is often thought to be separate from the brain, sometimes linked to the soul or to a magical type of tissue that we haven’t found. However, the best theories of the mind come from the idea that the brain’s interactions produce the mind, almost like a substance in and of itself. It is not the physical structure of the brain that gives rise to the mind but the pattern of activity that creates the mind. You can build a very simple information processing machine (a Turing machine) that can take a set of symbols (numbers, for example) and use any well-defined set of rules or operations to produce a solution in the form of a new set of symbols. Pinker discusses how the mind can process information and even create new information. In one example, a machine is given a set of statements about a family tree. It shows all parents and siblings, but it does not have a representation for aunt or uncle. The goal is to find out if someone is an uncle. The machine knows that an uncle is the male sibling of a parent, so it must find the person’s parents and then all male siblings to answer if a specific person is an uncle. It can then store the uncle information, thereby adding to information readily available without any processing. The computational theory of mind takes this approach to intelligence: that it is a form of computation. However, human intelligence can handle partial information and probabilistic information such that it can register the probability of a variety of outcomes instead of only handling information that will necessarily produce the same outcome with the same input.

There are many criticisms of the computational theory of mind, and there are still many areas of the mind we don’t fully understand and that are difficult to explain even with the computational theory of mind. The paradox of logic presented in Lewis Carroll’s “What the Tortoise Said to Achilles” shows that with any inference system, at some point, an action must simply be executed—the system does not follow rules all the way through. Warren McCulloch and Walter Pitts introduced the idea that neurons simply responded to the sum of inputs to their system to either fire or not to fire. If the sum of inputs exceeded a threshold, they fired. If it did not, they did not fire. McCulloch and Pitts showed that the firing and not firing behavior could be equated to logical statements. AND statements were neurons that needed input from A AND B to reach the threshold and fire. OR statements were neurons that needed either A OR B to provide input to fire, and NOT statements were neurons that fired when they did not receive an input and stopped firing when they received an input. Each neuron is connected to many other neurons, and it is the pattern of firing and not firing that produces specific responses or larger behaviors. Just with those three statements alone you can produce many logical statements and actions, but they are still not enough to cover all human consciousness.

For the human brain, the situation is even more nuanced, as neurons can respond to gradations of input. They can fire at different intensities to indicate the probability that something is true. Those probabilities build to indicate how likely it is that we should engage in a certain action and allow probabilities for other options to be considered. The network of neurons is also highly interconnected, meaning neurons that fire prompt many other neurons to fire. If those neurons also reach threshold and fire back, the neurons are reinforcing each other and engaging even more neurons in the event. The volume and complexity of options for input and output starts to look large enough to encompass human experience.

These networks of neurons are called auto-associators. These auto-associators are content-addressable memory, meaning they can receive the input of any property represented by a neuron in the auto-associator network and start activating the other neurons in the group. If those neurons also receive input for their property, the entire network is soon active, allowing rapid identification of something in the environment. Another property of auto-associators is their “graceful degradation,” meaning they do not require everything to match exactly to summon the correct inference. If someone has a typo in a presentation, we do not get stuck, unable to comprehend anything until the typo is fixed. We can use all contextual clues to recognize small errors, fix them automatically ourselves, and move on in a useful direction. As a similarly useful property, auto-associators can perform constraint satisfaction. When we hear a word that could be several words because of similar pronunciations, we can use clues from logic and similarity. The human brain can apply multiple constraints, and the correct option is the one that satisfies all constraints.

The auto-associators also use patterns of activation to indicate categories and to associate multiple individual objects in the same category. They use the basic principle that if two objects have at least one similar property, they likely have other similar properties as well. The more similar properties they have, the smaller the category they inhabit. The final property of auto-associators is that they can learn. Learning involves changing the connection weights to change the firing pattern based on new information. For instance, as children, we often learn that an initial assumption about an animal—for example, that cats are dogs—is wrong. When we learn that we are wrong, we adjust the weights to accommodate that new knowledge and separate dogs and cats.

These auto-associators that can learn are called perceptrons, and perceptrons have one problematic flaw: Perceptrons can’t handle an exclusive-or statement. If a neuron should fire only when A OR B input is present and not fire if A AND B are present, we essentially have two thresholds. To handle this situation, the system can have an intermediate, “hidden” unit that processes A OR B so it will only respond if A OR B is active. Another hidden unit could process A AND B and be inhibited if A AND B are present. These hidden units send signals to the next step, and if the next steps receive both A OR B and A AND B, the neuron doesn’t fire. If it receives only A OR B, it fires. This situation lets us create an exclusive OR using the same principals already discussed.

To move from simple constructions to more complex mental operations, we need the network to have a few more elements. Connectionism, proposed by David Rumelhart and James McClelland, argues that the simple networks already described account for human intelligence. The reason we are more intelligent than other creatures is because we have more hidden layers allowing exclusive-or-type propositions, and we have more finely tuned connection weights from interacting with humans around us (social learning). In general, connectionism views the brain as creating connections by using properties of objects and experiences instead of creating individual representations of each experience. Instead of having a representation of each dog you see, you have representations of “furry” and “barks,” and when you see enough of the representations that fall under the category “dog,” you say that what you are seeing is a dog.

Despite having some explanatory power, connectionism has some issues for standard examples of everyday thought. The first is the idea of the individual. If we are primed to just use properties to determine what something is, then how do we know individual people or dogs? Using the standard model of having representations of many properties, and in which the more properties we encounter the more fine-tuned we can make your judgment (e.g., from animal to mammal to cat), we would need to have a representation of something that was unique to that individual. Ultimately, we would have representations of everyone we encountered, contrary to the view of connectionism. As humans, we can go even further. We can distinguish identical twins.

Compositionality is another problem facing connectionism. Compositionality is the ability to create a representation in which the parts have a meaning and the way the parts are put together has meaning. The two sentences “The baby ate the slug” and “The slug ate the baby” use the same words but have very different meanings. How could we represent both sets of meaning from the same words? One way is to have one neuron for each possible sentence, but that is impossible because the number of possible sentences far exceeds estimates of the number of neurons in the human brain. Instead, we represent the concepts and their roles in the sentence separately. Then, another set of neurons fires that represent the complete sentence as a whole unit after the concepts and roles have been represented.

Surprisingly, quantification is another problem for connectionism. As humans, we can talk about a specific man performing an action or man in general performing an action, and we can tell the difference between those two statements and many other statements requiring knowledge about how many individuals are involved. Connectionism struggles to explain this ability. For instance, Neal Cohen and Michael McCloskey trained a network to add two digits. They first went through adding “1” to other numbers, and then they trained it to add “2” to other numbers. However, when they trained it to add “2,” it stopped being able to add “1” and could only add “2”. This ability requires the brain to hold information in mind or anchor information in a place and still develop new ways to do something. It needs to be able to hold onto the ability to add “1” and then add the ability to add “2” using the same basic machinery. David Sherry and Dan Schacter argued that humans have specialized memory systems for this purpose. Endel Tulving referred to “episodic” or event-based memory for specific events that have occurred in our life or in the world and “semantic” or generic-knowledge memory for knowledge of stable facts that are true about the world around us.

Another area humans handle quite well but connectionism can’t explain is fuzzy logic. We can handle categories with clear and distinct boundaries and inclusion criteria and categories that are fuzzy at the edges, with lots of examples that don’t share traits. We know that penguins are birds despite not being able to fly or sing. We know that vegetables come in many different textures and colors and tastes, but we still associate “green” and “leafy” with vegetables most often. We tend to think of green and leafy vegetables as better examples of vegetables, but that doesn’t mean we don’t know a carrot is a vegetable. We carry clear and distinct and fuzzy thoughts about the same category in our head at the same time, meaning we are representing this information in some clear and organized way even though it would appear to be very disorganized. Humans can create rule systems that supersede similarities used by the associators and group things based on explanations. Humans can carry around multiple rule systems in their heads and apply them as appropriate, knowing that in some areas two items are very similar, and in other areas they are very different.

Pinker discusses the various views of consciousness, including as a synonym for intelligence, as self-knowledge, as access to information, and as sentience. For self-knowledge, people typically define consciousness as, “building an internal model of the world that contains the self” (134). Self-knowledge is also the easiest version of consciousness to model. Even programs report back on their status and function, and if we have representations of other people, we can have representations of ourselves.

For access to information, we can see that humans can report anything they sense and perceive and anything they can hold in memory, but they can’t report on everything happening internally, like stomach acid or neurotransmitter levels. There are, therefore, different pools of information and different levels of access to that information. These versions of consciousness can be extended to computers, which know about themselves and the functions of things to which they are connected. They can report information on a wide variety of self- and other-knowledge. Finally, sentience encompasses the subjective experience of the world, the idea of “if you have to ask, you’ll never know”. This version of consciousness is the part that can’t be explained in models and appears almost magical.