Connectome (12 page)

Neurons are like scientists in this regard. If a neuron spikes in response to Jennifer Aniston but not other celebrities, we might think that the neuron's function is the detection of Jen. But this neuron is embedded in a network of many other neurons. It would be a mistake to think of this neuron as a lone genius, detecting Jen all by itself. Newton's words ring even truer for neurons than for Newton: “If a neuron sees further, it is only by standing on the shoulders of other neurons.” To understand how a neuron manages to detect Jen, we need to know something about the neurons from which it receives information.

The weighted voting model I presented earlier forms the basis for a theory of what happens. Let's describe Jen as a combination of simpler parts. She has blue eyes, blond hair, an angular chin, and so on (as of this writing, anyway). If the list is long enough, it will uniquely describe Jen and no other celebrity. Now suppose that the brain contains neurons for detecting each stimulus in the list. There is a “blue-eye neuron,” a “blond-hair neuron,” and an “angular-chin neuron.” Now here is the central hypothesis: The “Jennifer Aniston neuron” receives excitatory synapses from all of these “part neurons.” The threshold of the “Jennifer Aniston neuron” is high, so it spikes only when all of the part neurons spike, a unanimous vote that happens only in response to Jen. In short, a neuron detects Jen as a combination of Jen parts, which are detected by other neurons.

This explanation sounds reasonable, but it raises more questions. How does the “blue-eye neuron” manage to detect blue eyes, the “blond-hair neuron” detect blond hair, and so on? I'm reminded of the funny story that opens the book
A Brief History of Time
by the physicist Stephen Hawking:

 

A well-known scientist . . . once gave a public lecture on astronomy. He described how the earth orbits around the sun and how the sun, in turn, orbits around the center of a vast collection of stars called our galaxy. At the end of the lecture, a little old lady at the back of the room got up and said: “What you have told us is rubbish. The world is really a flat plate supported on the back of a giant tortoise.” The scientist gave a superior smile before replying, “What is the tortoise standing on?” “You're very clever, young man, very clever,” said the old lady. “But it's turtles all the way down!”

 

Likewise, my answer is “It's neurons all the way down.” A blue eye is a combination of simpler parts: a black pupil, a blue iris, a white area surrounding the iris, and so on. Therefore a “blue-eye neuron” can be constructed by wiring it to neurons that detect these parts of a blue eye. Unlike the old lady, I can avoid the problem of infinite regress. If we keep on dividing each stimulus into a combination of simpler parts, eventually we will end up with stimuli that cannot be divided further: tiny spots of light. Each photoreceptor in the eye detects a tiny spot of light at a particular location in the retina. There is little mystery in that. Photoreceptors are similar to the many tiny sensors in your everyday digital camera, each of which detects the light at a single image pixel.

According to this theory of perception, neurons are wired into a network with a hierarchical organization. Those at the bottom detect simple stimuli like spots of light. As we ascend the hierarchy, neurons detect progressively more complex stimuli. Neurons at the top detect the most complex stimuli, such as Jennifer Aniston. The wiring of the network obeys the following rule:

 

A neuron that detects a whole receives excitatory synapses
from neurons that detect its parts.

 

In 1980 the Japanese computer scientist Kunihiko Fukushima simulated an artificial neural network for visual perception, which was wired up with a hierarchical organization
governed by this rule. His Neocognitron network was a descendant of the
perceptron
introduced by the American computer scientist Frank Rosenblatt in the 1950s. A perceptron contains layers of neurons “standing on the shoulders” of other neurons, as shown in Figure 19. Each neuron receives connections only from neurons in the layer just below.

 

 

Figure 19. A multilayer perceptron model of a neural network

 

The Neocognitron recognized handwritten characters. Its descendants display more impressive visual capabilities, such as recognizing objects from photographs. Although these artificial neural networks still make more mistakes than human beings do, their performance is improving year after year. This engineering success lends some plausibility to the
hierarchical perceptron
model for the brain.

 

In the wiring rule introduced above, we focused on how a neuron receives synapses from neurons that are lower in the hierarchy. Alternatively, we can look in the opposite direction and specify how a neuron sends synapses to neurons higher in the hierarchy:

 

A neuron that detects a part sends excitatory synapses to neurons that detect its wholes.

The two formulations of the rule are equivalent, because a stimulus detected by a neuron somewhere in the middle of the hierarchy can be regarded either as a whole containing a number of simpler parts, or as a part that belongs to a number of more complex wholes. Again taking a blue eye as our example of a stimulus, we can see it as containing simpler parts like the pupil, the iris, and the white, or as being part of more complex wholes like Jennifer Aniston, Leonardo DiCaprio, and the many other people who have blue eyes.

So the function of a neuron depends on its output connections, not only its input connections. To clarify this counterpoint, let's embellish the story of Newton and Leibniz. Suppose you read in the news about the unearthing of old documents proving that some unknown mathematician invented calculus fifty years before Newton and Leibniz did. After failing to convince others to pay attention, she died in obscurity and took calculus to her grave. Should we now rewrite the history books, crediting this unsung scholar rather than Newton and Leibniz?

Such revisionist history might sound fairer, but it would fail to recognize the social aspect of science. Earlier I argued that discovery is not just an individual creative act of a lone genius, because any new idea depends on old ideas borrowed from other people. In the same vein, one might argue that the act of discovery includes not only the creation of a new idea but also the act of persuading others to accept it. To receive full credit for a discovery, a person must influence others.

Newton's place in history is defined by how he used the ideas of his predecessors and shaped the ideas of his successors. Similarly, I'd like to propose that:

 

The function of a neuron is defined chiefly by its connections with other neurons.

This mantra defines a doctrine I'll call
connectionism.
It encompasses both input and output connections. To know what a neuron does, we must look at its inputs. To understand the effects of a neuron, we should look at its outputs. Both of these perspectives were taken above in our two formulations of the part–whole rule of wiring introduced for perception. As we continue our exploration of connectionist theories, we'll encounter plausible explanations of memory and other mental phenomena, in addition to perception.

That sounds fascinating, but is there any solid evidence for these theories in real brains? Unfortunately, we've lacked the right experimental techniques to find out. In the case of perception, neuroscientists haven't been in a position to find the neurons wired to the Jennifer Aniston neuron, and to see whether they indeed detect Jen parts. More generally, if we accept the defining mantra of connectionism, it follows that we cannot truly understand the brain without mapping neural connections—in other words, finding connectomes.

 

Here's a wonderful thing about the brain: You can think about Jennifer Aniston even if you are not watching her on television or seeing her in a magazine. Thinking of Jen does not require
perceiving
her; you are thinking of her if you recall her performance in the 2003 film
Bruce Almighty,
fantasize about meeting her, or contemplate her latest love interest. Can thinking, like perception, also be reduced to spikes and secretions?

Let's return to the experiment of Itzhak Fried and his collaborators for some clues. Their “Halle Berry neuron” was activated by an image of the actress Halle Berry, suggesting that it plays a role in perceiving her. But the neuron was also activated by the written words
Halle Berry,
indicating that it participates in thinking about her as well. So it seems that the “Halle Berry neuron” represents the abstract
idea
of Halle Berry, which can arise from either perception or thought.

Both phenomena can be regarded as specific examples of a more general operation: association. Perception is the association of an idea with a stimulus, while thought is the association of an idea with another idea. So how do perception and thought work together when you're recalling a memory? Let's consider a scenario.

It's a fine spring morning, and you are walking down the street on the way to work. You catch the scent of flowers; within a few steps the smell becomes overpowering. You're not yet conscious of the magnolias blooming at the side of the road, but all of a sudden you're transported far away. You remember standing next to a magnolia tree, outside the red brick house of your first sweetheart. He is holding you in his arms. You feel shy and embarrassed. A plane is flying overhead, and you hear his mother calling for you to come have a glass of lemonade.

By the time the recollection is complete, you are thinking of many ideas: the magnolia, the red brick house, your sweetheart, the plane, and so on. For each of these ideas, let's suppose there exists a corresponding neuron in your brain. A “magnolia neuron,” a “red brick house neuron,” a “sweetheart neuron,” a “plane neuron”—all are spiking as you recollect your first kiss.

How was all this spiking triggered by the magnolia smell? The spiking of the “magnolia neuron” was caused by neural pathways from your nose. But how can we explain why the “plane neuron” is active even though there is no plane in the sky, and why the “red brick house neuron” is active even though there is no red brick house? This must be the result of thinking, not perception.

To explain all this activity, let's hypothesize that the neurons are excitatory and are mutually connected by synapses into a structure known as a
cell assembly.
The one shown in Figure 20 is just a small example, but you could imagine a larger assembly containing many neurons all connected with each other. Omitted from the diagram are connections to and from other neurons in the brain. These connections would bring signals from sense organs or send signals to muscles. Here we focus on the connections within the cell assembly, which represent the associations involved in thought.

 

 

Figure 20. A cell assembly

 

How do these connections trigger the recollection of your first kiss? Since the neurons are assumed excitatory, the activation of the “magnolia neuron” excites the other neurons in the cell assembly to become active. You can imagine it like a forest fire jumping from tree to tree, or a flash flood surging through a web of desert ravines. A similar spreading of neural activity allows the magnolia smell to trigger the recollection of all the ideas involved in the entire memory of your first kiss.

Memory is wonderful when it works, but we've all noticed and complained about its failures as well. In fact, a feeling of difficulty often accompanies the experience of memory, while perception usually feels effortless. If the brain stored only a single memory in a single cell assembly, perhaps remembering would be a trivial task too. But many assemblies are required to store many memories. If cell assemblies were like islands, completely independent of each other, having many of them would be no problem. But it turns out they need to overlap, and that's where the possibility of failure creeps in.