Connectome (45 page)

 

4. Neurons All the Way Down

 

[>]
   
make scientific observations:
Quiroga et al. 2005.

[>]
   
photo of Julia Roberts:
Fried's experiment was striking because it was done in humans. His results are less surprising if you're familiar with the work of his predecessors, who did similar experiments in monkeys and other animals. For example, Desimone et al. 1984 reported neurons that responded selectively to faces.

[>]
   
celebrity supercouple:
Actually there were a few spikes, though not many. Fried and his colleagues did find another group of neurons in the same person that was selectively (dare I say nostalgically?) activated by Aniston and Pitt together, but not by Aniston alone.

[>]
   
“celebrity neuron”:
In a famous paper Horace Barlow called this the “grandmother cell” theory of perception, joking that there is a neuron in his brain that is active if and only if his grandmother is present (Barlow 1972). Gross 2002, however, credits the “grandmother cell” theory to Jerome Lettvin.

[>]
   
small percentage:
This “small percentage” model actually fits the data better than the “one and only one” model. Before, I emphasized the neurons that responded to a single celebrity, but these were actually a small minority. Many more neurons responded to no celebrities in the experiment, and even fewer neurons responded to two celebrities. To see that this is consistent with the “small percentage” model, compare the random sampling of celebrities with throwing darts while blindfolded. Finding a celebrity that activates a neuron is like hitting the dartboard; both events have low probability. It's most likely that no dart will hit the dartboard. If you're lucky, one dart will make it. It's very unlikely that two or more darts will. That being said, the experiment cannot rule out the existence of neurons that truly respond to just one celebrity. To identify such neurons, it would be necessary to show patients a huge number of photos.

[>]
   
number of possible patterns:
Here we've simplistically defined the activity pattern to be binary: Every neuron is either active or inactive. We could refine the definition to include the rates at which the active neurons spike. Then the activity pattern would contain even more information.

[>]
   
Leibniz was wrong:
The philosophically sophisticated may disagree with my claim, saying that Leibniz was referring not to perception but to qualia, the subjective feelings that accompany perception. In other words, he was really referring to consciousness, and measurements of spiking haven't told us much about that.

[>]
   
This kind of mind reading:
Can fMRI also be used for mind reading? Recently some researchers have argued that fMRI could be used to detect when a person is lying (Langleben et al. 2002; Kozel et al. 2005). The standard “lie detector” used in criminal prosecution and employment interviews is the polygraph. This measures blood pressure, pulse, respiration, and skin conductivity, which are supposed to reveal the hidden emotional stress that usually accompanies the act of lying. There is widespread skepticism, however, about the accuracy of the polygraph, and because fMRI directly assesses mental state by measuring the activation of the brain, it could potentially be more accurate. In laboratory experiments, some researchers have claimed good results with using a brain scanner to distinguish between lying and truth-telling human subjects. Based on this research, businessmen have founded two new companies seeking to commercialize fMRI lie detection. It's still not clear whether fMRI will turn out to be superior to the polygraph, but that's irrelevant to the discussion here. The point is that fMRI researchers are hoping only for the crudest kind of mind reading. None of them would dream of using fMRI to read out a highly specific mental property like the perception of Jennifer Aniston.

[>]
   
“the shoulders of giants”:
Recently some revisionist historians have interpreted this remark as sarcasm rather than modesty, as it comes from a letter to rival scientist Robert Hooke, who was a hunchback. Newton and Hooke later became enemies because of a dispute over optics.

[>]
   “
receives excitatory synapses”:
You may have noticed something missing from this rule: inhibitory neurons. Most cortical neurons are excitatory, but we should not neglect the inhibitory neurons, as they surely have some function too. Recall that the “Jennifer Aniston neuron” did not spike for photos of Jen with Brad Pitt. We can emulate this behavior by adding to our construction an inhibitory synapse from a neuron that detects Brad. If this synapse is strong enough, then its vote will override the votes from the neurons that detect components of Jen, and keep the neuron silent if Brad is present. More generally, it has been theorized that inhibitory synapses are helpful for making fine distinctions between similar stimuli. Excitatory synapses may enable a neuron to spike for a certain type of nose, while inhibitory synapses enable it to
not
spike for similar types of noses.

[>]
   
hierarchical organization:
Actually the part–whole rule was used to wire up only every other layer of his network. The other half were wired by another rule: A neuron receives excitatory synapses from neurons that detect slightly different versions of the same stimulus. The neuron has a low threshold for spiking and therefore responds to any of the stimulus variations. This rule is required for achieving another important property of perception: its invariance to “irrelevant” differences between stimuli.

[>]
   
perceptron:
Some use
perceptron
to refer only to the case of a single layer of synapses, and specify
multilayer perceptron
for the more general case. But Rosenblatt originally meant the term to refer to a multilayer network, and I follow his usage here.

[>]
   
the layer just below:
The perceptron has a feature that is not consistent with the known connectivity of the brain. Its pathways go only from the bottom of the hierarchy to the top. In real brains, there are also connections going in the opposite direction. What could be the role of these top-down pathways in perception, and how are they likely to be organized? In the “interactive activation” model of McClelland and Rumelhart 1981, a letter-detecting neuron receives bottom-up connections from neurons that detect the strokes of the letter. (Such part-to-whole connections were discussed in the main text.) But this fails to explain a simple phenomenon: How do you know that the middle letter of
C–T
is likely to be
A, O,
or
U,
and not
E
or
I
? In the interactive activation model, a letter-detecting neuron also receives top-down connections from neurons that detect words containing the letter. In the above example, an
A
detector is assumed to receive a connection from a
CAT
detector. More generally, one can imagine the rule “A neuron that detects a whole sends excitatory synapses to neurons that detect its parts.” This allows a neuron to detect a stimulus by weighing evidence received from
both
bottom-up and top-down connections.

[>]
   
people who have blue eyes:
It's because many wholes can share a single part that a hierarchical representation is more efficient than a flat one.

[>]
   
connectionism:
The term
connectionism
more commonly refers to a 1980s movement in cognitive science that sought to explain the human mind using model networks of weighted voting neurons. Philosophers of mind argued over its merits relative to the “symbolic” approach of understanding the mind as a digital computer. As this heated debate recedes into history, it's better to use the word in the broader sense I've defined, as an intellectual tradition that dates back to the nineteenth century and is still evolving.

[>]
   
perception or thought:
The MTL is regarded by some as the top of the hierarchy hypothesized earlier (see Figure 51). At the bottom are areas of the cortex devoted to perception alone. Thinking does not activate the neurons in these areas, or at least not so much. The dividing line between perception and thinking does not appear to be sharp. Rather, the involvement of neurons in thinking appears graded, increasing gradually as one ascends the hierarchy.

[>]
   
never function perfectly:
According to some theorists, inhibitory neurons may be more precise at controlling the spread of activity than neuron thresholds, providing for superior memory recall.

[>]
   
information overload:
Inhibitory neurons increase memory capacity by retarding the spread of activity. To serve this dampening function, the connections of the inhibitory neurons don't need much organization at all. If each receives synapses from a random selection of excitatory neurons, it will be activated whenever the “mob” is active. If it sends synapses back to another random selection of excitatory neurons, it will exert a dampening effect on the crowd. An engineer would say that inhibitory neurons exert “negative feedback” on excitatory neurons. The household thermostat is the classic example of negative feedback. If the temperature of a heated room increases beyond a certain point, the thermostat turns off the heat; if the temperature decreases, the thermostat turns on the heat. In both cases the thermostat acts to oppose the change in temperature, in the same way that inhibitory neurons act to oppose changes in the activity of excitatory neurons. In this view, inhibitory neurons play a supporting role in brain function, so their connections don't have to be very specific.

[>]
   
left to right:
Note that this looks like the perceptron shown earlier, but turned on its side. Although a synaptic chain can be viewed as a special case of a perceptron, it's quite different from the typical perceptron, which is used to model perception. The neurons in one layer of a perceptron typically detect different stimuli, so each is wired to a different subset of neurons in the previous layer. (Or if they are wired to the same neurons, the strengths of the synapses differ.) All the neurons in one layer of a synaptic chain get activated together, so their connections with the previous layer need not be different. The synaptic chain has been formalized in mathematical models by a number of researchers (see, for instance, Amari 1972 and Abeles 1982). The American theoretical physicist John Hopfield developed related models in the 1980s.

[>]
   
theory of connectionism:
Donald Hebb proposed and named the cell assembly (Hebb 1949). Early computer simulations of model networks with cell assemblies were performed in the 1950s. The English theorist David Marr and the Japanese theorist Shun-ichi Amari were two prominent researchers who studied the equations of such models using pencil and paper in the 1960s and 1970s (see, for example, Marr 1971 and Amari 1972). But the real heyday of connectionism came in the 1980s, following the seminal papers of John Hopfield (Hopfield 1982; Hopfield and Tank 1986). Using esoteric mathematical techniques from a branch of physics known as spin glass theory, theoretical physicists had a field day calculating memory capacity through a statistical treatment of the effects of overlap between cell assemblies (see Amit 1989; Mezard, Parisi, and Virasoro 1987; and Amit et al. 1985). By the time this flurry of activity petered out in the 1990s, these researchers had discovered many interesting properties of the models. Also around this time, the PDP Research Group, a collective of cognitive scientists, published an influential two-volume manifesto containing many interesting connectionist models (Rumelhart and McClelland 1986).

[>]
   
“Problem of Serial Order”: Lashley attributed the “associative chain model” to the British psychologist Edward Titchener, citing a book from 1909. Actually both authors spoke of chains of psychological associations rather than neural connections. Strangely, Lashley did not use the word
synapse
in his article, although he was a neuroscientist. Nevertheless, the notion of synaptic chains is implicit in his writing.
Lashley attributed the “associative chain model” to the British psychologist Edward Titchener, citing a book from 1909. Actually both authors spoke of chains of psychological associations rather than neural connections. Strangely, Lashley did not use the word synapse in his article, although he was a neuroscientist. Nevertheless, the notion of synaptic chains is implicit in his writing.

[>]
   
huge variety of activity:
There would also have to be points where two chains converge into one, or we would quickly run out of neurons.

[>]
   
problem of syntax:
In a similar vein of criticism, some computer scientists have argued that relations between ideas are richer than simple associations. To say that the ideas of fish and water are associated does not do justice to their relationship. It's more richly descriptive to say that a fish “lives in” water. Computer scientists represent such relationships with a “semantic network,” which looks like a connectome except that each arrow is labeled with a type of relation.

[>]
   
addressed Lashley's second:
These connectionist models achieve greater computational power by introducing latent or hidden variables, to augment the variables that are used to represent explicit ideas.

 

5. The Assembly of Memories

 

[>]
   
two-and-a-half-ton:
The blocks varied in size; this number is an estimate of the average (Petrie 1883). Most blocks were limestone, but some were granite.

[>]
   
2.3 million: Petrie 1883.
Petrie 1883.