One of the biggest objections to the idea of grandmother cells is a matter of math: If every recognizable part of our vivid perception of life is detected by a grandmother cell, there don’t seem to be enough neurons in the brain to handle the job. Skeptics also point out that neurons responding strongly to one particular person or place often turn out to respond (albeit more weakly) to another stimulus as well. For example, in Quiroga’s study some neurons that responded strongly to Jennifer Aniston also responded to Lisa Kudrow, Aniston’s costar on Friends. Grandmother cells are supposed to respond to only one thing, yet these neurons seem to be reacting to a range of things.
Instead of grandmother cells, the doubters argue, our brains use a large network of neurons to recognize people and objects. Some clever brain-simulating computer programs bolster this view. The “neurons” in these programs are just simple processors that receive input from other processors and can send signals of their own. A network for recognizing faces has a layer of processors that act like eyes, capturing the image and sending signals to another layer of processors that analyze the data and send their output to another layer, and so on. Finally the network selects a name from a list. At first it can throw out random guesses, but researchers train the network by strengthening links that produce correct answers and weakening links that produce wrong ones. In the end, the network can become almost perfect at recognizing individual faces.
But if you open the hood on one of these programs, you will not find any processor behaving like a grandmother cell, responding only to a single face and being solely responsible for recognizing it. Instead, it is the pattern of all the processors taken together that corresponds uniquely to each face. This sort of network behaves like a crowd of people in a football stadium creating huge pictures by holding up colored squares. A single person may hold up the same color when producing different pictures; the unique picture emerges only through the collective behavior of the crowd.
Despite all that, Bowers is not convinced that large networks can explain the brain’s remarkable powers of recognition. “I find it striking that you can locate a neuron that responds to only one face out of a hundred,” he says. In the network-based model, each neuron should light up for most or all of the faces. He also disputes the claim that true grandmother cells must respond only to a single stimulus. A single neuron that is dedicated to one input might also respond to related inputs, but only up to a threshold. At the high level of stimulus needed to produce a clear interpretation (“That person is my grandmother”), the cell could still function as Lettvin envisioned.
On the other hand, Quiroga suggests that the grandmother cell interpretation is also incomplete. He says his own experiments probably did not detect grandmother cells in the strictest sense. The electrodes he used could pinpoint only about 100 neurons among the tens of billions in the brain. Just because Quiroga detected a single neuron responding to Halle Berry did not necessarily mean that other, faraway neurons were not firing as well. Quiroga also used only about 100 images in his experiments. A neuron that responded solely to Halle Berry among the items Quiroga showed might well have responded to many other images he did not show.
Using statistical methods, Quiroga and his colleagues calculate that about 0.1 percent of the neurons in the medial temporal lobe (about 1 million) respond to any given image. And out of the many thousands of images that our brains can recognize, he estimates that each neuron can respond to a few dozen.
That interpretation suggests a genuinely new insight into how our brain recognizes people and objects: not with grandmother cells but not with huge networks, either. Instead, the brain seems to use what is known as a “sparse-coding network,” in which small groups of neurons work together to recognize an object and in which each neuron may be able to join a few different groups.
Quiroga has begun testing this idea using improved techniques to extract more information from brain electrodes. This will enable him to eavesdrop on several hundred neurons at once. In such a large sample, he may be able to find two neurons responding to the same image in the same brain. If he succeeds, it will be the beginning of a whole new story. It may not be as colorful as the tale of Dr. Akakhievitch, perhaps, but it could at last bring us a lot closer to understanding what goes on in your head when your eyes see a pattern of light and dark and your brain shouts, “Grandma!”
