“Let’s go with our strength,” I said. “Our lab has spent the last decade studying the human reward system. We know a lot about how it works. There isn’t any reason to expect the dog’s reward system to be any different from the human one.”
“Reward-prediction error experiment?” Andrew asked.
The brains of all animals appear to act as prediction engines. Prediction, after all, is key to survival. If your brain weren’t predicting what would happen next, you wouldn’t be able to walk across the street without being hit by a car. Most of the brain’s predictions have to do with things in our environment, like cars, and things that other people are doing. The caudate nucleus (located within the basal ganglia) and the parts of the brain that feed into it are concerned with predicting rewards.
In the early 1990s, Wolfram Schultz, a Swiss neuroscientist, measured the activity of neurons in monkeys’ brains while they were trained on a simple classical conditioning task. When a light turned on in a monkey’s cage, it received a squirt of fruit juice in its mouth. Just like Pavlov’s dogs, the monkeys quickly began to anticipate the juice when the light came on. Schultz discovered that the neurons in specific parts of the brain followed the same pattern. Initially the neurons fired in response to the juice, but once the monkeys had learned the association with the light, the neurons fired to the light, not the juice. The neurons that showed this pattern were located in the heart of the reward system, the caudate.
Since Schultz’s discovery, neuroscientists have learned that these neurons don’t signal things that are pleasurable. Instead, they fire when something unexpected occurs that indicates something good is about to happen. If something is unexpected, then that means the brain made an “error” in predicting it. For this reason, scientists call these events reward-prediction errors.
We know where reward-prediction errors occur in the brains of monkeys and humans. Dogs, we figured, should be no different. And because the caudate is a well-defined structure, it made sense that we would be able to identify it in the dogs’ brains—assuming, of course, they held still for the MRI.
“We could train the dogs with a hand signal that indicated they would receive a hot dog,” I said.
“If the dogs learned the association between the hand signal and the treat,” Andrew agreed, “we should see caudate activity to the hand signal.”
“Just like Schultz’s monkeys,” I concluded.
Lisa spoke up and pointed out a flaw in this reasoning.
“How would you know that the dogs had learned the hand signal?” she asked. “After all, they’re not doing anything.”
She had a point. All of behaviorist learning theory depended on the manifestation of either a response, like drooling, or a behavior to indicate that the animal has actually learned something. We would have only the brain.
“We’ll have to rely on the dogs’ caudate,” I said. “A response there would be proof that they learned the signal. We could also look for other signs, like the pupils dilating in anticipation.”
There was another problem. An fMRI scan measures brain activity indirectly. What it actually measures are changes in the oxygen content of tiny blood vessels in the brain. When neurons fire, the surrounding blood vessels expand a little and let in more fresh blood for the neurons to replenish their energy storage. The scan picks up these changes in blood flow, and from that we deduce which neurons were active. But there is a catch. The brain is always on. It is a myth that we use only some small percentage of our brains. The truth is that we use all of it—just not all at once. Because the brain is always on, and blood is always flowing, fMRI can measure only changes in activity. When designing fMRI experiments, you always need a comparison, or baseline, condition.
Callie would be in the scanner, holding her head still and watching me. So many things could be going on in her brain there might not be a way to interpret the fMRI measurements. Even if we trained the dogs on a hand signal, we would still need something to compare their brain responses to. Ideally, the comparison condition would be almost the same as the thing of interest. You want to keep everything the same in both conditions except for the one thing that is being varied in the experiment.
To measure the response to a hand signal, we needed another hand signal as a comparison condition. This way, everything would be the same—holding still, watching the handler and even the handler’s movements. We would vary the meaning of the signals.
“How about another hand signal,” I suggested, “which means something else?”
“Like what?” Andrew asked.
“A different type of food,” I said. “Something the dogs don’t like as much as hot dogs.”
“Like what?” Lisa asked. “Sheriff likes everything.”
It was a fine line. We wouldn’t want the dogs to eat something nasty. We needed something that they would eat but not like as much as hot dogs. Dogs are mostly carnivores. It seemed logical that they wouldn’t value a vegetable as much as a piece of meat.
“How about peas?”
Everyone nodded as they envisioned how this would work. I held up my left hand in a “stop” gesture.
“Suppose this means ‘hot dog.’ ” I thought briefly about holding up my right hand for “pea,” but as we didn’t know the extent to which dogs distinguished left and right, this seemed like a bad idea. Instead, I held both hands flat in front of my chest, pointing toward each other. “And suppose this means ‘pea’?”
Mark nodded.
“Those signals should be easily distinguishable to a dog.”
The rest of the team agreed.
It was decided. The first canine fMRI experiment would be “Peas versus Hot Dogs.”
Over the next week, Andrew and I formalized the design of the experiment, which is in some ways like writing a screenplay. Every detail has to be planned in advance. The lab walls became our storyboard. We needed to decide how many times we would give peas and hot dogs and the order of their presentation. Dogs are very good at learning sequences of things, so we wouldn’t want to simply alternate between peas and hot dogs. If we did, the dogs would know that as soon as they got a pea, the hot dog would be coming next, and there would be no need to pay attention to the hand signals. To prevent this, the order would have to be random.
The most important detail, though, would be the timing of the experiment. Each repetition would have four elements. First, the dog would place her head in the chin rest. Because of the associated movement, this would cause artifacts on the scan being acquired at that moment. We would need to wait at least two seconds for the next scan to begin. Once the dog was settled in the chin rest and enough time had passed for the artifacts to decay, we would proceed to the second element, the hand signal.
Melissa and I would be giving the signals to our dogs, and all of our attention would be focused on Callie and McKenzie. It would be too much for us to randomly decide on the fly which hand signal to give, so Andrew would be standing next to us with a pregenerated list of the order of signals. He would hold up one finger for hot dogs and two for peas. The handler, facing the dog in the scanner, would then give the corresponding hand signal. Timing was critical.