So Callie had a strong tendency for side preference, which seems to be a normal variant among dogs. Still, I was disappointed that my feisty rescue wasn’t an Einstein. I had no idea about McKenzie’s preferences, but since at least one of our subjects couldn’t tell us the difference between peas and hot dogs, a change in the experiment was in order.
The next day, I reported my findings to the lab.
Andrew was disappointed. “If they don’t care about the difference between peas and hot dogs, how will our experiment work?”
“It won’t,” I replied. “If peas and hot dogs are the same to the dogs, then the hand signals convey no useful information. As long as they put their head on the rest, they know they will get a treat. They don’t care which.”
Nobody could understand why the dogs didn’t discriminate between the two foods. We were all stuck thinking like humans. We had to think like a dog.
“What if we just get rid of the peas?” I mused.
“You mean a reward versus no-reward experiment?” Andrew asked.
“Exactly. Even if the dogs don’t care about hot dogs versus peas, surely they care about hot dogs versus nothing.”
Andrew nodded in agreement.
It wouldn’t even require any new training. We already had the two hand signals. Left hand up meant “hot dog.” Now, two hands pointing toward each other would mean “no hot dog” instead of “pea.”
“Don’t you think the dogs will get irritated and stop doing the task?” Andrew asked.
It was a good question. If it were me in the MRI simulator, I would scoot right out of there as soon as I realized I wasn’t getting food all the time. Psychologists call this extinction, which means that if you stop rewarding a previously learned behavior, the behavior will eventually stop.
But dogs might see it differently. Not rewarding every trial might increase their motivation. This is called variable reinforcement—VR for short. VR is very common in animal experiments. A VR10 schedule means that sometimes the subject is rewarded but, on average, only once every ten trials. The unpredictability of VR tends to make animals more attentive and work harder to obtain the reward.
Something as drastic as VR10 would not work in our experiment. I just couldn’t see Callie sitting still for ten repetitions of hand signals to get just a tiny cube of hot dog. If I were in her position, I would begin to wonder after the third repetition without a treat. By about the fifth repetition without food, I would probably give up entirely and quit the experiment. I suspected Callie wouldn’t put up with it either. More important, it would result in an imbalance in the number of observations collected in the scanner. If we had ten trials of the no-reward hand signal to every one of the reward hand signal, it wouldn’t be an even comparison. We needed an equal number of rewarded and unrewarded trials for this to work. The solution was simple: VR2.
A VR2 schedule means that roughly half the trials will be rewarded (two trials for every one that is rewarded). This would give an equal number of observations for both hand signals. As long as we didn’t simply alternate, which would make it completely predictable for the dogs, then they should stay highly motivated.
That evening, I tried VR2 on Callie.
As usual, the rustling of the hot dog bag called her to the kitchen.
“Wanna do some training?” I said in my high-pitched doggie voice.
Callie cocked her head and tore off into the living room. When I got there, she was already in the tube with her head in the chin rest. To warm up, we went through several trials as usual. Left hand up, hold it for ten seconds, and then reward. When she seemed settled in, I flipped the two-hand signal that had previously meant peas. This time, after ten seconds, instead of giving her a pea, I just touched her forehead. She thought a pea was coming and tried to lick my hand. With nothing there, she looked puzzled.
I pointed to the chin rest and said, “Touch.”
Callie quickly placed her head down. To make sure she wasn’t confused, I immediately showed her the reward hand signal, and rather than wait ten seconds, rewarded her right away. The next trial, I gave the two-handed, no-reward signal and quickly ended the trial with a touch on the head. We repeated this for about five minutes, and amazingly, she didn’t get bored or leave the simulator. Instead, her posture and attentiveness improved. Her head positioning became more consistent, and her eyes were fixed in attention on my hands. Now when I showed the reward signal, I could see her pupils dilate, indicating a high level of positive arousal. And she remained motionless.
VR2 was a success! If Callie could catch on so quickly, I was sure McKenzie would too. And with her pupils dilating, it was clear that Callie now cared about the hand signals.
If this didn’t work, nothing would. We were ready.
18
Through a Dog’s Eyes
WE DIDN’T HAVE MUCH TIME to get Callie and McKenzie trained on the new version of the task. We could have taken longer with the training, but the logistics of finding a day when Mark, Melissa, Rebeccah, and the scanner were all available dictated the schedule, and the next available time that everyone could meet again was only two weeks away. If we missed the window in two weeks, we would have to wait another month to book a large chunk of time at the scanner. The pressure was on.
At least we knew the dogs could do this. Each time we had gone to the scanner, we had accomplished more than I had expected, and I was counting on this next time to be no different. The dogs knew what they had to do. The real uncertainty was how much data we would be able to collect and whether it would be enough to demonstrate caudate activity.
The fMRI signal is very weak. We measure activity as the relative change in signal intensity from some baseline level. But even in the best of circumstances the signal intensity rises by less than 1 percent. To make matters worse, the fMRI signal is noisy. The noise, which comes from heart rate variability, breathing, and even the electronics of the scanner, causes fluctuations in the signal that are ten times as much as the thing we are looking for. The signal-to-noise ratio (SNR) of fMRI is therefore quite low. Fortunately, the noise is random. If we collected enough repetitions during the experiment, we could average the fMRI signals from each, and the effects of noise would be diminished.
Often, when doing an experiment for the first time, you don’t know how big the signal is, so you have to make an educated guess at how many repetitions will be required to detect it. The Dog Project was on the verge of moving beyond a cute dog trick and into the realm of legitimate science. But to make this jump, we would first need to figure out how many repetitions would be required.
Andrew and I took a close look at what we had collected on the first scan day. Even though we had failed to find any differences between peas and hot dogs, there was still useful information in the data. We could estimate the SNR of the dog brain and, from that, determine how many repetitions Callie and McKenzie would have to do at the next scan.
Andrew zoomed in on the caudate of McKenzie’s brain. He pulled up a graph of the level of activity in the caudate for each scan that we had acquired. The first few scans had no signal because McKenzie hadn’t placed her head in the head coil until about the twentieth scan. But then it looked like noise. It was hard to tell how much of the noise was because of the usual sources or her moving slightly during the scan. The size of the fluctuations measured about 15 percent of the overall signal. This was much higher than in human studies.