Results are in!

I know it’s been quite a while since you heard from me, but that’s because Julia and I (and our pigeons) have been very busy for the past 7 months going through behavioral training and testing for our experiment that you helped fund!

Now, at long last, the results are in! There are a lot of experiments with a lot of results to report, so I will break them down into separate lab notes for each of the experiments. Then I’ll post a concluding lab note that will summarize the entirety of the results and tell the whole story, including what the future may hold.

Ready? Let’s dig in!

I’ll start by briefly restating the goals of the experiment, the experimental design, and state our hypotheses at the outset.

Goals: When we open our eyes, we see a world populated by objects. Objects appear as coherent sets of features and properties that become bound together in perception. There’s been quite a bit of work since the seminal work by Anne Treisman in the early 1980s on perceptual binding in humans (https://en.wikipedia.org/wiki/Anne_Treisman). There has also been some work on perceptual binding in monkeys. But very little work has been done in other mammals, or in birds. This was what we wished to study.

We were specifically interested in exploring whether the pigeon’s perceptual system can bind different features of an object. We picked object identity (specified by shape and color) and location as the two features of interest. Would pigeons show evidence of object-location binding? And if so, what are the necessary conditions to foster object-location binding?

Methods: To study object-location binding, we used an Object-Place Learning (OPL) task. In this task, a sequence of four objects is presented, one at a time. Each object was presented at one of four locations (or quadrants) on a touchscreen. All the bird had to do was peck at the object when it appeared. Doing so resulted in a food reward, mixed grain delivered from a food hopper located below the screen.

Here’s a picture of a bird in one of our touchscreen-equipped operant chambers (note, the task on the screen in the picture is not the OPL task we are discussing here, but it’s a lovely picture of one of our star birds, don’t you think?).

 

There were 12 birds placed on this task, and each was allocated to one of four treatment groups (thus, an n=3 per group). In Group Location, the location at which an object appeared repeated in each sequence, but the objects appeared in random order. In Group Object, the objects were displayed on the screen in a consistent order, but the locations at which they appeared were randomized. In Group Both, both object identity and location appeared in a repeating consistent sequence. Finally, in Group Neither, both the objects and locations appeared in randomized sequences. See the figure below that illustrate two possible consecutive trials, each with a run through the 4-item sequence, for each of the four training conditions.

 

Once the birds were responding relatively quickly to each presentation (this sounds simple, but took many months to accomplish!), we threw into each session some nonreinforced probe presentations to test what they had learned. There were four types of probe trial (see Figure below). One type switched the order of locations. Another switched the order of objects. A third type switched both, but preserved the object-location association. And a forth type (No Binding) switched both, but in such a way that the object-location association was not preserved.

 

What did we predict would happen? The predictions may seem a bit complex at first, but they fall into a few logical categories. First, we expected that if the sequence order of a feature’s presentation repeated consistently during training, then switching the order of that feature on a probe trial should cause the bird to respond more slowly to that probe. Such a reaction time (RT) cost would provide evidence that the bird had encoded the sequence order for that feature. For example, in the Location group, since the order of locations was consistent during training, switching the order on the Location probe trial should cause the bird to respond more slowly. But switching the object order for that same bird should not produce an RT cost because there was no consistent object order for that bird to encode (object order had been randomized, and only location order was consistent). A similar prediction could be made about Object probe trials for birds in Group Object.

Birds in Group Both should experience an RT cost for both Object probes and Location probes, while birds in Group Neither should not show an RT cost on any probe trials since object and location order was randomized during training.

Finally, to address the question about object-location binding, we predicted that only birds in Group Both would be able to bind the two features into a unified perception. How would this affect responding on the probe trials? Well, if the birds in this group had bound object and location together, then any probe trial that breaks this binding should result in an RT cost. Object, Location, and No Binding probes all break the object-location association. Only the Both probes preserve the object-location association, though the order of their presentation is switched, and thus some RT cost might be expected.

RT cost was measured as a percent increase in RT relative to training (baseline) trials.

What did we find? Here is a figure presenting the probe trial data.

 

Let’s take the groups one at a time, starting from the right side of the data figure. First let’s look at probe trial data for Group Neither where we didn’t expect any RT cost. None was found! This is a good manipulation check that giving the bird probe trials did not introduce any methodological issues or artifacts (i.e., screwing up the data accidentally). It also shows that when sequence order was random for both object and location information, “switching” the order on probe trials was undetectable because order was already random during training.

Next let’s look at Group Object. Only the order of objects was consistent during training, so we expected that if they encoded object order, switching object order on Object probe trials should produce an RT cost. No RT cost was observed, however, which seems to suggest that birds in this group did NOT encode the order of object during training!

Third, let’s look at the data from Group Location. Finally, we see that there was a significant RT cost on some of the probe trials! But which ones? Location, Both, and No Binding probes. All three of these probe trials involved a switch in location order, which was a feature that was consistent for this group. Switching object order, which had been random during training, had no effect on RT, as we predicted it wouldn’t. The difference between Groups Location and Object indicates that birds do attend to location order during training, but perhaps not object order. This is an interesting point to which we’ll return in a future lab note.

Finally, let’s look at the results from Group Both. A few very interesting things jump out at us. First, responding was slower on No Binding, Object, and Location probe trials than on baseline training trials. The large RT cost on No Binding and Location trials was not surprising because both types of probe trial involved a switch in location, and we had already seen in Group Location discussed above that birds appear to encode location order and slow down their response when location order is changed. What is surprising is that birds in Group Both were slower to respond on Object probe trials. This is surprising because none of the birds in Group Object showed any indication of having learned the order of objects! This surprising result suggests to us that something about having a predictable order of locations during training facilitated learning the order of objects as well! The way we learn about our world is not always linear, but often interacts with other aspects of the learning environment. The same is apparently true for pigeons as well.

The other very interesting result is that although marginally slower, responding on Both probe trials was NOT significantly slower than baseline training RT! Furthermore, responding on Both trials was significantly faster than on Object, Location, and No Binding trials! The only way to explain these differences is that, despite having the object order and location order both switched on Both probe trials, they were switched in such a way that their object-location associations (binding) were preserved. By preserving the object-location association, the birds were able to quickly recognize and respond to the probe even though its features were presented out of sequence. This is the VERY FIRST report of successful perceptual binding in pigeons (or any bird as far as we’re aware of). A great, heartfelt thanks goes out to all of our backers for making this discovery possible!

In the next installment we’ll present the results of a follow up experiment that manipulated the training trials in an attempt to make the object information more salient and learnable for the birds in Group Object.

Stay tuned!

Aaron Blaisdell