With any luck, by next week the starling you started training will successfully learn to discriminate between the S+ and the S- stimuli. But what has it really learned? Did it simply memorize the specific stimuli you trained it on, or did it learn the categories the stimuli belong to? It ought to be easier to remember “peck when you hear a song from bird A” than to try to remember a bunch of (unconnected) stimuli, but this is something we have to test. Another way of stating the question is whether your bird generalized from the exemplars you provided to the category the exemplars came from.
Assuming this is the case, a related (but somewhat more difficult) question is how the starlings represent the categories internally. There are a few theoretical possibilities, but we’re just going to consider feature theory, which proposes that categories are represented by the combination of features that define membership. For example, if you were the subject, and the experimenter trained you to discriminate between chairs and tables, you might classify novel exemplars that are large and lack backs as tables and exemplars that are small and have backs as chairs. In this case, you might be more likely to misclassify exemplars that are border cases (like stools).
Working with your group, your task for next week is to design two experiments, one to test for generalization and one to test a hypothesis of your choice about the feature (or features) that your bird used to discriminate the stimuli.
To test for generalization, you present novel exemplars of the category and see how the birds respond. This is going to be pretty straightforward, because you only used 5 of the provided S+ recordings in the original training. You also need some control S- stimuli, and there are lots of those, too. Choose three novel S+ recordings and three novel S- recordings.
The first thing you need to do is come up with a hypothesis about what features your starling (or starlings in general) use to categorize the songs of different individuals. Features can be pretty much anything: the motifs, the sequence of motifs, the average pitch, the rhythm, the loudness, etc. The only constraint on your hypothesis is that it has to involve a feature that you can figure out how to manipulate in the sound files. Audacity will let you do basic cutting and pasting, as well as a whole host of complicated transformations. There are also many other programs for sound editing that you can use. You may need to spend a lot of time looking at spectrograms and discussing possibilities to come up with a hypothesis.
Next, you need to generate specific predictions from your hypothesis. The two most obvious ones you can make are that (A) removing the feature(s) from the S+ stimuli will reduce the response rate and (B) adding the feature(s) to the S- stimuli will increase the response rate. Another way of stating these predictions is that morphing the S- and S+ stimuli into each other will cause the response rate to increase as you get closer to the S+ stimuli and decrease as you get closer to the S- stimuli. In contrast, if you change some feature that isn’t being used, you wouldn’t expect to see any effect. Write down your hypothesis and your predictions! Even if you’re wrong, it’s important to be clear about what you’re testing.
Finally, design the stimuli that will test those predictions. You want to have stimuli that test both ends of the spectrum between S+ and S-. If this doesn’t seem possible, you can just test one end, but the results may be more difficult to interpret. Design between 5-10 stimuli. It’s important to vary the stimuli so you don’t pseudoreplicate (we’ll discuss this next week). For example, if you’re changing the order of the motifs, generate stimuli with several different orderings.
Upload the stimuli you design to Collab Resources. Create a subdirectory under “Starling Probe Stimuli” and put your files in there. The stimuli need to be 44.1 kHz, 16-bit WAVE files (no mp3s!).