On Sunday, Diane and I had a visit from Jennifer Pettis, who recently graduated from Georgetown University with a master’s degree in biology. Jennifer’s major professor was Janet Mann, and Jennifer did her research on the bottlenose dolphins at Shark Bay (Monkey Mia), Australia.
What Jennifer has done that is different from earlier studies of biosonar in the wild is to correlate behavioral states with the properties of click trains in bottlenose dolphins. Jennifer made audio recordings via a hydrophone while videotaping shallow-water dolphin activity and making voice notes on the video record. She then classified the behavioral state of the majority of dolphins in the study area from the video record. Click trains from the audio record were then analyzed and tagged according to the noted behavioral states. One of the clearest findings of her study was that click trains with low inter-click intervals (ICIs) were far more often emitted during the foraging behavioral state. She compared this tendency to the terminal buzz noted in studies of bat biosonar. Many bat species emit a series of FM sweeps, such that the intervals between sweeps is longest in the hunting phase, is reduced somewhat during the approach phase, and is greatly reduced during close-approach and capture. In dolphins, it has been noted that when dolphins are very close to a target, they will often emit a “buzz” or “burst pulse”, which is a click train with a very rapid rate (has low ICIs). Jennifer’s study goes beyond the anecdotal evidence of those earlier observations to a quantitative analysis that couples the click trains to behavior.
I’m happy to say that I was able to assist Jennifer a bit last year. She was able to utilize a program I had written to pull data from wide-bandwidth recordings of dolphin biosonar. This was a program written in Borland’s Delphi, essentially providing a computer-assisted means of recognizing and tagging click trains. Within it, I have a parameterized algorithm that runs through selected data and returns a list of offsets where it finds candidate clicks. The algorithm was based upon analysis of the trained state of a simple artificial neural system (specifically, Widrow and Hoff’s ADALINE model). By hand-tagging several click trains, I had made a training data set for the ANS. My method of running the ANS, though, was to present a fixed number of samples to the ANS, and then increment the starting position by one sample to prepare for the next presentation. While the ANS had nice accuracy in its recognition of candidate clicks outside the training set, running it was very slow. Therefore the switch to a parameterized algorithm utilizing the features apparently identified by the ANS.
I was myself surprised that very little needed to be done to shift between recognizing clicks in wide-bandwidth data (sampled between 400 and 5oo kilosamples per second) and Jennifer’s audio data (sampled at 44.1 kilosamples per second). My program is based on being a computer-assisted rather than completely automated system, so the picked clicks are presented visually for user review and editing. The user can adjust the indicated offset for a click, or delete a selection from the list, or add a new selection to the list of clicks. Something I added for Jennifer was a display of the power spectral density plot for a selected click. This was needed because in Jennifer’s data it was quite common for more than one dolphin to be emitting click trains at the same time. Fortunately, it was often possible to distinguish such overlapping click trains one from the other based upon the similarities of power spectral density plots within each click train, as well as the timing of intervals in each click train.
So it was gratifying to see the results from Jennifer’s analysis. She ran through her thesis defense presentation for us, and we also took a pass through a chapter in her thesis that she is now preparing for publication.
After that, we stepped out to do some jump-ups with the hawks (a sort of calisthenics where the hawk is encouraged to fly up to the gloved hand several times to receive a tidbit of food on an intermittent reinforcement schedule). Both Rusty and Glamdring readily came to Jennifer’s glove. Whether they had recall of the hunting outing we had with Jennifer last year or not is something I can’t tell, but they showed little reticence today. A bit surprising, overall, since the birds were basically fed free choice over the Christmas holiday and are currently still well over their usual flight weight. Usually, at these weights the hawks are in what we call “fat and sassy” mode, and not terribly responsive.