| THE LURE OF AUTOMATED BAT ACOUSTIC CLASSIFICATION IN THE AGE OF AI | |||
| JOSEPH M SZEWCZAK; joe@humboldt.edu; | |||
| Bat echolocation recording and classification has become foundational to research and conservation efforts worldwide. However, the simple nature of bat echolocation calls, their situational plasticity, rapid attenuation, and convergent call types across species pose inherent challenges to bat acoustic classification. Effective classification in all natural settings must account for these effects with thorough and inclusive training data and analysis. The easily acquired, shorter-duration calls from hand-released bats create training datasets insufficient to classify the longer calls that our monitoring efforts tend to record on the landscape. Attempts at machine classification, such as the recent proliferation of AI approaches in backyard automated units, can render deceptively impressive classification performance on training data sets that inadequately cover species' call repertoires, but that promised performance evaporates against larger more representative data sets and field data with signals other than bats. In general, as the number of call samples used to train a classifier increases, classification performance descends asymptotically to a lower limit, and this effect can be used to assess the sufficiency of a training data set. As a further complication, few people have the dedicated training to recognize erroneous identification, and this has led to an unfortunate flood of poor data reporting. | |||
| |||
Speaker Bio: Joe Szewczak serves as a Professor of Biology in the Department of Biological Sciences at Cal Poly Humboldt. He earned a bachelor of science in engineering from Duke University, and a Ph.D. in comparative physiology from Brown University investigating hibernation physiology in the big brown bat. He has investigated bat physiology, ecology, distributions, bats and wind energy, and their vocalizations, which lead him to develop SonoBat software for analyzing bat and bird vocalizations. |