Search

Object classifiers that attempt to mimic dolphin echolocation require an auditory weighting function representative of dolphin peripheral auditory processing. An evolutionary program (EvPg) was used to fit the frequency-dependent output of a bank of bandpass filters to the auditory sensitivity of the bottlenose dolphin, Tursiops truncatus. Pseudo-Gaussian (PG) and rounded exponential (ROEX) functions were used to describe individual filter shapes. Variables determining the number of filters per model, overall filter shape and amplitude scaling were submitted to the EvPg for optimization. Maximum deviation (P e ) between model output and the sensitivity of the dolphin was used as a measure of similarity between the two, i.e., lower P e indicated a greater similarity. The number of filters converged upon 37 for all ROEX models and ≤ 45 for all PG models. The P e of the best-performing PG model was 0.08, and for all ROEX models was 0.13. Greatest deviations typically occurred below 5 kHz and above 1...

The relationships between acoustic features of target echoes and the cognitive representations of the target formed by an echolocating dolphin will influence the ease with which the dolphin can recognize a target. A blindfolded Atlantic bottlenose dolphin (Tursiops truncatus) learned to match aspect-dependent three-dimensional targets (such as a cube) at haphazard orientations, although with some difficulty. This task may have been difficult because aspect-dependent targets produce different echoes at different orientations, which required the dolphin to have some capability for object constancy across changes in echo characteristics. Significant target-related differences in echo amplitude, rms bandwidth, and distributions of interhighlight intervals were observed among echoes collected when the dolphin was performing the task. Targets could be classified using a combination of energy flux density and rms bandwidth by a linear discriminant analysis and a nearest centroid classifier. Neither statistical model could classify targets without amplitude information, but the highest accuracy required spectral information as well. This suggests that the dolphin recognized the targets using a multidimensional representation containing amplitude and spectral information and that dolphins can form stable representations of targets regardless of orientation based on varying sensory properties.

The development of a unique dolphin biomimetic sonar produced data that were used to study signal processing methods for object identification. Echoes from four metallic objects proud on the bottom, and a substrate-only condition, were generated by bottlenose dolphins trained to ensonify the targets in very shallow water. Using the two-element ('binaural') receive array, object echo spectra were collected and submitted for identification to four neural network architectures. Identification accuracy was evaluated over two receive array configurations, and five signal processing schemes. The four neural networks included backpropagation, learning vector quantization, genetic learning and probabilistic network architectures. The processing schemes included four methods that capitalized on the binaural data, plus a monaural benchmark process. All the schemes resulted in above-chance identification accuracy when applied to learning vector quantization and backpropagation. Beam-forming or concatenation of spectra from both receive elements outperformed the monaural benchmark, with higher sensitivity and lower bias. Ultimately, best object identification performance was achieved by the learning vector quantization network supplied with beam-formed data. The advantages of multi-element signal processing for object identification are clearly demonstrated in this development of a first-ever dolphin biomimetic sonar.

A biomimetic neural network was used to model the ability of a bottle nosed dolphin to recognize aspect-dependent geometric objects. Each echo train was recorded and an Integrator Gateway Network (IGN) was trained to discriminate among the objects using spectra of the object echoes. The IGN classifies objects using an average-like sum of the spectra from successive echoes. However, combining echoes may reduce classification accuracy if the spectra vary from echo to echo. The dolphin and the IGN learned to recognize the geometric objects, even though orientation was free to vary. The process of recognition using cumulated echoes was robust with respect to nonstationary raw input. The results were interpreted as evidence for the formation of aspect-independent representations of the objects. >

Steller sea lion (Eumetopias jubatus) mothers and pups establish and maintain contact with individually distinctive vocalizations. Our objective was to develop a robust neural network to classify females based on their mother-pup contact calls. We catalogued 573 contact calls from 25 females in 1998 and 1323 calls from 46 females in 1999. From this database, a subset of 26 females with sufficient samples of calls was selected for further study. Each female was identified visually by marking patterns, which provided the verification for acoustic identification. Average logarithmic spectra were extracted for each call, and standardized training and generalization datasets created for the neural network classifier. A family of backpropagation networks was generated to assess relative contribution of spectral input bandwidth, frequency resolution, and network architectural variables to classification accuracy. The network with best overall generalization accuracy (71%) used an input representation of 0-3 kHz of bandwidth at 10.77 Hz/bin frequency resolution, and a 2:1 hidden:output layer neural ratio. The network was analyzed to reveal which portions of the call spectra were most influential for identification of each female. Acoustical identification of distinctive female acoustic signatures has several potentially important conservation applications for this endangered species, such as rapid survey of females present on a rookery.

The Department of Defense lacks refined capability to assess and predict potential impact of tactical and experimental sound sources in the presence of marine mammals. Although there is sustained concern over the effects on marine mammals of man made sound in the oceans, there is very little direct information about what sound frequency-intensity combinations damage marine mammal hearing. Our broad objective is to transition information about effects of DoD sound types on marine mammal auditory anatomy and acoustic ecology to predictive models and mitigation tools. This effort responds directly to the DoD capability to comply with the National Environmental Policy Act requirements and will contribute directly to answering the National Council's Research Needs related to the effect of low-frequency sound on marine mammals (1994, 2000). This final report summarizes the accomplishments of SERDP Project CS-1082 spanning the lifetime of the project from FY98-FY-00.

Dolphins demonstrate an adaptive control over echolocation click production, but little is known of the manner or degree with which control is exercised. Echolocation clicks (N approximately 30,000) were collected from an Atlantic bottlenose dolphin (Tursiops truncatus) performing object discrimination tasks in order to investigate differential click production. Seven categories of clicks were identified using the spectral conformation and relative position of -3 and -10 dB peaks. A counterpropagation network utilizing 16 inputs, 50 hidden units, and 8 output units was trained to classify clicks using the same spectral variables. The network classified novel clicks with 92% success. Additional echolocation clicks (N24,000) from two other dolphins were submitted to the network for classification. Classified echolocation clicks were analyzed for animal specific differences, changes in predominant click type within click trains, and task-related specificity. Differences in animal and task performance may influence click type and click train length.

CS-1082 was a FY98 New Start. Our broad objective is to transition information about effects of DoD sound types on marine mammal auditory anatomy and acoustic ecology to predictive models and mitigation tools. Currently, the DoD lacks scientifically defensible information concerning the safe operation of many of their training and testing systems in the presence of marine mammals. There is only very little direct information about what sound frequency-intensity combinations may damage the hearing of marine mammals. This effort responds directly to the DoD capability to comply with the National Environmental Policy Act requirements and will contribute directly to answering the National Research Council's Research Needs related to the effect of low-frequency sound on marine mammals (1994). This project consists of three inter-related tasks. Task 1 consists of otopathological analyses of marine mammal ears. Task 2 consists of otopathological analyses of baleen whale ears, the results of which will motivate development of a biomimetic model of baleen whale hearing and responsiveness to DoD sound types. Task 3 utilizes predictions about sensitivity generated in Task 2, plus statistical sampling models and acoustical classification algorithms, to develop a capability to automate the use of the U.S. Navy's Integrated Undersea Surveillance System (IUSS) for mapping the distribution of whales.

This study demonstrated the ability of a false killer whale (Pseudorca crassidens) to discriminate between two targets and investigated the parameters of the whale's emitted signals for changes related to test conditions. Target detection performance comparable to the bottlenose dolphin's (Tursiops truncatus) has previously been reported for echolocating false killer whales. No other echolocation capabilities have been reported. A false killer whale, naive to conditioned echolocation tasks, was initially trained to detect a cylinder in a "go/no-go" procedure over ranges of 3 to 8 m. The transition from a detection task to a discrimination task was readily achieved by introducing a spherical comparison target. Finally, the cylinder was successfully compared to spheres of two different sizes and target strengths. Multivariate analyses were used to evaluate the parameters of emitted signals. Duncan's multiple range tests showed significant decreases (df = 185, p less than 0.05) in both source level and bandwidth in the transition from detection to discrimination. Analysis of variance revealed a significant decrease in the number of clicks over test conditions [F(5.26) = 5.23, p less than 0.0001]. These data suggest that the whale relied on cues relevant to target shape as well as target strength, that changes in source level and bandwidth were task-related, that the decrease in clicks was associated with learning experience, and that Pseudorca's ability to discriminate shapes using echolocation may be comparable to that of Tursiops truncatus.

Contemporary anti‐invasion mines have aspect‐dependent shapes, making discrimination of mines from nonthreat objects, and classification of mines, a difficult task for human mine countermeasures personnel. Shallow‐water (SW) noise and bottom reverberation substantially degrade the acoustic structure of mine echoes. Traditional MCM target strength and FFT classifiers are not effective under these conditions. The testing of a novel neural network classifier to solve the task of classifying mines in the SW acoustic environment was begun. Three mine types were ensonified at 1 deg rotations using synthetic dolphin clicks. A learning vector quantization network was trained to classify the mines using the Choi–Williams joint time‐frequency distributions of echoes. A training set of 36 echoes per mine (5, 15, 25,..., 355 deg) was created, with the remaining echoes (0, 10, 20,..., 350 deg) reserved for generalization testing. The network correctly classified the mines at novel orientations with 85% accuracy. Perfo...