{"title": "Natural Dolphin Echo Recognition Using an Integrator Gateway Network", "book": "Advances in Neural Information Processing Systems", "page_first": 273, "page_last": 281, "abstract": null, "full_text": "Natural Dolphin Echo Recog~ition Using an Integrator \n\nGateway Network \n\nHerbert L. Roitblat \nDepartment of Psychology, University \nof Hawaii, Honolulu, HI 96822 \n\nPatrick W. B Moore, Paul E. \nNachtigall, & Ralph H. Penner \nNaval Ocean Systems Center, Hawaii \nLaboratory, Kailua, Hawaii, 96734 \n\nAbstract \n\nWe have been studying the performance of a bottlenosed dolphin on \na delayed matching-to-sample task to gain insight into the processes and \nmechanisms that the animal uses during echolocation. The dolphin \nrecognizes targets by emitting natural sonar signals and listening to the \nechoes that return. This paper describes a novel neural network \narchitecture, called an integrator gateway network, that we have de(cid:173)\nveloped to account for this performance. The integrator gateway \nnetwork combines information from multiple echoes to classify targets \nwith about 90% accuracy. In contrast, a standard backpropagation \nnetwork performed with only about 63% accuracy. \n\n1. INTRODUCTION \n\nThe study of animals can provide a very important source of information for the de(cid:173)\nsign of automated artificial systems such as robots and autonomous vehicles. Animals \nhave evolved in a real world, solving real problems, such as gathering and interpreting \nessential information. We call the process of using animal studies to inform the de(cid:173)\nsign of artificial systems biomimetics because the artificial systems are designed as \nmimics of biological ones. \n\n273 \n\n\f274 \n\nRoitblat, Moore, Nachtigall, and ~nner \n\n2. INVESTIGATIONS OF DOLPHIN ECHOLOCATION PERFOR(cid:173)\nMANCE \n\nDolphin echolocation clicks emerge from the rounded forehead or melon as a highly \ndirectional sound beam with 3 dB (half power) beamwidths of approximately 10\u00b0 in \nboth the vertical and horizontal planes (Au, et al., 1986). Echolocation clicks have \npeak energy at frequencies from 40 to 130 kHz with source levels of 220 dB re: 1 JJ Pa \nat 1 m (Au, 1980; Moore & Pawloski, 1990). Bottlenosed dolphins have excellent di(cid:173)\nrectionally selective hearing (Au & Moore, 1984), spanning over 7 octaves, and can \ndetect frequencies as high as 150 kHz (Johnson, 1966). \n\n3. BEHAVIORAL METHODS \n\nWe have been studying the performance of a bottlenosed dolphin on an echolocation \ndelayed matching-to-sample (DMTS) task (e.g., Nachtigall, 1980; Nachtigall, et al., \n1985; Roitblat, et al., 1990a; Moore, et al., 1990). In this task a sample stimulus is \npresented underwater to a blindfolded dolphin. The dolphin is allowed to echolocate \non this object ad lib. The object is then removed from the water, and after a short \ndelay, three alternative objects are presented (the comparison stimuli). One of these \nobjects is identical to (matches) the sample object, and the dolphin is required to in(cid:173)\ndicate the matching stimulus by touching a response wand in front of it. The object \nthat serves as sample and the location of the correct match vary randomly from trial \nto trial. \n\nRecent work has concentrated on performance with three sample and comparison \nstimuli: (a) a PVC plastic tube, (b) a water-filled stainless steel sphere, and (c) a solid \naluminum cone (see Roitblat, et al., 1990a). On average the dolphin used 37.2 clicks \nto identify the sample, and an average of 4.2 scans to examine the three comparison \nstimuli. A scan is a train of clicks to a single stimulus ended either by the initiation of \na scan to another stimulus or by a cessation of clicking \n\nThe dolphin's scanning patterns were modeled using sequential sampling theory (see \nalso Roitblat, 1984). Simulations based on this model provide a reasonably good ap(cid:173)\nproximation of the dolphin's performance (Roitblat, et a1., 1990a). The simulation \ndiffered from the dolphin's actual performance, however, in that it was less variable \nthan the live dolphin. We return to the problem of accounting for this difference in \nvariability below after considering some models of the details of echo recognition. \n\n4. ARTIFICIAL NEURAL NE1WORKS \n\nWe have developed a series of neural-network models of dolphin echolocation pro(cid:173)\ncessing (see also Gorman and Sejnowski, 1988). We (Moore, et al., 1990; Roitblat, et \nal., 1989) trained a counterpropagation network (Hecht-Nielsen, 1987, 1988) to clas(cid:173)\nsify echoes represented by their spectra into categories corresponding to each of the \nstimuli in our current stimulus set. The network correctly classified more than 95% of \nthese spectra. This classification suggests two things. First, the spectral information \n\n\fNatural Dolphin Echo Recognition Using an Integrator Gateway Network \n\n275 \n\nOUTPUT \n\nfEAT\\JA\u00a3 \n\nGATEWAY \n\n\u2022 \u2022 \u2022 \n\n\u2022 \u2022 \u2022 \n\n\u2022 \u2022 \u2022 \n\nFigure \n\n1. \n\nA \n\nschematic \n\nof \n\nthe \n\nIntegrator Gateway Network. \n\npresent in the echoes was sufficient to identify the targets on which the dolphin was \necholocating. Second, only a single echo was necessary to classify the target. Al(cid:173)\nthough the network could identify the target with only a single echo, the dolphin con(cid:173)\ncurrently performing the same task emitted many clicks in identifying the same tar(cid:173)\ngets. Further investigation revealed that the clicks emitted by the dolphin were more \nvariable than our initial sample suggested (Roitblat, et a1., 1990b). This variability \nprovides one possible explanation for the high performance level, and low variability \nof our initial model. \n\n4.1 THE INTEGRATOR GATEWAY NE1WORK \n\nOur integrator gateway network incorporates features of the sequential sampling \nmodel described earlier, including the assumptions that the dolphin averages or sums \nspectral information from successive echoes and continues to emit clicks and collect \nreturning echoes until it can classify the target producing those echoes with sufficient \nconfidence. It mimics the dolphin's strategy of using multiple echoes to identify each \ntarget. Figure 1 shows schematic of the Integrator Gateway Network. \n\nNetwork inputs were 30-dimensional spectral vectors containing echo amplitudes in \n1.95 kHz wide frequency bins. The echoes were captured and digitized during the dol(cid:173)\nphin's matching-to-sample performance. In addition to the 30 bins of spectral infor(cid:173)\nmation, each echo was also marked as to whether the echo was (1.00) or was not \n\n\f276 \n\nRoitblat, Moore, Nachtigall, and ~nner \n\n(0.00) at the start of an echo train. Recall that the dolphin directs a series of clicks to \none target at a time, so it seemed plausible to include information marking the start of \na click train. The frequency inputs were then passed to a scalar unit and to the inte(cid:173)\ngrator layer. The integrator layer also contained 30 units, connected to the frequency \nunits in the input layer in a corresponding one-to-one pattern. The connections to the \nscalar unit were fIxed at lIn, where n is the number of frequency inputs. The weights \nto the integrator layer were fIxed at 1.00. The output of the scalar unit, i.e., the sum \nof all of its inputs, was passed to each unit in the integrator layer via a fIxed weight of \n-1.00. The effect of this scalar unit was to subtract the average activity of the input \nlayer (neglecting the start-of-train marker) from the inputs to the integrator layer. \nThis subtraction preserved all of the relative activity information present in the inputs, \nbut kept the inputs within a manageable range. \n\nThe elements in the integrator layer computed a cumulative sum of the inputs they \nreceived. The role of this layer was to accumulate and integrate information from \nsuccessive echo spectra. The outputs of the integrator layer were passed via fIxed \nconnections with 1.00 weights to corresponding units in the gateway layer. The inte(cid:173)\ngrator layer and the gateway layer each contained the same number of units. Each \nunit in the gateway layer acted as a reset for the corresponding unit in the integrator \nlayer, and connected back to its corresponding unit with a weight of -1.00. Each unit \nin the gateway layer employed a multiplicative transfer function that multiplied the \ninput from its corresponding unit in the integrator layer with the value of the start-of(cid:173)\ntrain marker. Because this marker had 1.00 activity at the start of a scan and 0.00 ac(cid:173)\ntivity otherwise, it functioned as a reset signal, causing the units in the integrator layer \nto be reset to 0.00 at the start of every scan; their previous activation level was sub(cid:173)\ntracted from their input. \n\nThe output of the integrator layer also led via variable-weight connections to each of \nthe elements in the feature layer. The same kind of scalar unit that intervened be(cid:173)\ntween the input layer and integrator layer was also used between the integrator layer \nand feature layer to subtract the average activity of the integrator layer, again to keep \nactivations within a manageable range. The outputs of the feature layer led via vari(cid:173)\nable-weight connections to the classifIer layer. The elements in these two layers con(cid:173)\ntained sigmoid transfer functions and were trained using a standard cumulative back(cid:173)\npropagation algorithm with the epoch duration set to the number of training samples \n(60). \n\nThe training set consisted of six sets of ten successive echoes each, selected from the \nends of haphazardly chosen echo trains. An equal number of cone, tube, and sphere \nechoes were used. The training set was a relatively small subset (4%) of the total set \nof available echoes (1,335). \n\n4.2 INTEGRATOR GATEWAY RESULTS AND DISCUSSION \n\nFigure 2 shows the results of generalization testing of the network in the form of a de(cid:173)\nrived confidence measure. The network was given all 30 scans (10 scans of each tar-\n\n\fNatural Dolphin Echo Recognition Using an Integrator Gateway Network \n\n277 \n\nSphere \n\nCcmI \n\nTube \n\ntf:9 \n\n9JtCIII1Yl Echoes \n\no \ni \u2022 \nr:t: I.' \u00a7 \u2022.\u2022 \n,,'\" \n1C \u2022\u2022\u2022 \no \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \n~ .~,~~~----~~--~------~----------~ \no i \nI \nQ: \u2022\u2022\u2022 \n8 \u2022.\u2022 \ni\u00b7\u00b74 \n~ \u2022\u2022 I \nfi I~---+--------------~----~---+----~--~----~ \n\u2022 \no \no i \n~ ... \n\nas '.4 \n~ \u2022. I \n~ .~~--~~~~~==~--~----------~--~ \n\u2022 \nO . \n\nIIcCIII1\u00a5l Echoes \n\nI \nII: \u2022.\u2022 \n\n.. \n\n., \n\n\u2022 \n\nSUccessive Echoes \n\n10 \n\n\u2022 \n\nI \n\n\u2022 \n\n\u2022 \n\n\u2022 \n\n\u2022 \n\n\u2022 \n\n\u2022 \n\n\u2022 \n\n\u2022 \n\n~ \n\n3D \n\nFigure 2. Results of generalization testing of the network in the form of the confi(cid:173)\ndence of the network in assigning the echo train to the proper category. See text. \n\nget for a total of 1,335 sequential echoes), and was required to classify each echo \ntrain. \"Confidence\" was defined as the ratio of the activation level of the correct clas(cid:173)\nsification versus the total output of the three classification units. A confidence ratio of \n1.00 indicates that only the correct unit is active. Confidence of 0.00 indicates that the \ncorrect unit is entirely inactive. Intermediate confidences correspond to intermediate \nlikelihood ratios (Qian & Scjnowski, 1988). \n\nRecall that echo trains varied in length under control of the dolphin. Therefore, it is \nnot entirely clear how to measure the network performance. According to sequential \nsampling theory (see Roitblat, et a1., 1990a) a rational decision maker collects echo \nevidence only until a sufficiently confident classification is available and then stops. \nTable 1 shows the number of clicks in each train that were required to reach a confi(cid:173)\ndence ratio of 0.96 and the classification that the network derived. Some of the scans \nended before the network could achieve this confidence level. Three erroneous clas(cid:173)\nsifications were made (90% correct). \n\n\f278 \n\nRoitblat, Moore, Nachtigall, and ~nner \n\nTable 1 \nNumber of Clicks to Network Confidence Criterion \n\nTarget Scanned \n\nSphere \n\nCone \n\nTube \n\nSphere \n\nCone \n\nTube \n\nIntegrator Gateway \n\nBackpropagation \n\nI6S \n9S \n7S \n6S \nI9S \n19S \n34S \n7S \n23C \n\n20C \n4C \n2C \n6C \nl4C \n6C \n6C \n4C \n6C \n3C \nllC \n\n40C \nl8C \n20T \n23T \n5T \n4T \n4T \n4T \n5T \n4T \n\nlS \n6S \nIS \n5S \nl4S \nI4S \n3S \nlS \n40T \n\nlC \n30C \nIC \n2C \n2C \n30Sl \n32Sl \n57S \n22S \n22S \n27T \n\n3S \nlS \nllSl \nIT \nl4T \nIT \nIT \nIT \n2T \nIT \n\nNote: Entries are the number of clicks needed by the network to achieve the 0.96 \nconfidence criterion. C indicates a Cone decision, S indicates a Sphere decision, T in(cid:173)\ndicates a Tube decision. lIndicates that the dolphin stopped echolocating before the \nnetwork reached its confidence criterion. On these scans, the decision is the one with \nthe highest confidence at the end of the scan. \n\n4.3 A SIMPLE BACKPROPAGATION NE1WORK \n\nThe integrator gateway network reflects the assumption of sequential sampling theory \nthat the dolphin combines information from successive echoes in deriving its identifi(cid:173)\ncation. In contrast, a standard backpropagation network does not integrate over suc(cid:173)\ncessive echoes, but instead attempts to identify each echo independently. A back(cid:173)\npropagation network can be used as a model of a system that emits multiple clicks be(cid:173)\ncause the echoes vary in quality. Rather than integrating the echoes, it simply waits \nfor a single adequate echo that allows it to meet its confidence criterion. \n\nWe trained a backpropagation network (using the fast-backpropagation algorithm to \nadjust the weights (Samad, 1988) on the same data that were submitted to the inte(cid:173)\ngrator network in order to determine whether the additional structure of the integra(cid:173)\ntor network contributed to its performance accuracy. The network contained exactly \nthe same number of inputs, hidden units, outputs, and adjustable connections as the \nintegrator network. The networks differed only in absence of the integration appara(cid:173)\ntus in the backpropagation network. \n\n\fNatural Dolphin Echo Recognition Using an Integrator Gateway Network \n\n279 \n\nSphere \n\n. .. \n\nCone \n\nSUccessive Echoes \n\n\u2022 \n\n.. \n\nSUccessive Echoes \n\nTube \n\nC \n\n!SO \n\nSUccessive Echoes \n\n\u2022 \n\n10 \n\nN- 1 \n\n&0 \n\n70 \n\n10 \n\n!III \n\n\u00bb \n\nso \n\n20 \n\n\u00bb \n\nFigure 3. Confidence of the backpropagation network in assigning the echo train to \nthe proper category as a function of the number of echoes received. \n\n4.4 BACKPROPAGATION RESULTS \n\nFigure 3 shows the confidence of the backpropagation network in assigning the echo \ntrain to the proper category as a function of the number of echoes received. Com(cid:173)\npared to the categorization performance of the integrator network, the backpropaga(cid:173)\ntion network was much more variable. As Figure 3 shows, the individual echoes were \nhighly variable, and frequently assigned to an erroneous category. \n\nThe performance of the backpropagation network when judged by the standards of \nsequential sampling theory are also shown in Table 1. This table shows the number of \nclicks necessary to first reach a classification with greater then 0.96 confidence. On \naverage the backpropagation network (11.57 echoes) reached its confidence criterion \nin the about the same number of clicks (t (df = 58) = 0.03, p> .05) as the integrator \nnetwork (11.67 echoes), but it produced more errors (X2 (df=1) = 5.96). \nThese data suggest that the integrator network added significantly to the ability to \nclassify sequentially produced echoes. By implementing a signal \"averaging\" mecha(cid:173)\nnism in the neural network the system could take advantage of the redundancy inher(cid:173)\nent in the use of multiple echoes from the same source and in the stochastic proper(cid:173)\nties of the noise in which those echoes are embedded. In contrast, the backpropaga-\n\n\f280 \n\nRoitblat, Moore, Nachtigall, and ~nner \n\ntion network is required to process not only the characteristics of the echoes them(cid:173)\nselves, but also the characteristics of the noise. This results in many spurious classifi(cid:173)\ncations. \n\nThe gateway integrator network adds a level of complexity to the standard backprop(cid:173)\nagation network architecture that contributes substantially to its performance. Its de(cid:173)\nsign is inspired by properties of the dolphin's performance (Nachtigall & Moore, \n1988) and it represents one step along a development path that seeks to include more \nand more of the mechanisms that we can identify from the neurobiology of echoloca(cid:173)\ntion and from the performance of dolphins in their aquatic environment. \n\nReferences \n\nAu, W. W. L. (1980). Echolocation signals of the Atlantic bottlenose dolphin \n(Tursiops truncatus) in open waters. In R. G. Busnel & J. F. Fish (Eds.) Animal \nsonar systems. (pp. 251-282). New York: Plenum Press. \n\nAu, W. W. L. & Moore, P. W. B. (1984). Receiving beam patterns and directivity in(cid:173)\ndices of the Atlantic bottlenose dolphin Tlll'siops tnmcatlls. JOllmal of the Acoustical \nSociety of America, 75, 255-262. \n\nAu, W. W. L., Moore, P. W. B. & Pawloski, D. (1986). Echolocating transmitting \nbeam of the Atlantic botllenose dolphin. JOllmal of the Acoustical Society of America, \n80, 688-691. \n\nGorman, R. P. & Sejnowski, T. J. (1988). Analysis of hidden units in a layered net(cid:173)\nwork trained to classify sonar targets. Neural Networks, 1, 75-89. \n\nHecht-Nielsen, R. (1987). Counterpropagation networks. Applied Optics, 26, 4979-\n4984. \n\nHecht-Nielsen, R. (1988). Applications of counterpropagation networks. Neural \nNetworks, 1, 131-139. \n\nJohnson, C. S. (1966). Auditory thresholds of the bottlenosed porpoise, Tllrsiops \ntnmcatltS (Montague) (Naval Ordnance Test Station Technical Publication No 4178). \nNaval Ordnance Test Station. \n\nMoore, P. W. B. & Pawloski, D. A. (1990). Investigations on the control of echoloca(cid:173)\ntion pulses in the dolphin (Tul's;opS tnmcatus). In J. Thomas & R. Kastelein (Eds.) \nSensory abilities of cetaceans. New York: Plenum. In press. \n\nMoore, P. W. B., Roitblat, H. L., Penner, R. H., & Nachtigall, P. E., Recognizing Suc(cid:173)\ncessive Dolphin Echoes with an Integrator Gateway Network. Submitted for publica(cid:173)\ntion. \n\nNachtigall, P. E. (1.980). Odontocete echolocation performance on object size, shape, \nand material, In R. G. Busnel & J. F. Fish (Eds.), Animal SOllar Systems, pp. 71-95, \nNew York, Plenum Press. \n\n\fNatural Dolphin Echo Recognition Using an Integrator Gateway Network \n\n281 \n\nNachtigall, P. E., & Moore, P. W. B. (Eds.) (1988). Animal sonar: Processes and \nperfonnance. New York: Plenum. \n\nNachtigall, P. E., Patterson, S. A., & Bauer, G. B. (1985). Echolocation delayed \nmatching-to-sample in a bottlenose dolphin. Paper presented at the Sixth Biennial \nConference on the Biology of Marine Mammals, Vancouver, B.C., Canada. Novem(cid:173)\nber. \n\nQian, N. & Sejnowski, T. J. (1988). Predicting the secondary structure of globular \nproteins using neural network models. Journal of Molecular Biology, 202, 865-884. \n\nRoitblat, H. L. (1984). Representations in pigeon working memory, In: H. L. Roit(cid:173)\nblat, T. G. Bever and H. S. Terrace (Eds.), Allimal cognition. Hillsdale, NJ: Erlbaum, \n79-97. \n\nRoitblat, H. L., Moore, P. W. B., Nachtigall, P. E., Penner, R. H., & Au, W. W. L. \n(1989). Dolphin echolocation: Identification of returning echoes using a counterprop(cid:173)\nagation network. Proceedings of the First International Joint Conference on Neural \nNetworks. Washington, DC: IEEE Press. \n\nRoitblat, H. L., Penner, R. H. & Nachtigall, P. E. (1990a). Matching-to-sample by an \necholocating dolphin. JOllrnal of Experimental Psychology: Animal Behavior Processes, \n16,85-95. \n\nRoitblat, H. L., Penner, R. H. & Nachtigall, P. E. (1990b). Attention and decision \nmaking in echolocation matching-to-sample by a bottlenose dolphin (Tursiops tnm(cid:173)\ncalus): the microstructure of decision making. In J. Thomas & R. Kastelein (Eds.) \nSensory abilities of cetaceans. New York: Plenum. In press. \n\nSamad, T. (1988). Back propagation is significantly faster if the expected value of the \nsource unit is used for update. \nInternational Neural Network Society Conferellce \nAbstracts. \n\n\f", "award": [], "sourceid": 331, "authors": [{"given_name": "Herbert", "family_name": "Roitblat", "institution": null}, {"given_name": "Patrick", "family_name": "Moore", "institution": null}, {"given_name": "Paul", "family_name": "Nachtigall", "institution": null}, {"given_name": "Ralph", "family_name": "Penner", "institution": null}]}