University Place

cc >> Welcome everyone to Wednesday Nite at the Lab. I'm Tom Zinnen. I work here at the UW-Madison biotechnology Center. I also work for UW Extension Cooperative Extension. On behalf of those folks and our other organizers, Wisconsin Public Television, the Wisconsin Alumni Association and the UW-Madison Science Alliance, thanks for coming to Wednesday Nite at the Lab. We do this every Wednesday night, 50 times a year. Tonight I'm delighted to be able to introduce to you Dr. Mark Berres. He's an assistant professor of avian biology here at UW-Madison. His research interests include the population genetic consequences of rarity in birds. Also, genomic approaches to enhanced disease resistance and productivity in poultry. Mark is a dedicated field researcher and has active projects throughout the world, including most recently Vietnam where he and his students are studying he effects of avian influenza on native pheasant populations, including the red jungle fowl, which is the direct ancestor of the domestic chicken. He is the instructor for many courses at UW-Madison including ornithology, birds of southern Wisconsin, avian physiology, an introduction to museum studies. He's also going to talk to us tonight about his new app called WeBird and the database that goes with it. One of the great things about this subject is that it's one of the fun-ist things to tweet about on twitter, because you get to use Tweety Bird and many other things. It's pretty amazing to think that in a few months, hopefully, we'll be able to hold our iPhones and Smartphones up to a bird and be able to identify it. I think that's just part of what Mark does, but it's certainly of interest to me, and I hope to many of you. Please join me in welcoming Mark Berres to Wednesday Nite at the Lab.

applause

>> Great, thanks Tom, and thank you for all attending this evening's presentation. I want to talk to you a little bit about WeBird, which has been a project that I've been working on, kind of on and off, and unofficially in my laboratory for about the past two years. While I wish a few months in the future, as Tom indicated, would be the release date, it'll probably be a little bit longer. But I'll have more to say about that, and in fact, this is a bit of science that I'll encourage you to actually help make it a reality. As you all know, undoubtedly so, I'm going to have a lot of avian bias here. So you it might not be surprising to hear me say that birds are a priceless part of our world's heritage. With no other animal has our relation been so connected over all throughout our human history. This is really not surprising as birds themselves are extraordinarily conspicuous. They're found everywhere throughout our planet. In part because of our close connection with birds our knowledge of them is more complete than compared to most other animal classes. Specifically, our understanding of avian population and community ecology, has increased to the point where birds are now accurate predictors of the health of our environment. They are also very successful in providing awareness to conservation issues. Birds are also extraordinarily and economically important. In the United States alone at least 48 million people participate in birding and birding-related activities. They contribute more than 82 billion, that's billion with a B, dollars in annual economic output. The magnitudes of these numbers, which are likely under-estimated, suggests that birders in the United States, and also elsewhere throughout our planet, could influence positive changes in cultural and political perceptions, of not only birds, but also of biodiversity in the environment as well. Yet in the United States, when at least compared to our neighbors in the neotropics, the majority of the nearctic is comparatively species poor. Nevertheless the nearctic is still home to a tremendous amount of diversity of native birds with nearly a thousand species inhabiting terrestrial, coastal and oceanic habitats. But alarmingly, nearly one quarter of bird species in just the United States alone have shown a steady population decline over the past 40 years. This really is a warning signal of the failing health of our ecosystems. Among these birds, 67 are currently federally listed as endangered or threatened. And at least 184 more species are a concern because of their restricted distribution, high threat risk or have already demonstrated early indications of population declines. The declining populations of birds is not a phenomenon only restricted to the United States. Thus, successful conservation of birds, I will argue, will require accurate information about the population status of every species, not just those within the United States. This is to ensure the survival of not only endangered species, but also those birds not considered currently to be threatened at all. But rather ironically, despite the declining trend in many bird populations, there's actually a significant growth of the popularization of bird watching. This is not just a phenomenon restricted to, again, to the US, but world-wide. A highly-publicized search for the ivory billed woodpecker that occurred in 2002, a popular field guide, Sibleys' Guide to the Birds of North America, actually makes the New York Times best-seller list. And hundreds of annual birding festivals held world-wide are genuine indicators of birding ardor. But while bird watchers, and more generally, nature enthusiasts, invest significant amounts of time, effort and money into their activities, threats against birds in particular, and the environment in general, continues. Before we talk about WeBird I want to ask an obvious question. Why is this? Apart for very important social and economic issues, I personally assert that a fundamental lack of biodiversity awareness is the root cause. And however complicated it may be, I certainly would argue that this result arises from a profoundly simple fact. Most people cannot identify animals and plants, even those that would be considered very common to their geographic area. One particularly telling paper that was published back in 2002 was quite telling of this phenomenon. In this paper the authors had established that learning to identify objects, in other words, species, is not limited to adults in general. And certainly not to children specifically. In this study, the authors showed accurate visual identification of nearly 120 Pokemon characters by children ranging in ages around five to eight. To give you an idea of the complexity of this task, these are some of the images of Pokemon characters that were able to have identification rates of 80% by five and seven year olds. That's pretty impressive. When you consider the variation in often very close color patterns and shapes and phenotypes, you might think that this probably presents a more arduous task than identifying birds in general. When I look at those, you know, it's very difficult for me to ascribe a particular name to them. But kids themselves do it very, very well. The results of the findings from that particular study, and others, indicate that children learn far more about Pokeman then about their native wildlife, even common native wildlife, in their area. The result is they actually enter secondary school being able to name fewer than half of plant and wildlife species common to their area. That's a subtle distinction. It's not a percentage of all of the plant and native wildlife species, but it's those that would be considered common, which is much, much less. Other evidence links loss of knowledge about the natural world to growing isolation from it. An increasing number of adults and children have been demonstrated, for those living in urban environments, show that they too have very poor plant and animal identification skills. This is consistent with an interpretation of a strong disconnection from the natural, and particularly local, environment. In conclusion, from these and a mounting series of other studies, are converging onto a common theme. Environmental awareness must first focus on local biodiversity. But I would also argue that people only care about what they know. So the ideal recollection for any bird watcher would be to be able to look at a bird such as this Baltimore oriole, perhaps recall the Latin, assign a gender to it and some plumage characteristics, and maybe some more broad-reaching classifications, such as, this is a New World oriole. Also, if you're really adept, knowing the song, and also the range of this oriole. It would really make you very adept, particularly if you could do this for all thousand or so species that are present within the United States. But there's a subtle problem here, but it's just staring us right in the face. How do we solve this? Well, first published in 1934, Roger Tory Peterson's A Field Guide to the Birds really was the first full-scale uniform method to identify birds by visual comparisons of so-called field marks. In other words, external phenotypic characters. His success really was based upon the usage of certain diagnostic plumage characteristics and other morphological features that differed from species to species. The success of that initial endeavor has really capitulated an enormous amount of published literature regarding bird identification. Many publishers have since followed suit, all of which are based on visual comparisons of field marks. At least, and there are some new authors coming out with a different perspectives, which I think probably appeal more to advanced birders, but nevertheless, in all cases, they use the same approaches advocated by Peterson, some quite awhile ago. Almost as popular as printed materials are CDs of vocalizations. And certainly, to be most contemporary, the increasing availability of digital media has prompted the majority of these publishers to also provide electronic versions of their printed and audio materials. These digital versions offer additional features certainly not available in print. For example, Green Mountain Digital, and these are the publishers of the electronic versions of the Audubon field guides series, they've got a new update to their app where EBird can be queried directly and can be GPS-aware. So you can just hit, what are the birds around me? It queries the EBird database, and in just a few minutes you're going to have a list of those species that are present within some specified distance of your location. It sounds pretty cool, right? Everybody should be an expert birder with all this material that's been available for more than 80 years. Well, not so. In fact, despite high participation rates and significant expenditures of bird watchers, combined with the pervasive nature of field guides and associated high-quality online materials dedicated to birds and also more app-like mobile resources, most enthusiasts of birds actually remain poorly adept at bird identification. On average, less than 20 species can be identified visually by a majority of birders. The number by sound is considerably less than that. Because of this, I posit, that not knowing the name of a bird and therefore recalling some information about it that would at least satisfy the curiosity of the individual engaged in that activity, creates frustration and quickly eliminates a desire to obtain additional information about it. Observations made in Birds of Southern Wisconsin, which is one of my upper division courses that I teach. Each spring we teach about 80 students. And without exception, many students simply indicted very clearly to us when asked, what is the biggest hurdle for you to achieve to be able to learn all of the birds and identify them visually and also audially? They respond by saying, I don't know where to start. That's despite all of the large amounts of materials available to help. Thus, electronic field guides have no definitive advantage over printed version. Access to even more sophisticated information sources appears to even further exacerbate this problem. Well, how do we solve this now? Adults and children are strongly influenced by the media, undoubtedly. This is a phenomenon that is known to extend into attitudes concerning wildlife. Logically, if the proliferation of electronic media, and certainly devices such as these continues, and environmental education is deemed essential to attenuate the continuing trend of biodiversity loss, we must consider the capabilities of specific media and use them appropriately. Because failure to do so can have unintended consequences early in life, all of which are more and more difficult the older one becomes. For example, most information acquired by school-aged children these days about how animals are, or operate or behave or just any information in general, is actually obtained from the internet, from website information. So research has demonstrated that exposure to this type of media created strong biases where considerations to actually protect a species were actually driven only by a few iconic and exotic species, the so-called charismatic mega-fauna. The diversity of species that should be protected, that was considered by the children, was very, very small, and definitely not local. Well, I am taking a new direction toward enhancing, just bird identification skills at this point, by attempting to create an application called WeBird which stands for the Wisconsin Electronic Bird Identification Resource Database. WeBird is simply a mobile Smartphone application designed to identify birds automatically by their vocalizations. The basic idea behind this is actually fairly simple. Information that's contained in bird vocalizations can be used to identify individual species. The fact of that matter that so many of us now have mobile computers makes the possibility of realtime identification while engaged in bird watching activity possible. Getting back to the identification procedure itself, if you take a look at the diagram presented to you here. This graph actually shows a power spectrum of typical vocalizations made by three species. We've got a burrowing owl, an olive-sided flycatcher and a Cape May warbler. In this particular example the audile frequency of species are fairly well separated from each other, and they also have power signatures that differ as well. WeBird itself uses this and other types of information to ascribe a particular vocalization pattern to a particular species. Or, as we will see a little bit later, even individuals within a particular species. Despite the complexity of this diagram, the use of WeBird is actually fairly simple. While in the field a user would simply record a bird vocalization which is then transmitted over a local data network to a remote computer where all of the calculations are conducted. We originally tried to do such calculations on the iPhone itself, because honestly these things are like mini-supercomputers. But it was-- It could not be done. So we do need access to a data network. Nevertheless, WeBird performs certain types of digital signal processing analyses that characterize audio features such as frequency, power and time. It essentially creates digital fingerprints that can be compared to vocalizations of other known species which are stored in the database. So the degree of similarity among candidate fingerprints is then evaluated with a statistically based significance test. The comparison with the best overall significance is simply chosen as the correct identification. So if a statistically significant match does happen to be found an identifier is simply transmitted back to the Smartphone and is displayed with imagery and text. Detailed descriptions and additional supporting information, such as range maps or more audio or imagery is also made accessible from a pre-loaded database that actually resides on the Smartphone. To give you a better look at what WeBird does, I want you to first take a look at these two spectrograms which are simply a digital representation of audio vocalizations of two Cerulean warblers. Let's take a look. These two birds were actually from a recording made in Maryland. It's the same individual on the top and the bottom, but as you can see, there are actually subtle differences in the call. Let's just play these. Sorry. Let's play these things.

bird chirping

You can't hear that, can you? Do we have an audio input?

inaudible

There it is. Let's see if it works.

bird chirping

I think it's worse. Is it possible to increase the gain on that?

inaudible

Sorry guys.

inaudible

That's not a bad idea.

bird chirping

The problem-solver. Great. Okay, so the Cerulean warbler is exhibiting its very characteristic territorial call.

bird chirping

What this recording is that you just heard was really song one, here, out of a bout of about 21 repeats of that particular song. So this was song one.

bird chirping

And this is song two.

bird chirping

A very keen ear would be able to tell, hmmm, yeah, they're very similar--

bird chirping

But, in fact, song 12 contains an extra note right here, an extra syllable, if you will. The point of WeBird is that, okay, it wants to perform some manner of identification using information that's simply contained in the power spectrum of these two calls. The way that it actually does that is that WeBird actually uses a series of algorithms that are able to accommodate the variation that's actually present in those two calls. Presently we use about seven acoustical measurements of the input data in order to accomplish this task. Without getting into the gory details, the matching score itself is really dependant on a degree of similarity. So WeBird, essentially through its comparisons of the two songs, computes a similarity score that's actually based on the number of changes that must be made in order to acquire the best theoretical match. If you look at this diagram up here, what we see on this axis is that this is song about number 12, and this is number 1, the two songs that you just heard. Now the representation here is essentially a colored diagram that represents a score. You can see we have negative scores and we can have positive scores. So if you were to essentially line up the two audio features together, and like I said, we're using seven of them, essentially a perfect match would be indicated by a presence of a comparison of a true similarity measure on this diagonal. So as the algorithm actually compares the early portions of each song, the similarities are actually very close to what maximally could be obtained. Thus we have very, very high scores. But at the point here you can see that it actually begins to diverge. This particular note that's present in this song is not present in that song. Therefore we have scores that are less than optimum. But yet, the point that WeBird is attempting make is that it tries to maximize the particular score at every position such that it is most like the optimal score. So there's manipulation of these seven audio features, both in terms of frequency and in time, and also intensity and other types of measures. A scoring system is applied. It's actually fairly simple in terms of an estimation of the score. The more changes that have to be made, or the more different any given comparison is provided, the poorer the score that is assigned. The algorithms are able to evaluate the best score deterministically. But it does take a lot of computational power. That's why we off-load all of the computations to a remote server. But in the end, in order to derive a score, you simply add up all of the scores along this black line here which corresponds to the best theoretical match that could be made given the two songs here. Our work with this actually suggests that, you know, it does a really pretty good job. Here's a situation unlike the first example where now instead of two songs from the same bird, now we have two songs from two different Cerulean warblers. Let's hear the one from Maryland.

bird one chirping

Now here's one from West Virginia.

bird two chirping

Do people buy that those are pretty different, but pretty the same? Yeah, I would. I would expect my students to be able to figure this out.

bird one chirping

And they do a good job.

bird two chirping

Okay. Well, if we look here along the same scoring space. Now I just omitted the diagonal of the perfect match just for ease of view. We see that we have the songs, they are traversing in terms of their matching scores through pretty high scoring spaces here. So between the Maryland and the West Virginia bird we actually achieve a very, very high score despite significant spectrally differences between these two songs. Now let's take a look if we do something completely different. Here's our Maryland bird again.

Bird one chirping

And now we're going to take a listen to a hooded warbler.

bird three chirps

Completely different. I hope everyone agrees. In fact, WeBird also agrees with that too. So if you look, yes, we do have matches here, but the problem is is that the are so different, the songs are so different spectrally to each other, from the metrics that we use, that the amount of change that it would require to make these things very similar, is just huge. In other words, it's unlikely to even happen. We can see here that the scoring space is significantly reduced. The overall match score itself is similarly reduced. But we're not quite there yet. I mean, yes, we can make this comparison. Now what we really want to ask is, well, is there any significance to this comparison of the difference between the two different Cerulean warblers, one from Maryland and one from West Virginia, and a Cerulean warbler from Maryland and a hooded warbler, which was also from Maryland. No big deal. To evaluate the match significance we actually employ a randomization model that allows computation of Z scores. In other words, a measure of statistical significance. In this example, the choice of the best match is really very clear. We're attempting actually to develop more exact significance tests that actually require less computational time. These particular tests, they can take a few minutes. In typical fashion, we actually ignore the confidence. When we're creating empirical estimates of differences in scoring values. So overall, the performance of WeBird is promising, but there's one caveat that is really going to be needed to be solved. That's where you guys come in. It's so good, I think, that we actually have evidence that the spectral analysis in matching models in WeBird are actually capable of distinguishing vocalizations even within individual birds. So the figure that what you're looking at now simply depicts a principle coordinate analysis plot of a spectral variation in the digital fingerprints for five singing bouts from two male Cerulean warblers. One in Maryland up here, that actually consists of four bouts, one, two, three, four. And also a bird in West Virginia that also had one bout of a number of songs. Overall, what you can see is that there's three clusters of these songs that are evident. The male from West Virginia, shown in blue at the bottom. And there's two additional clusters from the Cerulean warbler from Maryland. Each cluster is well separated from each other, and it's comprised of vocalizations made by each of the males in each singing bout. The largest cluster here that you see contains songs from three of the four bouts made by that male in Maryland. These were actually bouts that were recorded at dawn, mid-dawn and late dawn. The last cluster from that Maryland male, shown in light green, was actually recorded in late morning. Curiously, there appears to be structure between the three bouts of dawn song here, of dawn, mid-dawn and late dawn, on dimension three on the vertical axis. So while all of the songs that the Cerulean warbler was making in early morning were very, very similar, as evidenced by the tight clustering along dimension one and two, which are just variation axes. These also seems to be a strong vertical component which suggests that there is differential structure being used in those songs from those three time periods. Most significantly, what we see here is that during the late morning bout of that Maryland male, we see that there are a few bouts of a late dawn song which was performed earlier in the morning grouping with those songs that were sung in the late morning. This is actually evidence of a temporal shift in song type. This is a phenomena that's known to actually occur in Cerulean warblers, and many other species of birds. Like chipping sparrows for example. Early in the morning chipping sparrows tend to have a very higher frequency, a much more rapid and monotonous trill. More towards the afternoon they have a much lower frequency and more, what I would call, a lackadaisical monotonous trill. But nevertheless, this demonstration here shows that yes, WeBird performs especially well on very closely related individuals. It also works in terms of identifying songs from very different species as well. Generally speaking, media theorists submit that new media, and specifically mobil new media, affords educational and research experiences that in many cases were simply unobtainable. In the case of WeBird the novelty lies in the integration of automatic bird song recognition technology with crowd-sourced location-aware mobil devices. For informal users learning the species, names of all of these birds, and certainly other organisms, is a necessary yet difficult step to cultivate a positive relationship with nature, and more broadly, the environment. By providing feedback while actively engaged in bird watching I really hope that WeBird will help anyone use it, to be able to identify accurately and improve their identification skills of birds. And most importantly, to begin to foster a stronger connection with the environment. Since the vocalization data will have to be crowd-sourced, WeBird will provide scientists with a return of very, very high quality data that no team of professional researchers could ever hope to achieve. As more audio data of bird vocalizations is collected, the reliability and the scientific value of that data collected by WeBird also increases. Moreover, because WeBird actually collects and stores the audio data used to determine the species, the value of citizen-science collected data is really on par, precisely on par, with that collected by researchers and other scientists. This really has the important advantage of eliminating reliability concerns that many scientists and researchers have when using citizen-scientist collected data. The capabilities of WeBird will provide scientists with a new tool to also integrate into their research. Tasks ranging from remote field monitoring to fine-scale analysis of vocalizations will become more easy and more productive. We're attempting to pilot the use of WeBird through a collaboration among research scientists and informal education participants at the arboretum here on campus. We're preparing to deploy a beta version of this in the near future. But there are a few hurdles that we have to overcome. Please, somebody ask me during the Q and A period after the end of the talk. An improved method then, to identify and also monitor birds, such as I promote here, in the hands of the general public will not only increase the public's awareness of birds. Because of the fostering of the sense of satisfaction in terms of being able to identify that particular species of bird, is certainly going to enhance not only the casual user of WeBird, but certainly will be able to enhance scientific research as well. The end result from both types of usage, I believe, will be greater awareness, not only to scientists, but also the public, to the biodiversity concerns at both local and global scales. I appreciate your attention, and thank you for coming.

applause