Artificial Identification

The relatively recent rise in popularity of birding "tools" such as the Merlin Bird Identification app and Swarovski’s dG which identifies the wildlife it focuses on reflects a universal explosion in machine learning and artificial intelligence in general. Artificial intelligence or AI has taken over in most industries and is now used to complete incredibly complex tasks, ranging from automatically labelling CCTV footage to understanding and responding to human speech in Alexa and Siri. In some medical diagnosis tasks, machine learning has been found to perform better than doctors themselves.

None of these are a far cry from bird identification. If ANPR can identify the letters on a speeding car’s numberplate it won’t be long before a camera is able to reliably identify passing birds, especially given Merlin’s demonstrable identification skills.

Merlin relatively recently released the sound identification extension to its birding app which can identify (or at least give suggestions for) birds calling in real time. This is quite a remarkable ability. It is a tool which, at least in my experience, is not that widely used or known about in the birding community.
 
Is it something of an inevitability then that artificial intelligence will one day become a more commonplace tool for birders? Perhaps too, a weapon in the arsenal of conservationists? That weapon, however, may be a double-edged sword. I may not need to explain to many of you why artificial intellgience may not be the best thing for birding. Artificial intelligence has “artificial” right there in its name - it could jeopardise the natural wonder of birds; the wild, unpredictability of it all. Indeed, the Swarovksi’s dG was met with a lot of opposition amongst the birding community. If computers are identifying birds for us, we might lose the need for us to learn to do it ourselves...
 
But I don't think the situation is that bleak after all. It puts us on the edge of so many exciting opportunities for birding and conservation, and I don't think it will jeopardise any of our usual birding customs.
 
After all, the Merlin app has been around for many years, continually perfected and kept at the state-of-the-art, and yet it’s still no match for the majority of experienced birdwatchers. Indeed on the tough calls, the experts sometimes make it look like the birdwatching equivalent of a Commodore 64. There is no reason to think that further AI advances will obsolete anything at all. They will be tools admittedly (and exciting ones at that) but they won’t do the whole job for us. 
 
The thing about most modern-day machine learning algorithms is that they don't explain themselves. Sure, you might be able to build a tool which thinks it can identify Richard’s from Blyth’s Pipits, but it won’t be able to tell you why. And that, in birding circles, is a serious shortcoming. Trying to claim a BBRC record of an over-flying Calandra Lark, with the only field notes being “because it looks like one”, is ironically not going to fly, and it never will.

Merlin, for example, is based on a percentage match, and it is bound to make mistakes. This is simply an unfortunate fact of artificial intelligence algorithms. Provide it with a blurry indiscernible photo that an expert has dismissed as unidentifiable and an AI algorithm may come back with a 90% match to a Yelkouan Shearwater, with no explainable reasons for this decision. Literally modify one pixel and astonishingly, it could now be 90% certain it's a House Sparrow. Without going into the maths, machine learning models are unfortunately just vulnerable in this way to overconfidence and brittle decision-making.

Clearly, this is a Rock Pipit and the Merlin app would agree. However,  a deliberate but imperceptible change (known as an "adversarial attack") by someone who is trying to ruin the algorithm could cause a machine learning algorithm to believe this is a Puffin!

Critics of Merlin, and its siblings, use these errors and uncertainties as a reason to condemn these tools as a gimmick but I don't think that's very fair. After all, we don't need AI to make the difficult borderline calls; we have expert birders and rarity committees for that. AI will never be or need to be the final say on whether a rarity is accepted. 
 
AI identification could become hugely useful for so many other reasons. As a student studying maths and data science at university, this is an exciting fusion of two of my passions. My mind inevitably wanders in the many directions this untapped potential could go.

The ability to arm members of the non-birding public with a tool that can identify birds or sounds in their garden and alert them to an unusual visitor is an astonishing capability, one whose repercussions I think we are still yet to feel.

Imagine cameras placed at strategic locations inland, identifying the birds that they see. I would think huge numbers of rarities that ordinarily go unnoticed would be found. There is a school of thought that rarities should theoretically arrive somewhat evenly across the country – it’s just that at migration hotspots and offshore islands the number of birders and amount of available habitat are at such a concentration for them to readily be found. If this is true, then AI spotters would just mean more eyes and ears in under-watched locations. Sure, they will miss things and get things wrong but it will be an advantage over no eyes at all. A birder can then go and follow up a potential sighting. If a Red-flanked Bluetail can be picked up on a camera trap in a Worcestershire woodland then who knows what goodies lurk out there to be found.

 
And it’s not just rarities, this could give us a way of finding new breeding locations of endangered species that can then be protected.

With the improvement of camera technology, perhaps these programmes could even learn to read rings, and automatically send the results to ringing databases!

The opportunities are varied, but many amount to effectively increasing the number of birders out there: from spotting sites for sensitive species to protect to finding rarities to educating members of the public about the wildlife in their own back garden. 

Maybe this help in identification, during the early stages of the hobby, will help get new birders over that initial hurdle, when every bird is new and intimidating. If AI can be used to help engage more people in our hobby, at a time when increasing engagement in the environment is such a critical endeavour, then surely this can only be a good thing!

We won't need to convince people of the amazing and varied wildlife on their doorstep, because their phones will do it for us.