Video summary
The video explores significant and promising applications of artificial intelligence (AI) in wildlife studies, particularly through facial recognition technology used with camera traps. While acknowledging general concerns about AI, the speaker highlights its exceptional capability for pattern recognition as a tool to process vast amounts of data that would otherwise overwhelm human researchers spending hours reviewing footage. A key example presented involves capuchin monkeys, where an AI system was trained to distinguish between these primates and other animals like coatis or unrelated species in their natural habitat. This technology addresses the major challenge faced by field biologists: identifying individual wild animals without prior lab training, which is essential for consistent behavioral observation but difficult when dealing with transient populations of unknown individuals entering a study area.
To overcome this hurdle, researchers developed a two-phase experimental system that leverages AI to automate data collection in the wild. In the first phase, a machine equipped with a touchscreen and a banana dispenser was introduced to the monkeys; initially ignored by most, it eventually attracted an adventurous male who learned to interact with it for food rewards, prompting others to follow suit through observation or unique interactions like kissing the screen. Once the AI could reliably identify capuchins versus other species, researchers uploaded collected images back to the lab to train the system on individual faces. This allowed the machine to progress from simple detection to recognizing specific individuals, enabling a second phase where experiments could be deployed autonomously. In this stage, the device would not just dispense treats upon any touch but only when an identified monkey performed a specific task or met certain conditions, effectively turning wild primates into subjects for cognitive and social testing without constant human supervision.
The implications of such technology extend beyond primate research to broader ecological management and conservation efforts. The speaker notes that while primates are uniquely curious and willing to engage with experimental setups due to their intelligence, similar systems could be adapted for birds like cockatoos or even insects attracted to screens. Furthermore, the AI's ability to identify specific individuals offers practical solutions for population control issues, such as preventing accidental double-castration of cats in New Zealand by recognizing which animals have already been treated. Ultimately, this approach bridges a critical gap between laboratory knowledge and real-world animal behavior, allowing scientists to gather high-quality data quickly while observing how animals naturally act outside the controlled environment of a lab, thereby providing a more accurate understanding of their true behaviors and social dynamics.
Read the full video transcript
All in all, I think it's bad. But
>> [laughter]
>> There are there
certain applications of AI that I think
are really cool. And one of the big ones
in is in any sort of
facial or individual recognition in
studies of wildlife. Because camera
trapping, you know, we've had citizen
science where people spend hours and
hours like combing through camera trap
footage and they try to say like oh well
this raccoon looks like this raccoon. I
think it's the same guy. And we know now
that you can train AI to scrub through
just insane amounts of data like that
because that's what AI is good at is
pattern recognition. That's what it's
good for. It is not good at drawing new
conclusions. It is not good at
fabricating things, but it is very good
at pattern recognition. And so if you
can find a way to leverage that to help
with ecological study, there's a lot of
really cool applications for that. So
This is a study where they looked at
capuchin monkeys
or capuchins depending on how you want
to say it. You know, what I was raised
in the 90s and I said capuchin
>> in a zoo, so you actually think about
you think about these things.
>> Yeah. You know, it's it's whatever you
want it to be I suppose. But um
>> I say Indiana Jones monkeys. No, that's
not right.
>> Yeah.
>> [laughter]
>> Or you know, Marcel from Friends. Um
anyway,
So they trained this AI to uh
recognize among some different lab
monkeys, six lab monkeys. It was able to
recognize which of these six lab monkeys
it was looking at. Okay, so that's
pretty well established. AI's been doing
that for a long time. But then um they
kind of did this
two-piece AI design for camera trapping
and doing behavioral experimentation
with wild monkeys, which is something
that is really difficult to do.
Because usually you have to like have a
monkey in a lab, you have to train it to
use a touchscreen, you have to teach the
either the researchers or an AI which
monkey is which, then they they learn
how to use the touchscreen, you try to
teach them a task, they do the task,
then you can associate, okay, this
monkey does it this way, this monkey
does it this way, right? So, that's how
it works in a lab.
But in the wild, you could have 10 or
100 monkeys coming and going. Um you
don't know them as well because they're
wild monkeys. There might be a monkey
that you've never seen before and you're
like, is that Steve or is that a new
guy? I don't know. And so, it's like
really hard to to measure that behavior,
but also you don't get consistent um
exposure to the same monkeys in a over
and over. So, teaching them a behavior
and um uh
a behavior that then can lead to an
experimental process is really really
difficult. So, you know, that's like a
really long way for me to say, they
trained an AI to uh recognize what a
capuchin looks like as opposed to any
other animal.
And that is where they started was just,
you know, since we have all these
individuals, we might see some for the
first time. We can't train the AI on
individuals and name them. Like, they
just need to know what is a capuchin,
what is not a capuchin.
And so, if a capuchin comes up, it would
essentially turn on the machine. And the
expectation was that the monkey would
touch the touchscreen and they would be
delivered a little dried banana.
And from that, the
the first couple days, like no monkeys
wanted to touch it or go anywhere near
it, but eventually uh kind of um
a more adventurous larger male came up
and started like hitting the machine and
banging it on the back and trying to
figure out what was going on cuz it
smelled the the banana.
And eventually he touched the screen and
the banana came out and like, bam, I got
it. Other monkeys watched him doing
that, they were able to come up, figure
out how to do it. Some monkeys figured
out that they could kiss the touchscreen
and get a banana, which I think is so
funny.
>> Oh my gosh.
I would like a
I would like a banana.
>> Yes, exactly. [laughter]
Banana, please.
Um
but so then uh from that, they were able
to have this machine recognize
monkeys or capuchins versus other
animals like coatis and stuff like that.
But then they were able to based on
frequency have it start to recognize
individuals. And when they when the
machine recognizes an individual, it
could then
essentially deploy an experiment. Um so
it could recognize a condition and then
switch modes from just touch you get a
treat to some sort of condition and you
get a treat. So this is very
preliminary, but what's so cool about it
is that like
this is completely you set it up and you
walk away and it collects data. And it
can recognize capuchins versus other
animals. It can recognize individuals
when given enough exposure. And then it
can also do kind of a two-phased
experimentation approach. So it can have
phase one, which is to teach a wild
monkey how to use the machine. And then
the second is the actual experimentation
case. So
um it's a really cool kind of
complicated two-step system
um for using AI to study cognitive or
social or any number of abilities in
wild primates, which I think is so cool.
>> But it does also like I mean primates
are special in their cognitive faculties
and their curiosity and their uh
probably their willingness to engage
with this kind of a setup.
>> Yeah.
>> I mean you could talk about any other
animal and you know, a little wild boar
would come along and be like
you know,
not climb up onto the platform and not
be able to engage.
>> coatis came up on the platform quite a
bit. They recognized, oh, that's not a
capuchin and it didn't turn on the touch
screen.
>> And they didn't get any bananas.
>> Nope.
>> And nobody broke the box.
>> No, they and and coatis are related to
raccoons. Like they're
they could definitely do some damage,
but they yeah, it was it was well
proofed.
>> All right.
So, they were able to specify like what
animal they're looking at a particular
species because of
AI. The AI was able to
did it it did identify them eventually?
>> Uh so so basically after they collected
a bunch of information from this touch
screen just like touch it get a treat,
they were able to take that data, bring
it back to the lab, train the AI on the
monkeys it saw and put it back out
there.
So, the AI was able to categorize the
kind of the cataloged data that it
collected so that then it could have
this two-step system. So, in this case
we're looking at you know, Pedro shows
up and it's like, oh, I know Pedro.
Here, you can do this this test on like
pick the triangle or whatever, you know,
like
This is all very this is a proof of
concept. They didn't actually do any
data collection yet, but essentially the
machine could recognize, I know this
monkey. I'm not just going to give it a
treat if it touches the screen. I'm
going to give it a treat if it touches
the screen the right way or if it
>> What would you do?
What would you do if you were like just
going through, you know, your normal day
and there was a thing
that suddenly it was like
Yeah, oh, get a free thing if you do
this and you did it and like suddenly
it's really nice to you and eventually
you find out that it's learning that
you're Blair.
>> [laughter]
>> Are are about
>> Blair likes bananas.
>> Cuz that's what the internet does.
>> I know. They're making the
>> It sends you your targeted ads.
>> They're doing targeting. They're doing
They are doing targeting for primates in
a jungle.
>> Mhm.
Yeah.
>> Fascinating.
>> But I think it's great because like
Yes,
primates are special, but birds would
definitely do this. We know pigeons do
this, right? So, lots of bird species
would do this. I think quite a few other
mammal species would do this.
>> Who put it in Australia with all the
cockatoos?
>> Yeah. We know that um
uh certain bugs actually respond to
screens. We know this, right? So, like
there's there's a lot of potential for
uh for use. Another thing that occurred
to me is like
in um New Zealand where they're trying
to um
chemically castrate cats.
>> [laughter and gasps]
>> Um they'd be able to recognize Oh, I
know this cat. I already got this one.
>> Don't accidentally do it again.
>> Yeah.
>> [sighs]
>> Double castration, man.
>> Yeah. But um you know, it's just I think
it's a it's a neat use for all the you
know, the trash talk I do with AI. I I
do think this is a really cool way to
gather a lot of data really quickly and
also bring behavioral experimentation
into the wild, which is a big missing
missing piece of a lot of researches. We
know how animals act in a lab, but we
know that's not also how they act in the
wild. There's a disconnect.