Sophie Bushwick: Can you tell what a cat is thinking, just from looking at it?
Tulika Bose: Probably not! Cats evolved to be solitary hunters stalking their prey, not social animals like us humans.
Bushwick: And that poker face might be handy while you’re out stalking prey, but it’s a real problem if humans are trying to figure out whether their pets are in pain. So that’s why researchers are trying to read more into feline feelings
Bose: — [sings] using artificial intelligence!
Bushwick: —[sings] using artificial intelligence! As they love to do. I’m Sophie Bushwick, tech editor at Scientific American.
Bose: I’m Tulika Bose, multimedia editor.
Bushwick: And you’re listening to Tech Quickly, the version of Scientific American’s Science Quickly podcast that just wants to boop AI on the nose.
Bose: Cute.
[INTRO MUSIC]
Bushwick: Right now, if you want to see whether a cat’s in pain, you have to take a quiz called the Glasgow Feline Composite Measure Pain Scale.
Bose: Interesting.
Bushwick: It’s also just called Glasgow scale.
Bose: Ok, hmmm.
Bushwick: But The question is, what’s on the scale?
Bose: Yeah, what is on the scale?
Bushwick: Well, I just happen to have printed out a couple copies for us to peruse!
Bose: You did not.
Bushwick: I did indeed.
Bose: Wow
Bushwick: Take a look, take a look at this.
Bose: This looks like a Buzzfeed quiz for cat pain, but uh —
Bushwick: Yeah, the least favorite kind of Buzzfeed quiz.
Bose: Yeah, the least favorite. Ok, wow. So I see some little cartoon cats here.
Bushwick: Uh huh.
Bose: And I see their little ears squashed, or upright —
Bushwick: — And then you have to say which one is your cat.
Bose: Yeah! And then, I see uh, a second little set of cartoon cats here, and their little mouths look to be squished in different ways too.
Bushwick: Yeah, so you choose the picture that best matches the cat. And you also answer questions about its behavior —
Bose: — The cat —
Bushwick: How it responds to being pet —
Bose: Ok, ok.
Bushwick: And then each one of those questions gets a little number associated with it. Higher or lower, depending on how much pain it indicates. And then you add it all up at the end to get your pain score.
Bose: [Flips page.] This looks kind of hard actually.
Bushwick: It’s tricky! I mean can you tell the difference between a cat that is tense / crouched —
Bose: No!
Bushwick: And one that is rigid / hunched?
Bose: No that would be like the same thing to me!
Bushwick: It sounds exactly the same.
Bose: Yeah, exactly!
Bushwick: So it’s pretty subjective, you have to make some judgment calls…
Bose: mmhmmm
Bushwick: And if you’re not a veterinarian, who sees cats and uses this scale all the time, it could be an issue.
Bose: Absolutely. More subjective right?
Bushwick: Mmhmm exactly.
Bose: So the idea is that AI might be able to do better?
Bushwick: Or at the very least, it could do it faster and way more easily than an untrained human could.
Bose: So I know freelance science writer Leila Okahata wrote about a study that put AI to the pain-detection test.
Bushwick: That’s right. Researchers started with photos of 84 cats that were being admitted to a veterinary/animal hospital in Germany —
Bose: Interesting —
Bushwick: — and then they gave these cats a pain score based on the Glasgow scale that we talked about —
Bose: Right —
Bushwick: And also on the amount of pain you’d expect a cat to be in for whatever ailment or injury brought them in, you know like a bone fracture. That’s gonna be pretty painful.
Bose: Yowch.
Bushwick: You mean “meow-ch”?
Bose: Uh, no, no I don’t.
Bushwick: [Laughs]
Bose: Ok. diabolical laugh. Diabolical laugh.
Bushwick: I’m not laughing at the cat’s pain. I’m laughing at the cat’s pun.
Bose: Uh, wow. Did you just make another pun?
Bushwick: I can’t help myself.
Bose: Oh my gosh, ok, ok.
Bushwick: [Continues laughing.]
Bose: Ok, ok.
Bushwick: I’m just littering the script with them! [Continues laughing.]
Bose: Yes you are!
Bushwick: [Continues laughing.] Littering? Litterbox?
Bose: Oh my God. You did it again.
Bushwick: Yeah. Anyway, the photos are labeled with the pain scores, and the researchers fed them into two different machine learning algorithms.
Bose: Interesting.
Bushwick: One broke down each photo in great detail, by examining four dozen specific sites on the cats’ faces. These are called called landmarks.
Bose: Got it.
Bushwick: The second AI program was a deep learning model that just looked at each photo as a whole.
Bose: Got it. Moment of truth though — could the AI read the cats’ faces?
Bushwick: It did pretty well!
Bose: Huh!
Bushwick: The landmark-based model was 77 percent accurate —
Bose: Interesting —
Bushwick: — and the deep learning one was a little less good. That was 65 percent accurate. But they think the deep learning one could improve if it has more photos to train on—this particular kind of model is really data hungry.
Bose: It needs more pictures of cats.
Bushwick: Exactly. Everyone needs more pictures of cats.
Bose: We all do.
Bushwick: Including AI.
Bose: Right! So can we learn anything from the AI about reading our cats’ faces?
Bushwick: Yeah, that’s one of the cool things about this study! So it turns out that the eyes are not the windows to your cat’s soul. When it comes to accurate pain recognition, you need to look at your cat’s mouth.
Bose: It’s meowth.
Bushwick: Aaaaah! You did it!
Bose: I did it. Can we go beyond pain here? Like, is there an AI that can tell me if a cat is happy, or sad, or about to knock a glass of water on my laptop?
Bushwick: That’s the eventual goal! A researcher who’s not involved in this AI study recently published a paper showing that cats can produce almost three hundred facial expressions.
Bose: Yeah like I knew this cat that had this expression when it was going to take a bite of MY sandwich—
Bushwick: [laughs] That’s very specific.
Bose: Yeah. But I feel like that was its sandwich-eating, sandwich-stealing —
Bushwick: Sandwich-eating sandwich steal-ing face. The classic expression.
Bose: Yeah! It’s like I’m going to swipe your food, stupid human.
Bushwick: [laughs] I mean I feel like all cats have that expression at all times.
Bose: Yeah! Just like, hello, dumb human. Feed me. Give me your stuff.
Bushwick: [laughs]
Bose: I’m taking your bed.
Bushwick: [laughs] — and your laptop.
Bose: — everything that you own.
Bushwick: — Everything that you own is mine now.
Bose: It’s mine.
[CLIP] Cat meow
Bushwick: So yeah, once you know that you can recognize these different expressions,
Bose: Right.
Bushwick: This researcher wants to join forces with the AI researchers to create a model that can detect more specific cat emotions based on whatever face it’s making. The goal is to be able to look at an image of the cat and have the AI detect whatever emotion that cat is feeling.
Bose: Got it, got it. Ok — But like I’ve heard that emotion recognition AI is really flawed though.
Bushwick: Oh, for sure. For instance, what do you think I’m feeling now? [makes face]
Bose: Hmmm. [pause] Sad. You look sad.
Bushwick: Yeah, Wrong! I’m feeling smug!
Bose: Oh!
Bushwick: — and sneaky, because I just tricked you into thinking I’m feeling an emotion that I’m not.
Bose: [laughs.] Wow, a politician.
Bushwick: I mean think about it! If I just smile I could be like I’m being a politician, I could be saying ‘oh I feel happy’ or I could be saying ‘Im just being polite’ —
Bose: Ohhh.
Bushwick: Or it could mean, I’m holding in a fart right now!
Bose: [Laughs.]
Bushwick: And the AI — it can’t tell that we’re a bunch of lying liars who lie like that.
Bose: Interesting! So the idea is that cats don’t lie about their emotions the way we humans do?
Bushwick: Well, it’s more that with cats and with other pets we have no way to ask them how they feel
Bose: Right.
Bushwick: So even an imperfect tool, like an emotion recognition AI — it’s valuable because it can help us cross that communication barrier.
Bose: Got it.
Bushwick: The researchers also want to expand their work to study emotions in other animals.
Bose: Right.
Bushwick: We’ve also talked on this podcast about using AI to decipher wild animal communication.
Bose: Interesting!
Bushwick: So we might just be a few datasets away from letting our pets practically “talk” to us.
Bose: Ok that would be so cool!
Bushwick: Yes. I want my pet to talk to me.
Bose: Same!
Bushwick: With its furry little face.
Bose: Um — I would love to know when my sandwich was about to be stolen —
Bushwick: [laughs]
Bose: Because this has happened numerous times, Angela!
Bushwick: I do worry that when I think my cat — my dog is gazing lovingly into my eyes, it’s really just like [give me a treat!]
Bose: Give me food! Do we really want to know? Do we really want to know.
Bushwick: Yeah. I mean once we open up that Pandora’s catbox —
Bose: Oh my God, litterbox, catbox.
Bushwick: Schrödinger’s box? Let’s not mention Schrödinger. That guy was a jerk.
Bose: Yeah, no. Furry. Furry little box.
Bushwick: [laughs]
Bose: [laughs]
[OUTRO MUSIC]
Bushwick: Science Quickly is produced by Jeff DelViscio, Tulika Bose, Kelso Harper and Carin Leong. Our show is edited by Elah Feder and Alexa Lim. Our theme music was composed by Dominic Smith.
Bose: Don’t forget to subscribe to Science Quickly wherever you get your podcasts. For more in-depth science news and features, go to ScientificAmerican.com. And if you like the show, give us a rating or review!
Bushwick: For Scientific American’s Science Quickly, I’m Sophie Bushwick.
Bose: I’m Tulika Bose. See you next time!
[The above has been a transcript of this podcast.]