This is a download from the FM 89.9, the Business Station. The FM 89.9, the Business Station. This is Matt Spleyne, and I'm Richard Bradbury. We're exploring the freaky side of AI and robotics on Matt Spleyne.
As Matt Armitage tries to explain why science is slick. I was going to be worried when I use a word freaky and Matt in the same sentence. But Matt, let's get your scary AI stories out of the way first. I think they're all scary this week, but let's go with the scary robot story first.
So one of the ways that humans get fitter, of course, is to work out. We know we can run and walk and do yoga, and of course we can build our muscles by lifting weights. We take it for granted though that machines, robots especially, are made to be as fit and strong as they can be. So, you know, pretty much out of the box.
But what if we were able to give machines the same evolutionary tools that we have, that ability to get stronger and fitter to improve over time? Because it's something that we already expect from artificial intelligence. With AI, that's kind of the whole point. So why would we put limits and constraints on the machine bodies that we wouldn't and don't put on their intelligence?
Maybe it's because in any movie or TV show where that happens, things don't go well for people, Matt. Those are just cowardly details. You know, you've got to take risks in this life. A team of roboticists at the University of Chicago has come up with a soft gel that hardens like a muscle and a pressure.
Now, it's made out of a cellulose mixture that's embedded with nanoparticles of zinc oxide and when vibrated, the nanoparticles emit an electrical charge and that charge turns the gel into a polymer which gains strength as it's shaken. Wait, other than scaring people and giving you something to laugh about, what real world use does this have? Well, admittedly, the idea of robots working out in the park in the morning as they plot their machine uprising does kind of amuse me. But at the moment, the changes to those robot muscles are permanent.
The gel hardens in the way the movements saw the vibrations are made. So in human terms, it's a bit like someone who makes their shoulders and chest massive but doesn't bother with their arms. So the team is trying to engineer a way to make these changes reversible. Currently, robots using a gel can or do become stronger as they perform repetitive tasks.
But as I said, those muscles stay hardened to suit that specific task. If you can reverse that process, then those machines can shed that strength and they can rebuild it in ways that will then suit another task, which makes them, of course, much more adaptable. I mean, I've got a feeling that this episode is heading into unpleasant lands. Would you like a nice reassuring story about honeybees?
Absolutely, please, you know, sweeten me up a little bit, please. Okay, honeybees and sinister robots, it is. Of course. Now, we all know honeybees.
They fly around, they land on flowers and they collect the nectar, also gathering and spreading pollen as they go. Now, we've long thought that the way that these move from plant to plant and land on those plants is mainly due to instinct. Now, we understand some of the steps that are involved in that process. Now, this story and most of the stories for today come from New Scientist, by the way.
I'll put the links up on culturepop.com. So the insects use something called optical flow that enables them to get around. Essentially, this is the rate at which things move through their field of view. So the closer you get to something, the faster it moves through your field of view.
I mean, we do much the same. Signaling that now is the time that the bee should slow down and get ready to land on that delicious-looking flower. And then there's the secondary aspect, divergence. That's the change in size of the object as you move towards it.
And then that would allow them to land. Well, yeah. So it was thought that these two factors combined would provide enough information for the honeybee to land and do its thing. And that's how researchers at the Delft University of Technology in the Netherlands approached the problem when they attempted to mimic the behavior with drones.
See, I told you it was about scary robots. Only they've found that those optical cubes weren't enough. As the drones slowed and moved into land, they would simply oscillate. So they'd start to rise and fall above the thing they were starting to land on.
Even by recreating that divergence aspect, the optical flow wasn't providing the drones with enough information to actually settle and land. So what do we think the honeybees are doing? Well, obviously we don't know for sure. So the Dutch team cracked the problem by adding a machine learning component.
Now, the drone doesn't just judge distance. It learns about the surface it wants to land on. It might be a plant, the bark of a tree or a blade of grass. So it learns what that surface looks like at different distances.
And that starts to make sense of the oscillations the drones were doing because that behavior, that going up and down, can be used to build those memories about the surface textures from different heights. And with those combinations of information, the team was able to land the drone smoothly. So that's what the honeybees are doing? Well, we don't know, but the Dutch researchers can only conjecture.
They think the bees may do something similar. A team at the West Faean University of Applied Sciences in Germany is doing some tests based on those learnings. And they've observed that honeybees that enter a new environment go through what does appear to be a learning phase. So they also want to find out if the bees do their own version of their oscillating behavior above new plants, which may indicate that they're creating mental maps of surface texture in the same ways as the drones.
What would the next step of that understanding be? Be? Be? Well, that, yeah, yeah, you got me.
That brings me to another story, also about honeybees. So one of the big things that CES, the consumer fair in Las Vegas a few years ago, was insect-like robotics. And when I say big, I mean, actually genuinely big, they were enormous insectoid robots and the whole thing had to be housed outdoors in one of the car parks. So I only saw the videos and that was enough to frighten me.
It was like one of those Japanese monster movies from the 60s had come to life. But from a practical point of view, basing robots on insects and animals has advantages. Standing upright on two legs is incredibly difficult. Very few animal species can manage it for any length of time.
Yet for us, it's second nature. And our physiology has evolved so successfully that walking on all fours like a dog or a monkey for anything but brief periods is incredibly uncomfortable and painful. And what do enormous robot beetles have to do with honeybees? Well, you know that these are trick questions, don't you?
You know, this is where we transition from robots to artificial intelligence. In the same way that insectoid robots are easier to create than humanoid ones, the same may be true of artificial intelligence. The first wave of AI in the mid to late 20th century thought we could create intelligent machines by giving them a set of instructions to follow. But as we've subsequently learned, not every choice or task can be broken down into simple steps.
Some of them are hard to break down at all, such as getting an autonomous car to come to a halt at a stop sign because we struggle to articulate the steps that would lead us to stop. It's a lot more complex than we think it is, largely because, you know, we don't think about it at all. It's just something that we do. And the second wave of AI would be deep learning.
Yeah, in that instance, the machines are given a set of rules that allow them to learn how to complete a task. And that's something that has really exploded over the last 10 years as computing and processor speeds have taken it out of the realm of supercomputers and into the land of smartphones. And that's what enables us to have Siri and Alexa and Google Assistant in our pockets. But we're still left with a massive efficiency gap, if you will.
These computer brains require a lot more power and energy to compute than we do. And they can be fantastic at one task, say beating the world go champion and then completely flounder because you've changed a single variable like the number of squares on the board. And that sends them back to start that process of learning over again from scratch. The next wave of AI then would be modelling brains themselves.
Yeah, and as we don't really understand how the human brain works exactly, it may well be insect brains that we have to start with. So those Japanese monster movies you mentioned were actually creating those monsters. Humanity might be wiped out by 20-meter-long self-determining machine cockroaches. Well, I really wish I could say yes to that because it would give me a lot of pleasure.
But no, this comes from an op-ed in the New Scientist again by James Marshall from the University of Sheffield. He and his colleagues have reverse engineered part of the visual cortex of the honeybee as well as its navigation and memory centers. So you can see how this plays into that story. Right.
Yeah, the algorithms they developed are incredibly efficient. They use about 1% of the computing power of deep learning. And they run at about 100 times the speed. So this has enabled them to build a drone in their lab that is fully autonomous and can avoid obstacles.
And that speed and flexibility makes them much more able to deal with unexpected or unfamiliar situations, much as we talked about honeybees in unfamiliar landscapes, also learning their surroundings. What are the major advantages of this reverse engineering approach? Well, according to James Marshall, and this is why he thinks it might herald a new wave or a new way of approaching AI. It's because you aren't learning from the ground up as you do with deep learning.
Reverse engineering allows you to benefit from the millions of years of evolution that have made those brains so efficient. So in the same way that honeybee can mysteriously work out exactly how to land on a flower, we may be able to harness those powers and efficiencies for the next generation or step forward in AI computing. Hmm. When we come back then, reading your mind and your veins, you're tuning to Matt's Blaine here on BFM 829.9, the Business Station.
For more, BFM 89.9, the Business Station. BFM 89.9, the Business Station. Welcome back to Matt's Blaine. I'm not Matt.
I'm Richard Bradbury. Now, in the, if the first half of the show didn't scare you off, welcome back. What AI unpleasantness is up next? Around Christmas, we did an episode about AI in popular culture, especially art and music.
I think it was episode 1, 5, 2, that's entertainment for anyone who missed it. Was that the one where our chat box took over for a little bit of the show? Well, that's the one. We'll have to drag the Matt Bot Infinity and the Brad Bot Pro back to the show soon.
But there are lots of advances being made when it comes to music and AI. We'll have to look at the whole concept of making music, another creative concept that specifically targets and exploits the algorithms of the content platforms on a future episode. But this is a story about an AI that's been trained to look at your brain and identify what song you're listening to. So in a way, technology like this could feed into that boredom index that we talked about on the episode last week.
But of course, only if you were wearing smart brainwave sensing devices like the ones that the big tech companies are currently developing. I can't imagine anyone doing that. Yes, exactly. It comes back to something we've talked about in the show before.
You know, what happens when your thoughts become data, things that can be stored in mind by private companies. But, you know, obviously now it's not the time to go over that. So a team at Delft University of Technology in the Netherlands took EEG readings from 20 people. They listened to 12 different songs through headphones.
So cuttings from those readings were then used to train an AI alongside the music they corresponded to. They then took other segments of those readings that hadn't been fed to the machine to see if it could identify the music correctly. And how accurate was it? Scarily accurate, about 85% accurate.
The AI was able to identify patterns that corresponded to the samples that it had been fed. The research does make it clear that these results only hold for each individual, however. Using the patterns from one individual to identify the music someone else listened to returned a really low accuracy rate of only around 10%. Is that because we all experience music differently?
Well, the researchers think so. Music triggers a different emotional response in all of this. You know, the song that one person wants to walk down the aisle to inspires complete dread in somebody else. Right.
Assuming then that this research isn't just a feed-some corporate database. What insights are they hoping to get from this? Well, I think one of the reasons is to understand why music is so important to us. You know, why do we create music?
Why do we listen to it? How does it affect us? And I think they're hoping to discover some universal constants in the way that we respond to music. But music itself is really only one aspect of this whole kind of emotional model.
In the past, we've talked about AI being used to generate art. So recently, there's been a lot of talk about NFTs, non-fungable tokens, which can be used as a kind of crypto-based authenticity certificate by a new wave of digital artists. Now, that's all incredibly complicated. So we'll talk about that on an episode in a few weeks' time.
But what I want to get to now is the way that AI is muscling its way into the art appreciation scene. As in writing reviews about work? Well, not quite, but sort of if that's any kind of an answer. The traditional approach for an AI to look at an artwork would be to treat it like one of those capture verification modules that we all do online, you know, to try and describe what's going on.
So, for example, Vaz with 12 Flowers, lots of yellow. That might be an accurate description of Van Gogh's sunflowers, but it completely misses what we feel about that piece. So similar to what we were saying about music and the EEGs, we all respond to art differently. Researchers at Stanford University built a data set of artworks around 6,000 people were then asked to write their opinions about pieces of art in the Artemis data set and include a short description of the piece they were looking at.
Each piece was analyzed and commented on by at least five people, and that's the information that was used to train this AI. Yeah, but if everyone's response to an artwork is different, then kind of the machine just can't wait with writing anything about the emotions as long as the caption was factually accurate. Well, in theory, yes. You know, the team themselves acknowledge that, of course, there's no right answer.
It's very subjective. But you will probably see some trends emerge from that data. So the people that find Rothko's soothing, for example, the people who find Gilbert and George unsettling, not many people are likely to describe much as the scream as calming and peaceful. So what the team wanted to see was if the machine could pass off its opinions and descriptions as believable, you know, a kind of art history during tests.
And they found that the machine pen captions passed that during test about 50% of the time, which is actually quite a high bar for machines to pass as humans. So the Stanford team hopes that over time this is going to improve. The data has been made open source, so other researchers can now use and improve on the research that they've started. Right.
I mean, it's great that a machine can pass as a human art critic, but again, it comes back to why? Well, music and art are things that we have these very visceral reactions to those insights may make it easier for machines to interact with us in the future in ways that we find more palatable. You know, the old cliche of the robot saying to the injured human, but why does it hurt? You know, even if the machine doesn't understand the emotion itself, it will understand the causes and the outcomes of the emotion.
So it understands the environment of the emotion. So that the digital assistant of the future will understand what it means to have a hangover say, or even if you look at the kind of modeling scenarios to understand how the design of things or items in public spaces might affect us. But also, and I think this may be the one that people will be reaching for to enable machines to make creative works, or at least works that seem to be creative and that resonate with us. And again, that brings us back to what I was saying before about music producers who are creating music to fit the algorithm.
And this is the flip side of that. This is algorithms creating art to meet our emotional needs. Okay, I'm going to give you one last opportunity to unsettle everybody, and then I'm going to extra disinfect this studio, Matt. What would you say if I told you that I could use your hands to identify you?
I'd say that you had a fingerprint scanner and that no one's interested. You'd be correct, and at the same time also wrong. So we've talked about a bunch of ways that our bodies can be used as unique biometric identifiers. So we spoke about gate analysis, for example, last year, identifying you from the way that you move.
We spoke about facial recognition payments systems a couple of weeks ago. That was episode 158, access denied, where we talked about the many problems that we have with passwords. And we also mentioned emerging systems like ear-shaped detection. But it turns out that the pattern of veins on the back of your hand is also unique.
Look, I'm sure a lot of parts of our bodies are unique. Why would this be any more useful than a face scanner or a fingerprint? Well, we spoke on that password episode about the ability to spoof biometric systems, particularly consumer grade machines. And when you're talking about payment gateways, you want tech that's cheap, reliable and robust because you're going to have millions of millions of units.
Your average corner shop can't afford to rent computing time on Google's DeepMind to process payments just to keep its transactions secure. So we spoke about being able to trick some facial recognition systems with a photo or lift a fingerprint and use it on a fingerprint scanner. The vein reading system uses technologies that are already in many of our phones, such as infrared sensors and LiDAR. So the camera lens is actually able to separate veins and tissue.
And that's what this system looks for. It looks at the pattern of the blood in your veins. And that makes it something that's very hard to spoof or to duplicate. And just how accurate are we talking about?
Well, the model was developed by a team at the University of New South Wales in Australia. For the 35 volunteers they tested, the system was actually 99.8% accurate, which is pretty incredible. It certainly beats sex panther cologne, which 60% of the time works every time. I'll apologize for that.
I watched them come out again last week. In the case of payment technologies, it's much easier to wave your hand at a camera than wait for your hand. And of course, in these touch-sketch times, it avoids having to touch a sensor with your hand, which is, of course, what fingerprint scanners rely on. So one of the applications that the makers think it could be used for is those digital smart locks.
A camera that looks at the back of your hand as you turn the handle, but another application is a little bit more unsettling. It could potentially be linked to CCTV cameras in the street or anywhere really, identifying your swinging hand as you stroll past. Well, you know how to make something feel good about themselves. I'm not sure what to do.
I do my best. Thanks very much for that. Of course, that was Matt Armitage on a Matt's Blade here on BFM 89.9. You can find Matt on Instagram and Twitter at culture.
You can also head over to culturepop.com for transcripts of these shows and information about CulturePop and its consulting services. Thanks again, Matt Armitage. My name is Rich Brebbury for BFM 89.9, the Business Station. Thank you for listening to this podcast.
To find more great interviews, go to bfm.my or find us on iTunes, BFM 89.9, the Business Station.