By Dr. Lance B. Eliot, the AI Trends Insider
I live in a gated community that has a main gate for residents and guests, and provides an impressive driving entrance into the neighborhood that displays ornate iron gates and a spectacular water-sprouting fountain. There is a secondary gate at the back of the community. This secondary gate is intended for various contractors and vendors that need to gain access into the neighborhood and it also provides a fast way to get into and out of the neighborhood for the residents, though the gate is very simple in appearance and can be considered of utility value only. For many years, I’ve routinely used both the main gate and the secondary gate.
One day, I drove into the community via the main gate and wondered where some of the roads in the community led to. I usually just drive straight to my home and don’t diverge from an otherwise highly efficient and optimized driving route from the gates to my home. Why bother, I had thought, in terms of exploring the rest of the neighborhood. I knew the aspects that I needed to know, namely, where my home is, and how to get into and out of the community. Nothing more to be concerned with, I figured.
As I drove down a winding road in the neighborhood, it gradually looked like it would reach presumably a dead-end at one of the edges of the surrounding protective fence. Instead, I discovered that there was a gate at the end of the road. The gate was nearly out of view and only those that likely lived on that particular street even knew it was there. A secret passage! I inched my car up to the gate and it opened, leading out to a public street that is seldom used. I felt like I was a secret agent that now knew of way to go into and out of the gated community virtually unseen. How exciting!
This story illustrates some interesting aspects. First, let’s consider how I found this hidden treasure. Question for you — had I been focused on a goal to find another way into or out of the gated community? No. In fact, I was completely unaware that such a third gate even existed, which is kind of surprising since I had lived in this neighborhood for several years. You would have thought that I would have at some point either encountered the third gate, or maybe the guards at the other two gates would have mentioned it, or my neighbors would have said something about it, etc.
What was it then that led to the discovery? Answer: my curiosity. I was just generally curious about the community and happened to have some spare time to indulge my curiosity. I really had no idea what I might find by driving around somewhat aimlessly. If you had asked me beforehand, I would have bet that I wouldn’t find anything particularly interesting. I assumed it would be just home upon home, all pretty much looking the same. Nowhere in my mind was the idea that I would find another gate.
It was my curiosity that caused me to look around. It was curiosity that led to this surprising and quite helpful discovery. Without that curiosity, I would have remained blissfully unaware of the discovery and not have been the wiser that it existed. Though I admit that this discovery was not as though I had found a new subatomic particle or that there are aliens from Mars here on our planet, it was nonetheless a productive result of my curiosity.
You might be aware of the famous quote by Albert Einstein about curiosity: “I have no special talent. I am only passionately curious.” I think we might all agree that his modesty about not having any kind of special talent is a bit over-the-top in terms of humility, but anyway, we would likely all agree that he definitely had a large heaping of curiosity. Talent that has no curiosity is bound to be kept in a cage and not going to reach its true potential. Curiosity alone, without any kind of talent, probably would be somewhat wasted or might lead to troubles.
We all know that curiosity killed the cat, or so they say. That’s why cats seem to have nine lives. Being curious is both an advantage and a disadvantage. A person that uses curiosity can make new discoveries, similar to what happened with my finding the third gate in my community. Of course, there was perhaps a chance too that my curiosity might have led to some calamity. Suppose I had driven down a street that had a group of thugs waiting for someone to unknowingly drive along, and they were waiting to stop the car and beat the driver to a pulp. In that case, I would have been cursing my curiosity.
In the case of my driving around the neighborhood, I had somewhat calculated the risks associated with indulging my curiosity. Where I live, the odds of a pack of thugs is close to zero, so that really wasn’t an event that would occur (well, I did have a bunch of rambunctious children one day toss water balloons at my car, but I dare say those aren’t thugs by the strict definition of the word). I did waste some of my personal time by driving around, and presumably could have used that time for some other purpose. I did use up some gasoline as I drove around, plus I was putting more wear and tear on my car. All of those costs, or one might say are disadvantages associated with my application of curiosity in this instance, but they were relatively mild, and so the opportunity (or advantages) to see the neighborhood and take a chance on finding whatever I might find was a reasonable one.
I have explained here that curiosity can be a carefully utilized tool that is chosen from a cognitive toolkit by using reasoning about the potential Return On Investment (ROI) of deploying the curiosity. That’s not always the case that someone or something opts to presumably do an elaborate ROI calculation before invoking curiosity. Does a cat that leaps from the top shelve onto a nearby table in order to explore the table make use of a mental calculus about the motivations of curiosity? We don’t know, but maybe it is fair to suggest that if there is some kind of calculus, it probably involves limited formulas. It might be a simplistic calculation of seeing something that catches the eye of the cat and so it then opts to go for it. If you watch cats carefully, you’ll notice that they do seem to try and be a bit more elaborate in gauging the use of their curiosity, such as looking around first to see if there might be an obvious danger sitting on the table or that they don’t have solid enough footing on the shelve to make the jump.
For humans, we generally believe that curiosity is crucial for human development. Research in psychology and also in neuroscience is replete with studies that purport to show the importance of curiosity in how we learn. Curiosity is considered a cornerstone for doing explorations of our world around us, and for the desire to formulate hypotheses and test them to see where they lead. Some argue that curiosity is found in all animals, and that in humans we take it to another higher level of usage by being able to increase our knowledge and insights. At times, humans exhibit perhaps the same basic level of curiosity and get the same benefits (and costs) as animals, such as the human that leaps from the top of his house over to a nearby tree to see if it can be done (saw this on YouTube, somewhat like my earlier indication of a leaping cat), but at the same time we have instances such as Albert Einstein that leveraged curiosity to find new ways to describe our understanding of energy and matter. It seems doubtful that many cats or even dogs can use curiosity in quite that same way as Einstein did.
You can think of curiosity as typically depicted as being either so-called “state curiosity” or being “trait curiosity.” State curiosity is the type of curiosity that seeks to employ curiosity simply for the sake of using curiosity. If I have curiosity, I might use it because I can. Perhaps I have no particular goal in mind for using the curiosity. I just let it run its course. In the case of my story about the gated community, you could say that perhaps I was making use of state curiosity and driving around just out of overall curiosity.
In the case of trait curiosity, there is supposedly some form of learning goal that underlies the use of the curiosity. Let’s say that in the case of the gated community, I had wanted to know where each of the roads led, as a precaution if I should ever get lost in the neighborhood and wanted to have an already formulated map of the streets. In that scenario, my curiosity had a particular focus of learning. I knew that I wanted to create a map of the community and my curiosity spurred me to do so.
It can be a fine line between the state curiosity and trait curiosity. At one moment, my curiosity is being spurred by this desire to make a map, and the next moment I come upon the third gate. If I was strictly focused on only making the map, I probably would have not tried to use the gate and instead merely continued driving around. In short, there are times at which curiosity can be a mixture of state and trait, and they can intertwine around each other in a fluid fashion.
In the field of AI, we are quite interested in curiosity. Why so? Well, if you believe that human intelligence depends to some extent on the cognitive aspects and capabilities of curiosity, and if you are trying to create artificial intelligence that is based on what human intelligence consists of, you naturally would want to know how curiosity works. By figuring out how curiosity works in human intelligence, you could try to mimic it in a machine and then maybe that helps to produce a true AI system. Can humans exist without curiosity and be considered intelligent? If so, then presumably AI wouldn’t need any semblance of curiosity in order to be as intelligent as humans. On the other hand, it does seem like curiosity is pretty commonly found in humans, whether innately or learned, we aren’t quite sure, but anyway it is present and so we in the AI field believe it worthy to pursue it.
Curiosity at times could be viewed as a means of reducing uncertainty. There is uncertainty in your environment and thus you use curiosity to reduce that uncertainty. My uncertainty was about the roads and nuances of my gated community. I reduced my uncertainty by driving around. Or, you could say that I increased my certainty about my environment. I became more certain about what was there. This could be a means to boost my chances of survival, allowing me to be better able to survive in my environment. If you are familiar with Maslow’s hierarchy, you could say that curiosity can have a payoff at any level of the hierarchy, and not just be solely a benefit for the survival level of the human condition.
You could also claim that curiosity allows humans to seek new opportunities. One theory called the optimal-arousal theory posits that we use curiosity for finding something that can gain us an advantage of some kind. My curiosity about my gated community could have been due to my desire to find a better house and maybe during my driving around I’d find a house for sale that I otherwise did not know about. Rather than looking at curiosity as reducing uncertainty, this theory is a camp that asserts we use curiosity to gain new ground.
The neuroscientists and biologists and chemists might say that curiosity is driven by our neurotransmitters. There seems to be evidence to suggest that indulging your curiosity can release dopamine and opioids. You could say that we have curiosity and use our curiosity as a pleasure seeking animalistic mechanism. Just like your enjoying ice cream on a hot day, your curiosity is a satisfier that causes your human body and your human mind to have a good time.
In the above sense of reducing uncertainty, of seeking opportunities, of wanting to be pleased, AI tends to so far be modeling curiosity as a rewards and penalties type of mathematical function. For example, researchers Pulkit Agrawal, Deepak Pathak, Alexei Efros, and Trevor Darrell at the University of California Berkeley AI Research Lab have explored how curiosity works by establishing an elaborate points system for rewarding successful behaviors and for penalizing unwanted behaviors, doing so in such tasks as an AI system learning to play the video game of Super Mario Brothers. Some characterize this as a carrot-and-stick approach to doing machine learning. Underlying this application of curiosity is the notion of bounding the curiosity by both intrinsic motivation and by extrinsic motivation.
What does all of this discussion here about curiosity have to do with self-driving cars? At the Cybernetic Self-Driving Car Institute, we are developing software components that go into a self-driving car application portfolio and provide an added cognitive capability shaped around curiosity. In essence, we want to provide self-driving cars with a cognitive tool consisting of curiosity.
You might be surprised to hear this. What, you say, a self-driving car being curious? Crazy idea, you might at first suggest. Well, remember earlier when I mentioned the idea that curiosity seems to be an essential element of human intelligence, and that AI if it is to be considered equivalent to human intelligence we might therefore need to embody curiosity into AI – let’s tie that aspect to the fact that a Level 5 self-driving car is supposed to be able to drive a car in whatever manner that a human could drive a car. For the AI to do so, it can be argued that the AI needs to thusly have curiosity.
When I have given presentations about our self-driving car “curiosity invoking” software component, I’ve had some that tried to instantly discount the idea, doing so by saying that if a self-driving car got curious, maybe it would wonder what it is like to drive off the end of a pier, and the next thing you know, a self-driving car full of human occupants decides to drive off the end of the nearest pier. I get that point, and I would also say that’s a pretty stupid implementation of curiosity. In that sense, I agree that we don’t want unbounded curiosity in our AI.
This is true pretty much of any AI that uses curiosity. You likely don’t want the curiosity to run unbounded and unfettered. You could certainly also make the same case about humans and animals. If a human or an animal allowed curiosity to get the better of itself, you’d indeed have a many a case of the infamous curiosity that killed the cat. Indeed, there are daily stories in the news about people that did something untoward, presumably allowing their curiosity to go wild, and get themselves into really bad situations. The other day, a man was curious about where his toilet pipe led to, and so he reached in with his arm and got stuck. They had to break apart the toilet to get his arm out. Not the best use of curiosity, I’d say.
Our curiosity software for self-driving cars is bounded by parameters that avoid aspects such as getting curious about driving off a pier, or any similar kind of dangerous or ill-advised action. There are “curiosity traps” that we know to be watchful for about making use of curiosity. For example, curiosity can be excessive, and as such the cost can exceed the benefit of invoking it. If I had driven around in my gated community endlessly, it would have been an of excess curiosity and gotten the better of me. Curiosity can also lead to dead-ends, and as such, this also needs to be detected.
The way we’ve been developing the cognitive curiosity capability involves self-curiosity and also human-directed curiosity.
In the case of self-curiosity, the AI seeks to employ curiosity on its own, without a human urging it to do so. We continually have the curiosity component running in background, and it then shares what it finds with the other components, such as the self-driving car strategic AI component and the tactical components, which then can decide whether to make use of what the curiosity component has found.
For example, the curiosity component is looking at the path that the self-driving car is taking, and seeking to identify alternative paths. This is more than the already traditional approach of finding paths that are better optimized paths. Our curiosity component is looking at other paths that aren’t necessarily more optimized for route planning purposes. Instead, these other paths might offer some other valued aspect, such as a more scenic route, or a route that could be later used when traffic is snarled.
It also is observing other cars in the surrounding environment, and trying to figure out what those cars are up to. As I’ve mentioned in my other writings, self-driving cars can be enhanced by how they observe and potentially learn from other cars, whether other human driven cars or self-driving cars. The curiosity component is peeking at the sensor data and sensor fusion of the self-driving car, and trying to spot seemingly curious behaviors of other cars, which then it can bring to the attention of the other elements of the self-driving car that are strategically and tactically driving the self-driving car.
In addition to self-curiosity, there is also human-directed curiosity. The human in the self-driving car is able to interact with the component dealing with curiosity, and inspire or help shape the focus of the curiosity component. If I were in the self-driving car as an occupant, I might encourage the self-driving car to take its time getting to my destination and show me around to some novel places. The curiosity tool then overtly gets invoked and becomes a key component for what the self-driving car creates as the traversal plan.
There are some that believe our self-driving cars will be working all the time, meaning that even when you as the owner aren’t in the self-driving car, it nonetheless might be driving around. One reason for the self-driving car to be driving around involves the notion that we might all become Uber-like and have our self-driving cars be doing ride-sharing when we are not using the self-driving car. Another less obvious aspect involves having your self-driving car learn more about its environment, doing so without you needing to be in the self-driving car. Imagine, for example, if I had a self-driving car that I had parked at my home in my gated community. It might have on its own opted to explore my gated community and found the third gate (self-curiosity), or I might have tasked the self-driving car to go ahead and drive around and let me know what it finds (spurring the curiosity component to the forefront via human-directed curiosity).
Curiosity can be focused or unfocused. It can be working in the background in case it finds something worthy of consideration, or it can be at the forefront. It needs to be roped in and not allowed to overtake the rest of the system. The ROI for making use of curiosity has to be kept in mind and be an overt aspect of weighing the value of deploying curiosity or not, and the context and situation will all be factors that can shape the circumstance of when curiosity makes sense to use or not.
We believe that curiosity is and will be an essential tool in self-driving cars. Whether you agree or not to that notion, I think that at least you might agree that curiosity as an essential element in AI has merits. Let’s all use our curiosity to at least see how we can further extend AI to embody curiosity. I hope I’ve piqued your curiosity to do so!
This content is originally posted in AI Trends.