Dr. Lance B. Eliot, AI Insider
One of the most discussed advancing frontiers is plasticity.
At the forefront of the fields of cognition, biology, social ecology, physics, chemistry, computer science, neural science and studies of the brain (involving neuroplasticity), and many other disciplines, plasticity refers to the adaptability of an organism or equivalent to be able to change and adapt to its environment or habitat.
There have been recently reported cases of phenotypic plasticity in certain kinds of toads, roundworms, lizards, and other creatures that has caused some evolutionary biologists to take a second look at Darwin’s theories of evolution.
Prior to Darwin, there was some naturalists such as Jean-Baptiste Lamarck that postulated it might be possible for evolutionary change to happen in the midst of a single lifetime and not need to work itself out over multiple generations. It was Darwin and others of his ilk that asserted that the “single lifetime” approach was essentially infeasible and unlikely, and that the notion of a multi-generational playout was seemingly more logical and likely.
What has caused a bit of a stir in the standard Darwin theory is that there seem to be some animals that defy the “you cannot change in your lifetime” provision. In a particular species of toads, the spadefoot toad, when they produce their tadpoles, apparently the offspring tend toward eating algae, they are calm and mild mannered tadpoles, and are small-jawed. It is reported that if the water body the tadpoles are in contains let’s say fairy shrimp, some of the tadpoles “transform” into aggressively devouring carnivores and display bulging jaws along with a fierce demeanor.
So, when the environment is the normal and expected calm pool of water and there is nothing carnivorous to eat, the tadpoles are relatively docile algae eaters. If instead the water contains large crustaceans such as shrimps, a change in their normal environment, some of those same tadpoles become intense meat eaters that will take on any comers, which gives them an added advantage in that environment.
Plasticity-First Form of Evolution Comes Into Play
One explanation about the transforming tadpoles and other such creatures has been the suggestion that there might be a plasticity element involved in this. The plasticity theory keeps Darwin’s theory intact. Some are referring to the “discovery” or more like the scientific realization and emergence of plasticity as a sign that maybe there is a plasticity-first form of evolution.
This is one possible explanation for the tadpoles too. Perhaps they have a dominant trait built-in of being polite and vegans, but they also have a hidden trait of being fierce carnivores when needed. Upon experiencing an environment for which the hidden trait has value, some of those tadpoles display the hidden trait. For a human observing the tadpoles, it seems strange and unpredictable that some would “transform” in their given lifetime, when in fact it is simply that they’ve been triggered to use a hidden talent that was there all along.
We can recast the topic plasticity into another realm, namely the nature of the human brain. The human brain appears to be capable of changing and adapting, doing so in neurobiological ways and also in more abstract cognitive ways. There is a continual effort underway of forming and adapting amongst the synapses that connect the neurons in the brain, which we assume is the brain’s way of reorganizing itself and learning and changing.
For those of you versed in Machine Learning (ML) and Deep Learning (DL), you likely know that right now most of the computational models used for crafting Artificial Neural Networks (ANN or sometimes shortened to just NN) are typically rigid and locked-in once they’ve been initially trained.
You toss a million pictures of cats at a deep learning system and once you are satisfied that it seems to pattern-match relatively well in terms of discerning what a cat looks like, doing so by having adjusted automatically or semi-automatically the number of artificial neurons, the layers, and their connections, you then will tend to deploy that deep learning system “as is” and let it do its cat identification magic.
The finalized or deployed version takes as input a new image that might or might not contain a cat in it and ascertains to some probability that there is a cat in the picture or not in the picture and indicates where the cat seems to be.
In today’s deep learning implementations, it is rare that you would have the deployed artificial neural network change and adapt while it is deployed. You more likely might do a retraining if you believe that the deep learning needs further depth or refinement. This would be done in a controlled setting usually, and not in a live environment.
If we are all ultimately aiming to have “true” deep learning and do so by properly modelling and mimicking how the human brain really works, it would seem like we ought to be building into our Machine Learning and our artificial neural networks the plasticity capability that real brains seem to have. In the real-world, the brain is continually changing and adapting, and so should our deep learning models.
What does this have to do with AI self-driving cars?
At the Cybernetic AI Self-Driving Car Institute, we are developing AI software for self-driving cars. One aspect that we are building into our AI systems is a form of DL neuronal plasticity. We believe it is essential as an element for advancing AI and likewise ML and deep learning capabilities of computing.
Less Effort Going into Use of ML and DL for Sensor Fusion
In terms of a driving tasks stack, by-and-large today’s use of ML in self-driving cars is primarily focused at the sensors level of the AI self-driving car automation. There is much less effort underway in terms of using ML and DL for the sensor fusion portion and even less so for the AI action planning and virtual world model updating and analysis.
This initial preoccupation with the sensory data makes sense. The multitude of sensors and their data capture provides an exquisitely rich source of voluminous data and it is relatively easy to come by. Furthermore, vast swaths of data is customarily needed to best make use of today’s ML and DL capabilities, it is their lifeblood, so to speak.
Numerous efforts are taking place at improving the ability to use ML and DL to examine visual images that are captured via the camera and video recording devices on AI self-driving cars. Likewise, via the use of ML and DL, patterns can be found in the radar collected data, the LIDAR collected data, the ultrasonic collected data, and other such data sources. An AI self-driving car needs to figure out what is surrounding the car and then make use of that informed “awareness” to decide what actions the self-driving car should undertake.
A self-driving car that cannot detect its surroundings adequately is going to fail.
It takes though a lot more than just seeing or detecting something to be able to drive a car.
Even if you see the pedestrian crossing the street, you need to put two-plus-two together and realize that there is a chance that the path of the car is going to intersect with the pedestrian, and the car will end-up harming the person. Upon that realization, you then need to try and decide what to do. Should you slow down? Should you swerve away from the pedestrian? Radically hit the brakes? Maybe speed-up?
The AI action planning portion of the driving task is when the driving behavior becomes sacrosanct.
Action Planner Functions Today Are Rudimentary
For modeling of human driving behavior, most of the auto makers and tech firms have to-date been using a rather rudimentary and programmatic approach to having the AI action planner perform its function. They have crudely been programming the more simplistic aspects of human driving decisions into the AI system. If there is a pedestrian in the road up ahead, and if the self-driving car is going to intersect, first calculate if the self-driving car can stop in time. If stopping in time is not feasible then consider a swerving action. And so on.
The AI action planner element:
- Currently tends to be rigid and programmatically depicted, rather than being fluid and based on Machine Learning or Deep Learning aspects derived from human driver behaviors,
- Generally, tends to be based on simplistic hard-coded rules by the AI-developers about how driving is supposed to happen versus based on real-world data of how drivers actually drive
- Will be a key and severe limitation or constraint toward achieving true Level 5 self-driving cars since it will inhibit or undermine the AI to be able to step-up to the myriad of innumerable ill-defined driving situations that will be encountered on public roadways.
Our AI development effort involves using a repository of driving behavior templates, traits as it were, which are based on human driving experiences, and as pattern-matched via the use of Machine Learning and Deep Learning.
In essence, apply the same kind of ML/DL techniques to the detection of objects in the sensory data, but use it for the formulation of driving behaviors based on voluminous driving behavior data rather than sensory images data, and then apply those driving behavior traits to roadway situations as they arise while driving the car.
In addition, this use of ML and DL is not just as a pre-training and pre-deployment kind of operation. Instead, the ML and DL continues while the AI is driving the self-driving car. Learning on the fly is considered an equally valid avenue of learning. Admittedly, in the case of driving a car, some rather significant “guardrails” need to be embodied into the AI system to prevent it from learning “the wrong thing” and taking an untoward driving action accordingly.
Humans of course continue to learn about driving when they are driving a car.
Each time you get behind the wheel, there is an opportunity to learn something new about driving. There is a plasticity in your driving behavior, which makes sense when you contemplate the matter.
When you start to drive as a novice in your teenage years, you have a great deal of plasticity since you are rapidly trying to absorb a swirl of driving tactics and strategies, along with devising tactics and strategies that aren’t otherwise already brought to your attention.
There is “supervised” leaning in which someone explains to you a driving tactic or strategy, such as a driving instructor or perhaps a caring parent that is helping you learn to drive. And there is “unsupervised” learning that involves your own efforts to glean what is happening as you drive, and not only cope with the moment, but also turn the moment into a permanent member of your driving behavior (as a newly formed or revised trait or template) that will become part of your overall mental repository of driving templates or traits.
Let’s consider two use cases. The first will involve a novice teenage driver. The second use case will involve a seasoned driver.
I was helping my teenage children learn to drive, which is both an honor and somewhat scary. You realize rather quickly that there is little you can do from the front passenger seat if your offspring happens to make a wrong move while driving the car.
When I first learned to drive, my high school had specially equipped cars that had dual controls, one for the teenager at the driver’s wheel and another set of controls for the driving instructor sitting in the front passenger seat. Everyone going to the high school was able to take a beginner’s driving course. This made things somewhat easier for parents at the time.
In terms of the driving instructor, I’m not suggesting that the dual controls made life any easier for that teacher, since I can only imagine what his or her life must have been like to work with teenagers all day long in a car that can get into life-or-death predicaments, regardless of the instructor also having access to the driving controls. Forever bless those instructors!
Anyway, after having practiced on local streets with my children driving, it seemed time to try using a freeway. Up until that moment, the fastest we had the car going was maybe 45–50 miles per hour. Now, once we got onto the freeway, it would be more like 60–70 miles per hour. A lot faster than 40–50 mph, even though I realize you might argue it is only “a few mph faster” (it is exponentially higher, on a frightening perceptual scale, I assert). There’s a lot less time to take needed actions. A lot higher chance of things going awry. Fatherly love made me take the chance.
When they reached the on-ramp, they each drove up the ramp and tried to enter into the freeway traffic at the top speed they had already gotten used to, namely the 45–50 miles per hour. I had chosen a time of day when there wasn’t much traffic on the freeway so that we’d be able to drive along steadily and not simply be mired in the usual Southern California bumper-to-bumper snarl. As such, the prevailing traffic was easily doing 65 to perhaps 75 miles per hour (yes, those higher speeds exceed the legal speed limit, but the speed limit is considered more of a suggestion than an imperative here).
I realized immediately that we were going to enter into traffic at a much lower speed than the prevailing traffic. I’m sure you’ve done this before or seen it done by others. The driving problem this creates is that you might end-up merging in front of cars that will have to pump their brakes to keep from ramming into you, or you might cause other cars to have to do a dance trying to get away from the slower going car, all of which could cause a cascade of crashes.
I urged that they push down hard on the accelerator pedal and give us a flash of speed to try and match the prevailing traffic speed. I’m sure that some teenagers would love to do this, willingly and gladly putting the pedal to the floor. My children were more conservative and cautious, thankfully so, and I had to really emphasize the need for speed. Fortunately, we made it okay and nothing untoward happened.
The story might end there, except for the valuable insight it provides about driving behavior and the learning of driving tactics and strategies.
Young Drivers Adapt to Speed-Matching on LA Freeway Ramps
Shortly after that one incident, we ended-up in other situations whereby the need to match the speed of prevailing traffic arose. For example, as they tried to make it to the desired exit ramp, they were in a faster lane and had to slightly decrease their speed to match the cars in the slower lane that led to the exit ramp. I could see them concentrating on what to do and then adjusting their speed accordingly. When we got off the freeway, the off-ramp was a fast turn directly into a busy highway, and they once again had a look of concentration and matched their speed to the prevailing traffic.
They each had adapted to the “new” environmental conditions that involved as a potential “solution” a speed-matching approach (the word “new” in this case refers to their first time driving on a freeway and at predominant high speeds).
Based on the one instance of coming onto the freeway, they had each crafted on-their-own a mental template or trait that imbued them with the driving tactic that when the circumstances warranted it, they considered a “matching the speed” maneuver. Notice that I had not said to them “whenever the situation arises, such as getting onto the freeway or getting off the freeway, adjust your speed to the prevailing traffic.” They devised this notion on their own, merely by my impetus to them to speed-up at the first occasion.
You could say that they learned in a somewhat supervisory fashion, since I did give them a tip or hint and it was presumably my nudge that started them toward the tactic.
It is also interesting that they could have gained a narrower lesson learned in that suppose their thought was that if you need to go faster then go faster. In the aspect of trying to later on get to the exit ramp, they had to actually go slower to match to the slower moving traffic. If the hard-coded rule was go faster, it would not have lent itself to the broader notion of matching the prevailing speed.
These human drivers learned an important driving behavior, which I’m sure became part of their overall driving lexicon.
Did they have to drive a thousand times on thousands of on-ramps to derive the lesson learned? No. I mention this because the prevailing approach to Machine Learning and Deep Learning requires humungous volumes of data. Presumably, the only way a conventional ML or DL could have devised the match-the-speed template or trait would be to have had thousands or maybe hundreds of thousands of traffic flows data to try and pattern onto.
We don’t think that’s needed for doing driving behavior adaptability for an AI system. It helps to have such data, but it isn’t a prerequisite and nor is it the only way to learn.
One thing the kids did have was plasticity. They came onto that on-ramp with a limited set of prior driving experiences. They had to be prepared to change, in the sense of perhaps learning something new or adjusting things that they had earlier learned. They were being confronted with a new environment, a new driving environment from their perspective. It would require honing new driving skills to survive. And, they needed to do so in real-time, in the real-world, in a situation involving real cars and real life-or-death matters at-hand. Adapt or die, I suppose one might say.
The next use case involves a seasoned driver. Me. I’m going to describe it rather briefly here since I’ve already extensively covered the use case in my other writings.
As a seasoned driver, there is not much that I could likely learn anew about driving, though there are always those moments whereby a driving tactic or strategy can be further refined or extended.
You never know when you might get a chance to learn something new for your driving repertoire. Some seasoned drivers that I know have never driven in snow, and thus upon their first encounter with trying to drive a car on snow, they might rediscover the joys of learning something new (to them) about the driving task.
In any case, on my daily commute to work, I drive in the hustle and bustle of Southern California traffic.
Here, especially it seems, everyone wants to get to where they are going in the fastest possible way. For some drivers, they believe that by riding the bumper of the car ahead of them, it is going to magically make things go faster. I’ve debunked this notion overall by examining traffic data and simulations and analyzing it to showcase that this driving tactic not only at times will not work as intended, it can backfire and make traffic go slower, causing at times for the driver to take even longer to get to where they are going. They ironically worsen traffic and make it go slower, in spite of their (false) belief that they are going to speed things up.
Nonetheless, the average pushy driver thinks (rightly or wrongly) that they will get traffic to go faster if they “push” the car ahead of them by coming right up to the back of the car and motivate the driver therein to go faster (or, presumably, get that driver out of the way so that the “faster” driver behind them can get further ahead).
I am accustomed to this driving behavior.
So much so that I anticipate it. I know that a high percentage of drivers here in Los Angeles are going to ride on my tail. No matter what speed I might be going, even if going over the speed limit, these other speed demons are going to go to the bumper. Unfortunately, this kind of driving behavior can have adverse consequences. For example, the driver being tailed now has to be watchful of trying to use their brakes, since the car behind them has little buffer distance to also slow down or stop.
I realize that some drivers figure that if the driver behind them is stupid and doesn’t allocate enough buffer distance, it is the fault of that driver and nothing else is to be done. For me, and for any truly defensive oriented driver, it is crucial to not simply let other “dumber” drivers dictate our options, but it is best to consider how to drive in a manner that takes into account those other drivers and their driving foibles.
After years of my adapting to this driving environment of pushy drivers that constantly are riding on the bumpers of other cars, it had become ingrained in my driving style. My adaptations included numerous driving tactics. For example, you can avoid a pushy driver by potentially spotting them in your rearview mirror long before they get behind your car, in which case, you can then get into a position that will likely preclude them from getting directly behind you, if you plan out the movement of nearby cars and the maneuvering of your car in a chess-like way. And so on.
What makes this driving behavior template or trait of interest herein is that when I recently took a vacation and went to a location that did not have these same kinds of pushy drivers (or, had them but to a much lesser degree), my driving continued as though I was still in the same environment. Each car that I saw coming along, my assumption was that this was most likely a pushy driver, regardless of how they were actually driving, and I silently and subliminally was invoking my pushy-driver control tactics.
This aspect that I fell into is a mental trap known as prevalence-induced behavior.
Conclusion — Aim for Artificial Neuroplasticity
I’ll tie together the giraffes and the tadpoles with the aspects of driving and driving behaviors. They all interrelate by the matter of considering what kinds of traits we have, some of which are innate, some of which are learned, along with the plasticity of being able to change and adapt to our environment. If Darwin were still here, I’m sure he’d be interested in this topic too.
To further advance AI, I’d wager that we’ll need to make progress on Machine Learning and Deep Learning that will incorporate plasticity. We need to be able to construct artificial neural networks that can change and adapt and adjust as the environment changes, in real-time, in a real-world context, and essentially on-their-own as we’ve hopefully imbued them with the capabilities to do so.
In that sense, we should all be aiming to have artificial neuroplasticity, which, since real neuroplasticity occurs in the brain, we likely will need to do something likewise in the computer if we are going to reach AI brain-like capabilities.
For driving purposes, the AI action planning is where the crux of driving and driving behaviors resides. Being able to see and sense the driving environment provides the so-called table stakes for playing the self-driving AI game, but to really succeed in AI self-driving cars will require the AI to be able to drive with driving behaviors, ones that are honed and pre-tuned, and others that will arise as the driving situation emerges and the driving environment changes (as perceived by the AI).
If those tadpoles have the ability to change how they act and look, doing so after sensing the environmental conditions that warrant a change, and presumably bringing forth some kind of latent traits that can be triggered and showcase the plasticity of these toads, I’m voting that we can do the same kind of thing with driver behavior templates and traits, for which the AI self-driving car would use and refine, based on the driving environment and the plasticity that we’ve built into the AI. Score one for the humans and let’s show those malleable tadpoles what we can really achieve.
For free podcast of this story, visit: http://ai-selfdriving-cars.libsyn.com/website
The podcasts are also available on Spotify, iTunes, iHeartRadio, etc.
More info about AI self-driving cars, see: www.ai-selfdriving-cars.guru
To follow Lance Eliot on Twitter: @LanceEliot
Copyright 2018 Dr. Lance Eliot