The Real Risks of Smarter Machines
When people ask me what I’m working on, I’m often confused about the depth I need to go to in my response. ‘Artificial Intelligence’ is way too broad for my personal satisfaction, and image understanding probably too specific. Nevertheless, every single time, I do get this completely unrelated follow-up question that infuriates me to my core.
‘Is AI going to take over the world?’
As someone who’s dealing with machine intelligence at an investigatory level, this question really bums me out. And I can’t even blame the skeptic — most people think artificial intelligence is some unknown, mysterious entity which is conspiring infinitesimally, and will eventually kill us all, since it can predict that Sausage Party is the next movie we’d want to watch after we’ve binge-watched Evan Goldberg flicks all night.
What most people fail to realise is that however idiosyncratic we think we are, our preferences and basic intelligence generally follow ubiquitous patterns, that computers can easily identify once they’ve seen them enough times.
That’s what makes predicting your favourite music, or suggesting the correct phone app to use while you’re taking a dump — an easy task for machines. That said, I don’t imply that all prediction tasks are alike in nature and difficulty — but expecting people outside the realm of investigation to comprehend that is a stretch.
The key to understanding what our current techniques in artificial intelligence are strong at is knowing that computers learn especially well in two primary settings — i) controlled environments, and ii) supervision.
We saw that Google’s artificial Go playing machine decimated the best human player recently. Chess was solved long before most of us hit puberty, and there are so many recent papers about beating humans at Doom. The thing is, in games, you know the possible outcomes, the operating environment and the actions you can take, which make the task of modelling the problem much easier. Once we’re able to model the world our game is in, simulating and learning from the simulations is the next part — and we’ve made leaps and bounds in the hardware required to do large scale learning, which made breakthroughs like the technology behind AlphaGo possible only recently (the theory was pretty much present this whole time).
Games are a perfect example of a supervised controlled environment. An environment where you can estimate a penalty you get with each action you take, and can hence learn efficiently from your mistakes. Another example of a supervised controlled environment would be the movie prediction that we were talking about. There’s a large catalog of both users and movies, and given a pattern of user choices, we need to predict the next one.
In supervised controlled environments, we know what kind of information we’re getting, and can handle similar bits of information in the same manner. We create representations for these groups of objects, which will help us eventually identify the exact math to carry out when we have to make a prediction. This is a very narrow subset of general learning — the kind intelligent beings like humans do.
Humans interact with our environment and develop logic as well as intuition, and largely without supervision. Basic processes such as identifying objects and understanding physics occur every waking second, and more often than not, we learn more about new information by interacting with it. That’s where computers fail miserably right now.
Today, if you want a machine to recognise your car in all your pictures, you have to tell it what it has to look at (your pictures) and what your car looks like. And once you’ve given it many images of your car, then it can certainly identify it. This is supervision in learning, the supervision being you pointing out what your car looks like in some images before it understands what to look for. What computer scientists are trying to figure out right now, is learning without, or with little supervision — unsupervised learning. We want to, ultimately, make machines that can understand the notion of (in this case) object-ness and scenery on its own, without us explicitly telling it what it is.
Learning without, or with very little supervision is a harder problem to solve, and most of the research in artificial intelligence is focused on tackling this. Our machines are getting smarter, yes, but mostly in supervised, controlled environments. We first need to figure out how to make robots learn without supervision, and then in uncontrolled environments to make systems which are even close to general human intellect.
So yes, even the question about robots killing all humans, or even talking about the ‘intent’ of a robot is still far away in the future. However, a much more serious threat from AI is looming, and can have severe repercussions.
In the very first talk at this conference I was attending, one of my previous advisors mentioned something that for the first time, made me question the applicability of AI research.
In traditional, earlier artificial intelligence techniques, we could very easily understand why the algorithm was doing what it was doing. For example, we were to build a machine that would, just by measuring people’s height and weight, tell them if they were overweight or not. For this, we would just have to calculate the person’s BMI, and if it’s above a certain threshold, call them overweight.
This is a very primitive, artificially intelligent agent. If it were to call someone fat, unlike a hurtful teenage bully, it would actually have a logical justification behind its action — the person’s BMI is in the average ranges for overweight people.
Most machines employed these days obviously aren’t this simple. They take in complicated, large inputs (high-definition images, for instance) and have to do fine-grained predictions about their contents. Here, simple methods like a threshold or a decision tree don’t work. Increasingly, systems use a set of algorithms known collectively as deep learning, which are computationally intensive techniques that can identify and learn, with large amounts of data, detailed patterns in a manner similar to humans.
These systems are remarkably good at their job, despite being slow at learning since they need a lot of data to learn.
However, there’s a catch — once they do give us an outcome, albeit the correct one, most of us don’t know how exactly the machine came to it.
This isn’t as alarming as it sounds —well, initially. In machine learning systems, we have two types of data — features and labels. Features are the observed variables of a system, and labels are what we’re supposed to predict. For example, in the earlier obesity detector, our features were the person’s height and weight, and labels were one of overweight or healthy for each person. For detecting cancer cells from images, our features would be the images of several organs, and our labels would be whether the said image has a cancerous cell or not.
Machine learning algorithms would normally solve this problem by assigning ‘weights’ to each feature, summing them up, and finally making a decision based on the value of the sum. For example, if you were asked to predict if an apple had gone bad, you would probably look at the apple’s smell, its color and finally, if you do touch it, then the feel of the skin, mentally assigning weights to each of these features.
Computers follow a similar high-level idea, except that these ‘weights’ are mathematically obtained by different optimisation techniques. In deep learning, however, we aren’t sure of what features we want to look at, let alone assign weights to these features. So what do we do? We let the computer learn the best features on its own, and combine them in the best way to make predictions, in a manner that is in some sense similar to the way the human brain does.
This idea gives us astonishing results — in computer vision (that is, the study of making computers understand visual data), especially, the advent of high-performing GPUs and novel architectures made learning basic image-level concepts a piece of cake. However, the caveat — these features that we were talking about, which these algorithms learn, make less intuitive sense than what they used to with traditional techniques.
Most people won’t see this as a problem — it technically isn’t that big a problem at the moment, where the tasks that AI is being used to solve are concrete, such as identifying people and objects in images, tracking faces and generating speech samples. Here, we have a fair idea of what kind of representations the algorithm is learning (in fact, this demo is a recent development in this regard). However, when we start using deep learning for tasks which have more risk associated with predictions, each prediction needs to be justifiable.
Consider that you’re a bank, and you have detailed transaction and credit histories of all your customers. You use a sophisticated deep learning algorithm to identify loan defaulters. Given that you’ve got a huge database of user behaviour patterns, your algorithm will probably fetch you a really good accuracy on this task, except that once you do suspect a future defaulter, you don’t exactly know what exactly caused suspicion, making justification of the prediction more difficult.
Most deep learning systems don’t have good techniques for understanding their decision making capabilities, and this is currently an area of active research. For some task-specific deep networks, especially in computer vision, we have made progress in understanding these systems — localising, to a good degree, what provokes an algorithm to do what it does. But for a general scenario, there’s still work that needs to be done.
Machine learning systems suffer from one severe handicap — the manual input required to make them correctly separate the signal from the noise. Or, in technical terms, overfitting. What I mean by this technical mumbo-jumbo is that when a model has to fit a certain set of gathered data, to make a prediction on the newer, unknown data, it can learn to become too comfortable with the gathered data it already has. Consequently, it won’t do well when deployed in the real world.
This means that models will generally start looking for patterns that don’t really exist in the real world, but exist in the collected data the algorithm has to train on. There are several ways to understand overfitting, and there are some really good real-life examples of overfitting systems for the curious, but a simple case would be filling your suitcase up entirely with summer clothes when it’s summer where you are, but it’s 11 degrees in Amsterdam, and you find yourself shivering like a leaf.
The reason for going into this aside about overfitting was to highlight the importance of accountability in machine learning systems. If we don’t have ways to understand and see what these algorithms are learning, we won’t have good ways to tell if they are overfitting. An example of the ill-effects of something like this is a machine which is predicting suspicious behaviour online from browsing history, and since most users it’s seen are 19-year olds from, say, the United States, it’s biased against any 19-year old user from the United States, even though their search history mostly has PewDiePie videos.
The repercussions of this will exponentiate with the increase in applicability of deep learning systems in difficult inference tasks. For example, we see a lot of research going into medical image prediction — an application where decisions require greater accountability and understanding. Additionally, in cases where the scale of predictions is large enough to prohibit manual examination of the predictions, we will need systems which allow us to understand and adjust what these algorithms are taking into consideration.
This threat is looming, but is an area of research, and, as more time is invested in it, better solutions are likely to emerge. However, we must realise the importance of model accountability, especially as we develop new systems for making our lives easier. I’d like to end with an example —
If a human crashes a car, we can hold them accountable, and understand why the accident happened — they were probably inebriated or texting someone while holding soup.
But if a self-driving car crashed into another one, killing a passenger, who would you blame, and why? How would you make sure it doesn’t happen again?
These incidents have occurred several times recently, and with more systems using AI, more mistakes will happen. To rectify, we need to understand what we’re doing wrong — that’s one of the major challenges AI faces today.
