Six easy and not-so-easy AI pieces

There is a hard divide in AI. One on side are the problems that we know how to solve given time, money, and a little bit of luck. On the other side are problems that we can barely define. This essay explores six problems from both sides of this divide. The title itself is a riff on Feynman’s essays on physics.

Photo by Franck V. on Unsplash

Artificial Intelligence (or even intelligence for that matter) is poorly-defined. That’s not to say that people haven’t tried, but such definitions are either vague or overly specific. The vague definitions define intelligence as some super-set of intelligent abilities (for example, everything a six-year old can do) that, while sufficient, don’t tell us more about what intelligence is (Is it the six year old’s ability to speak that is important? or their ability to learn? or their mobility? or their curiosity? etc). The overly specific definitions, on the other hand, are so one-dimensional that the moment you solve the problem, it becomes not-AI (chess is a great example of this AI to not-AI transition; autonomous driving will soon become another). So, by definition, there are no easy problems in AI. That leaves us the hard problems, and the very hard problems. Or, to borrow Feynman’s nomenclature, easy pieces and not-so-easy pieces. In this article, I take easy problems to mean problems that we can solve in the near future (probably the next 20–30 years). Solutions to the not-so-easy pieces are at least an order of magnitude farther out (100–200 years) with most estimates being little better than guesses or a gut feeling.

So, by definition, there are no easy problems in AI. That leaves us the hard problems, and the very hard problems.

Six easy AI pieces

There are many many easy problems in AI of course but I selected these six because these are normally considered hard problems. In the bigger AI picture, however, these are easy. Solving these problems do not necessarily bring us closer to artificial intelligence.

1. Autonomous Driving — Autonomous Driving (AD) is hard. Very hard. But, it’s getting clearer everyday that we can and will solve this problem in the next 10 years, 20 tops. That’s not just because we have prototypes driving around on the streets. It’s also because it’s one of those AI problems where the value-prop is clear enough that there is money to be made in developing even limited solutions. Of course, like for the remaining easy pieces, the solutions may not always look like we imagine currently. AD will most likely involve changes to our driving infrastructure. Large scale instrumentation of roadways, better inter-vehicle connectivity via new technologies like 5G, and changes to public policy will all come together with core technological advancements to make autonomous vehicles a part of our daily life.
2. Smart Homes/Factories — Smart homes, smart cities, Industry 4.0. These are all realizable today. What’s stopping us is not the technology but a lack of compelling use cases because the cost of integrating these technologies into existing infrastructure, and because the economics of automated factories vs outsourcing (specifically for industry 4.0) don’t measure up very well. It is going to take a lot of money to retrofit existing factories to do things in a “smart” way, but is it worth it if you can still manufacture stuff cheaply elsewhere in the world? And, to the average homeowner, what really is the benefit of a “smart” home?
3. Explainable AI — Machine learning, and specifically deep learning, has made great strides in the past decade in very specific problems and domains. But the inner workings of these techniques are opaque and their success often feels like magic. This black-box like effect is problematic in certain domains like healthcare or any part of the government like the justice system for instance, where it is important to know why an AI model made the prediction it did and to be able to assure people that there is no inherent bias in how laws are applied. Lady justice is supposed to be blind but can we be sure that these techniques are too? Explainability is a difficult problem for which we currently don’t have any good solutions. But, there is also no real demand for explainable AI. Sure, everyone says they want it, but there is still so much low hanging fruit in terms of applying AI that explainable AI is mostly restricted to academic research (a darker take is that companies don’t really want explainable AI because it allows them to hide behind the opaqueness of their algorithms). The good news is that we have made great strides in some areas of explainable AI. Not all methods are black-box-like and there are ongoing efforts to understand newer techniques like Deep Learning. Plus, we are finding (in hindsight of course) that in many cases the performance of these newer techniques can be achieved using older, more explainable, methods.
4. Structured AutoML (“easy” to get 90% of the way) — There are at least three ways in which people use the term AutoML. There is AutoML as hyperparameter search. There is AutoML as automating the process of data science (including pipelining and feature engineering). And, finally, there is AutoML as automating the process of machine learning itself which includes learning new algorithms. The first two I call structured AutoML since it operates within a certain structure that has been pre-defined by someone (AI/ML researchers, data scientists etc). The third one is much harder and closer to the not-so-easy piece of AI. Most current AutoML work focus on solving problem 1 although nearly all of them have ambitions to solve problem 2 as well. In general, 1 and 2 can be thought of as automating some of the design choices that data scientists make when building models. This is clearly a non-trivial problem and even pure hyper-parameter optimization is sometimes more art than science. But, with newer techniques and more computing power, I don’t see a reason why we can’t do most of 1 and 2 in the next ten or twenty years.
5. Natural Language Processing (including speech processing) — We’ve begun to take for granted that talking to Alexa or Google Home is natural and, while there are many places to improve, for a limited vocabulary and specific domains, these systems are exponentially better than anything we had five years ago. Google’s ad of their assistant making salon appointments, while only a demo in very controlled experimental conditions, is a harbinger of things to come. Maybe not next year but in five years it will be natural to have your AI assistant call and make appointments for you. That said, one area where AI still has a long way to go is machine translation. Most of the simple stuff, like translating news, everyday conversations, technical reports, articles, etc, we can do automatically in the near future. However, translating a novel or a poem in a way that captures the essence (and not just a literal translation) is not easy to do. But, to be fair, even the multi-lingual among us find such translation hard going.
6. AI creating AI (in an evolutionary way) — One of the two holy grails of AI (the other being consciousness) and, to some, the beginning of the end of the human race. Once AIs create AIs, there is nothing stopping the creation of Skynet and it won’t be long before all decisions are taken out of the hands of humans and made completely by machines. We have a long way to go before any of that happens, so why is this one of the easy pieces? Any complex problem can be broken down into simpler more manageable pieces. Like in NLP, we will begin to see AI systems that are built solely for the purpose of building other AI systems. The difference is that these systems will be extremely focused on optimizing a reward function, one that is carefully crafted by an ML engineer/scientist for a particular set of problems. Sure, someone may decide to build a system that crafts reward functions but it will have a reward function that was carefully crafted by an ML engineer / scientist. It’s turtles all the way down but we’ll discover a few of those turtles in the next few decades making this problem solvable in a narrow sense.

Large scale instrumentation of roadways, better inter-vehicle connectivity via new technologies like 5G, and changes to public policy will all come together with core technological advancements to make autonomous vehicles a part of our daily life.

Six not-so-easy AI pieces

1. Reasoning — 99.9% of AI systems in production are inference or pattern matching systems (I’m hedging my bets that the other 0.1% consists of some causal reasoning systems deployed somewhere). Pattern matching is only the first step in a process that moves us to actual AI. Reasoning systems is the next big challenge. How can we build systems that reason, i.e., break down complicated problems into sub-problems, iteratively solve these sub-problems, and then put it all back together at the end? While there is a long history of building such systems (starting with Newell & Simon’s General Problem Solver), these solutions work well for limited cases and only when the information needed is well structured, and organized. There is also still no solution to the Symbol Grounding Problem — the gap between how we perceive the world and the atomic elements of reasoning. We know that these elements (and reasoning systems) must be built on top of pattern matching systems. Pattern matching systems that are similar but probably not exactly the same as the systems we currently use and employ but the “how” of that is still unclear. Only once we have such reasoning systems, can we move on to systems that decide and set their own goals instead of solving problems given to them. We are making some steps towards solving these issues. Current research into attention mechanisms are an important step in the process because it allows systems to focus on parts of problems instead of blindly taking data and making a prediction. Research into causal reasoning systems is another area with potential but it will take us closer to 200 years than 20 to build such systems.
2. Understanding — We don’t even know how to define understanding (much like intelligence). We can provide approximate definitions but there are always exceptions. We can say (like Justice Stewart’s definition of obscenity) that we’ll recognize it when we see it, but we could just as well be anthropomorphizing the result of what turns out to be less understanding and more reflex action. Is knowing our preferences understanding? If so, AI systems already understand, because they reroute us to avoid traffic jams, suggest new shows to watch, and recommend good deals to us all day. On the other hand, does anyone really understand anything? Between these two extremes lies an extensive history of thinking machines and a wide variety of definitions for understanding that we don’t even come close to achieving currently or in the near future
3. Graceful degradation — Most computer systems are extremely brittle. Except in safety critical areas such as nuclear power plants, space shuttles, and, hopefully in the future, autonomous cars, there is very little verification and validation of such systems. But, true intelligent systems need to be flexible and resilient. One of the finest abilities of living organisms is the flexibility to adapt to their surroundings. This flexibility takes a lot of forms but one critical aspect is graceful degradation — the ability of an organism to not stop working when things go wrong but to be flexible and take into account the changed circumstances. We do this everyday. We use one hand to open a car door and the other to swing our backpack into the back seat. If we are carrying a box with both our hands, we wedge the box against the side of the car, open the door a little and use our elbows to open it fully. Or we may place the box on the ground before opening the door. Or we may ask someone to help us. There are many ways to solve this problem and we do it unhesitatingly instead of giving up immediately. To be fair artificial systems do this too. An autonomous car will pull over to the side of the road when it can no longer drive autonomously. But, this is an example of hard-coded behavior built into that system and highlights an essential difference between life and artificial life. We are nowhere close to replicating graceful degradation partly because there are humans waiting to take over when the system is failing, and partly because these systems are rarely used in critical scenarios where such failure has harsh consequences, but mostly because solving it requires solving the more general problem of reasoning. Fields like life-long learning and explanatory AI are just the slightest hint of a first step in this direction but we have a long long way to go before we can show the kind of adaptability of even the simplest animals.
4. Right vs. Wrong — Even in the legal sense, this is a non-trivial problem. We are at the very early stage of AI being applied in very specific situations like deciding sentencing guidelines and bail amounts. But, when you move into right vs. wrong in the moral sense, we are faced with the problem that everyone has their own moral code that is the result of one’s unique experiences. Can an AI “grown in the lab” have such a moral code? Does an AI even need a moral code if it adheres strictly to a legal code? We don’t even know the right questions to ask, let alone the right answers.
5. Consciousness — The second holy grail of AI, consciousness has long fascinated a variety of scientists — from philosophers and biologists to computer and neuroscientists. Like understanding, finding an acceptable computable definition is hard. But what is harder is deciding whether it is even useful. Is consciousness what makes us human? Or is it being human that makes us conscious? Can we make a fully functioning AI system that is not conscious? There are even levels of consciousness and long-standing arguments about whether the harder levels are within our abilities to create.
6. A four year old — This one is a little bit of a cheat because it combined pieces 1 through 5, and then adds quite a few more into it. A one year old would suffice, or even a newborn. But there is something that happens around 3 or 4 that has profound implications for how we see and perceive ourselves. We develop a theory of mind — the ability to put ourselves in other people’s shoes and then reason about the world as if we were viewing it from their point of view. It is neither necessary nor sufficient for intelligence but there is something critical about this ability that allows us to become social animals and better problem solvers. In addition to this theory of mind, there are a host of other abilities that four and five year olds manifest that we are not even close to understanding much less replicating as a system. These include the mundane abilities of interacting with the environment (not just bumbling about in it), curiosity (recent papers on curiosity are just modern versions of the explore vs exploit paradigm), making social attachments, self-preservation, voracious appetite for learning. And last but not the least, every single idea mentioned above and thousands more are packaged into a tiny flexible package that has everything needed to not just survive but thrive in this world (and accommodate growth as this child matures into an adult). We are pretty far from achieving any of this.

This list is not perfect in many ways. There are problems that I have left out (some mentioned below). Some of the items can be combined into a single item. AutoML and AI creating AI can be considered a single item mediated by a good dose of life-long learning. On the other hand, some items could be broken down into pieces that are just as not-so-easy to solve — Understanding or a four-year old for example. And, finally, some of the items are vague, like Consciousness, and there is not even agreement that these abilities are crucial to artificial intelligence. The point of this list is that while our current successes in AI might seem remarkable we are far from achieving true artificial intelligence.

Honorable mentions

1. Art — is subjective to a great degree. Who’s to say machine-created art is not real?
2. AI that can do any different things (multi-task) — Most current AI models are specific to a particular task. In the near future we will see networks that can perform many tasks. But networks that can perform an arbitrary number of different tasks are a long way off.
3. Self-learning systems — depending on how this is defined, it could either be easy (AutoML) or not-so-easy (Reasoning)
4. Deciding priorities — Robot is taught not to roll over the expensive persian rug. Owner is dying on the rug but the robot has to roll over the rug to get the pills to them. An extremely unlikely scenario but we face (are assailed by?) many many small decisions every day and a robust prioritizing mechanism is needed to avoid analysis paralysis. It may turn out to be a trivial piece because you just flip a coin every time you are faced with a decision and that will work just fine (hey someone should try doing that for a day and report back). Or it could be an easy piece that piggybacks on the probabilities and confidence values that are inherent in ML approaches. In the worst case, it might be a not-so-easy piece that requires some combination of coin-flipping, probabilities, and reasoning.
5. Life-long learning — Networks and models that can be kept updated over long periods of time are achievable in the short-term. Even those that can learn a new skill or two are possible but achieving true flexibility in learning close to anything including using techniques such as self-learning are very far away.

Disclaimer: The views expressed in this article are my own and do not necessarily represent the views of my employer.