The Buzzkilling Dystopians
By Alberto Moel, Vice President Strategy and Partnerships, Veo Robotics
Welcome back, dear reader, to yet another installment of our thoughts on the raging debate concerning the macroeconomic impacts of automation. As always, please refer to our 2x2 matrix of the possible positions.
If you are in need of a refresher, please review our previous posts on pop culture automation, the techno-optimists, techno-pessimists, and traditionalists. And if you’re still in the mood, check out what automation could have in store for developing nations.
In today’s post we discuss the dystopians, who sell the most books by featuring apocalyptic visions of AI taking over the world and consigning humanity to the scrap heap, or perhaps to permanent servitude or slavery.
The dystopians have a very straightforward argument on the inevitability of the robopocalypse: Computers are getting more and more powerful and are increasingly capable of performing a greater number of previously human-only tasks. As computers increase their skills, these capabilities will unify and stack up until machines are far smarter than people, at which point human labor (and perhaps humans) will become unnecessary.
My favorite dystopian is, of course, Yuval Noah Harari. Here are some Debbie Downer quotes for you to ponder from his buzzkill of a book, "Homo Deus”:
As old professions became obsolete, new professions evolved, and there was always something humans could do better than machines. Yet this is not a law of nature, and nothing guarantees it will continue to be like that in the future.
The most important question in 21st century economics may well be what to do with all the superfluous people. What will conscious humans do, once we have highly intelligent non-conscious algorithms that can do almost everything better?
In the 21st century we might witness the creation of a new massive class: people devoid of any economic, political, or even artistic value, who contribute nothing to the prosperity, power and glory of society.
Yikes! Let’s assume that this outcome has a non-zero probability of occurring. After all, human intelligence and consciousness are housed in actual physical matter, made of atoms and subatomic particles and all that, so why wouldn’t it be possible to replicate that in a machine?1
Many otherwise smart people seem to think this non-zero probability of AI eating the world is high enough and happening soon enough to be of concern.2 On the other hand, the probability of humanity being wiped out in a meteorite strike is also non-zero. But is it something we should be worrying about? Probably not.
Let’s pick apart the dystopians’ argument by considering the following:
- How certain machine learners’ exceptional performances were achieved and where they fell short of human capability;
- How machine intelligence is different from what we commonly "know"3 as human intelligence (and what is missing); and
- Why these two elements mean that the possibility of machines somehow “catching up” to humans is highly unlikely, even far into the future.
What machines are capable of now
Let's take a closer look at three examples of machine learners that have widely made the news and wowed the general public: the DeepMind DQN machine learner for video games, the AlphaGo Go "world champion," and state-of-the-art image and character recognition learners. What these three have in common (with pretty much all other leading-edge learners) is that they are notoriously data- and power-hungry and have limited true "human" intelligence.
DeepMind and the Atari Challenge
In February 2015, DeepMind announced that it had trained its "deep-Q network" (DQN) to play 49 classic Atari 2600 console games, achieving 100% of human-level performance or better in 29 of the games4 The DQN learned to play the games by combining a deep convolutional neural network acting as a pattern recognizer and a 20-year-old model-free reinforcement learning algorithm (Q-learning).5
It was pretty clear that although the DQN was presented with a "blank" rulebook of the games, it was learning to play them very differently from how humans do. For each game, the DQN was trained on 200 million frames, the equivalent of 924 hours (38 days) of game time. And as the DQN played and trained, it revisited each of these frames an average of eight more times. For comparison, a benchmark professional gamer received two hours of training per game.
Even with the extensive training program, the DQN initially achieved lower than 10% of professional human performance. It wasn't able to beat a human until after its program had undergone more tweaking and it had re-trained from scratch.6 And the DQN learned very slowly. After 231 hours of its new training program, its performance was only 46% of that of a human. After 116 hours it was only at 19%, and after two hours it was at a paltry 3.5% (close to random play, 1.5%).7 Even non-professional humans can learn to play better than chance with a few minutes of play.
And what if we changed the rules of the game slightly so that the objective wasn't to get the highest score, but the lowest? Or maybe get the closest to 3,000 points without going over? These variants (and all possible ones humans can think of) would all require the DQN to undergo extensive re-training, as would, for example, changing the background color or the shape of the objects in the game. Not only would the DQN be stumped by all of these variants, it wouldn't even be able to come up with them to begin with.
DeepMind and AlphaGo
The game of Go is very simple to learn but very difficult to master. It is considered more difficult for a machine to play than chess as the number of moves and possible branches make it impossible to capture deterministically.
AlphaGo is a combination deep neural network and tree-search reinforcement learner that plays the game of Go. In March 2016, it beat 9-dan master Lee Sedol 4 to 1 in a widely publicized match.
In its initial training, AlphaGo was fed 160,000 individual games played by human experts, accounting for 28.4 million positions and moves. It then underwent reinforcement learning, where it played 30 million games against itself.8 Between the time of that initial training and the match with Lee Sedol, AlphaGo played successively stronger versions of itself, in the end playing 100 million or more games.9 In comparison, Lee Sedol probably has played around 50,000 games in his career.
How good would AlphaGo be if it had only played 50,000 games in training? If the Atari learner example is any reference, it would probably be pretty lousy. Humans can learn the rules of this simple game just by watching other humans play a few rounds and can quickly adapt their strategies to an infinite number of game variants, where the rules are adjusted just slightly.
Again, just like DeepMind and Atari, AlphaGo not only can't adjust to these variants without extensive re-training, it can't even conceive of these variants. Humans can because we have meta-knowledge that Go is a game, and that the objective is to beat the opponent within the constraints of the game rules, whatever they may be.
Image and character recognition
We know current machine learners are now as good as humans at large-scale image classification tasks, and that's no small feat. These algorithms have become very good at recognizing and classifying objects, but they are still pretty poor at capturing the physics of the scene and causal relationships between identified objects, much less inferring intent from actors in the images.
For example, an otherwise sophisticated image recognition learner identified an image of a plane about to crash as "an airplane is parked on the tarmac at an airport," and one of a man being thrown off a motorcycle as "a man riding a motorcycle on a beach."10
Humans also learn from far fewer examples and can learn many more general classes of objects. This is pretty clear in handwriting recognition. Humans can not only recognize a new character from a single example, but also learn a concept, which allows us to apply the learning in new ways, such as generating new characters given a small set of related observations. More broadly, humans learn a lot more from a lot less, and machine learners still have much to, ahem, learn.
An aside on power consumption
And to wrap up this critique, there is the issue of power consumption. For example, getting AlphaGo to "work" required the full-time effort of some very smart people over many months and a huge amount of hardware. AlphaGo ran on 1,920 CPUs and another 280 GPUs, and at 200 W per compute element. That's 440 kW of power. Lee Sedol, by contrast, is content to get himself going with a couple thousand calories a day, or about 100 W of power—a factor of 4,400 less than what AlphaGo needs (and that 100 W is not just powering his brain, but also the rest of his body).
So saying AlphaGo is better than Lee at Go is like saying an autonomous Formula 1 racecar is faster than a human on a bicycle. Of course it is! It was built specifically and exclusively for that purpose. Is the human better at pretty much everything else? Definitely. In an apples-to-apples comparison, something certainly seems amiss, so we now turn to what is missing.
Where current machine learners fail
Despite recent progress, humans are still better at learning concepts, making sense of scenes, acquiring and deciphering language, and spatial understanding and movement. And let's not even mention creativity, common sense, and general-purpose reasoning and intelligence.
How is human learning different from machine learning? Precisely, what are current machine learners missing? If we can reliably answer these questions, we can probably find engineering solutions to making machine learning less "machine" and more "human."
The leading machine learners all approach learning as statistical pattern recognition and classification, treating data as the primary unit of observation and analysis. Machine learning has evolved in the direction of discovering and classifying these patterns using large amounts of data.11
This is not consistent with how humans learn, which can be characterized more by rapid causal model building from intuitive (and innate) concepts rather than gradual improvements in pattern recognition. These models not only seek to predict nature, but also to explain it.
More broadly,12 truly human-like thinking machines will need to:
- Be able to rapidly build causal models that explain rather than just predict, including the creation of imaginary counterfactuals that would never occur in nature;
- Have a good grasp of basic theories grounded in physics and psychology, to build on the explanatory power of the models; and
- Be able to generalize to new tasks and situations by having the ability of learning-to-learn.
Humans can build causal models on the fly
Humans can process a new scene, spoken word, or physical object in a fraction of a second, from very limited amounts of previous experience. This richness and efficiency suggest that humans are building causal models of the world in real-time, and that this process is representative of how humans learn.
The "output" of these causal models is often symbolic and relational, which means the human brain is going beyond "neural networks" and applying other forms of learning. In particular, these causal models can include “imaginary” counterfactuals where impossible situations that will never happen can be modeled in the human mind.13
Humans are "pre-programmed" with firmware at birth
There is ample evidence that we are born with some substantially pre-trained "firmware" that gives us a serious leg up in the learning game.
Even infants yet unable to speak can distinguish animate agents from inanimate objects.14 Infants can also infer an agent's goals and objectives, and whether those goals are harmful or beneficial to them.15 Building on this basic firmware, young children have an intuitive knowledge of physics, geometry, navigation, numbers, and psychology.
Humans are able to generalize to new tasks and situations
Humans can make inferences that go far beyond the data, which means strong prior domain knowledge is involved. One way this happens is through "learning-to-learn,"16 where learning a new task can be accelerated through previously learned tasks.
Humans can make inductive analogies from previous knowledge and extend learning to new situations or novel concepts very easily. This may sound patently obvious, but machine learners developed to date have extremely limited capability to do this, and we are only now starting to see progress.
Machine learning is not artificial general intelligence
Current machine learning techniques and engineering coupled with increasing computational power, improved algorithms, and bigger and more available training datasets will continue to make great strides over the next few years.
But let's get one thing clear: The existing knowledge base of machine learning isn't going to lead to any form of artificial general intelligence. A machine learner achieving human or super-human performance at a specific task (say, playing checkers or assembling cars) doesn't mean that the same machine learner is capable of human-level general intelligence, or even of super-human performance in a completely different task (say, driving the car it just built).
This machine learner is also not capable of generalizing, having independent thoughts of its own, or reaching general inferences from and imagining counterfactuals to observations in the material world. It is also nowhere near to exhibiting consciousness or experiencing emotions or sensations. Machine learners will, over time, be more human-like in their performance and capabilities. But they cannot think like humans do, and probably never will.
Therefore, despite the dystopians’ grumblings and hand-wringings, we really don’t need to worry about one day bowing down to malicious robot overlords. What we should actually be worried about (besides some badly-programmed AI wreaking havoc), are the broader effects of gradual automation on the world’s economies.
So we return to our regularly scheduled programming in our next blog post and consider the causal relationship between employment and robot adoption—which way does causality go? Does robot adoption lower labor demand, or does lower labor supply spur robot adoption?
1 This is the argument that Nobel-prize winning physicist (and seriously smart person) Frank Wilczek makes in Possible Minds: 25 Ways of Looking at AI, a collection of 25 essays on AI edited by John Brockman. This argument has been also extended into a full-length book by physicist Max Tegmark (another seriously smart person) in Life 3.0: Being Human in the Age of Artificial Intelligence.
2 Unfortunately, the dystopian argument is so easy on the brain that it’s picked up an air of inevitability. As Rodney Brooks writes in his May 17, 2019, blog post AGI has been delayed: “Here is a press report on a conference on ‘Human Level AI’ that was held in 2018. It reports that 37% of respondents to a survey at that conference said they expected human-level AI to be around in 5 to 10 years. Now, I must say that looking through the conference site I see more large hats than cattle, but these are mostly people with paying corporate or academic jobs, and 37% of them think this.”
3 In the true Justice Potter Stewart definitional sense when referring to obscenity.
4 Mnih et al., Human-level control through deep reinforcement learning, Nature 518, 2015.
5 Watkins, C. J., and Dayan, P., Q-learning, Machine Learning 8, 1992.
6 Wang et al., Dueling network architectures for deep reinforcement learning, 2016.
7 Schaul et al., Prioritized experience replay, 2015.
8 Silver et al., Mastering the game of Go with deep neural networks and tree search, Nature 529, 2016.
9 Lake et al., Building machines that think like people, 2016.
10 Karpathy, A., and Li, F., Deep visual-semantic alignments for generating Image descriptions, 2015, as cited in Figure 6 of Lake et al., Building machines that think like people, 2016.
11 As the late great Patrick Winston would say, it would be more helpful to describe the recent developments as being advances in “computational statistics” rather than in AI.
12 In the interest of full disclosure, these ideas aren't mine. I am not knowledgeable enough to come up with them by myself. They come from Lake et al., Building machines that think like people, 2016.
13 This argument is explained in book-length form by Judah Pearl and Dana Mackenzie in The Book of Why: The New Science of Cause and Effect.
14 See for example Johnson, S. C., Slaughter, V., and Carey, S., Whose gaze will infants follow? The elicitation of gaze-following in 12-month-olds, Developmental Science 1, 1998.
15 See for example Hamlin, K. J., Ullman, T., Tenenbaum, J., Goodman, N. D., and Baker, C., The mentalistic basis of core social cognition: Experiments in preverbal infants and a computational model, Developmental Science 16, 2013.
16 Harlow, H. F., The formation of learning sets, Psychological Review 56, 1949.