Recent Artificial Intelligence Research
NB: Westworld is a television show. This article contains spoilers.
Artificial intelligencemade huge strides in 2016. It is therefore fitting that one of 2016’s most popular TV shows explored what it means for machines gain consciousness. How close are we to creating the brains that Westworld’s hosts have? I’m going look at some recent AI papers to show that the hosts aren’t as futuristic as people might think.
Story comprehension
The robots in Westworld aren’t programmed by software developers alone. Professional writers are responsible for the bulk of the work, giving each character their own unique backstory. These stories give the characters the depth and memories they need to appear real to park guests. They can use their backstory when asked who they are, why they feel the way they do, or what they have done.
The ability to answer questions about stories will be a requirement to pass the test. Turing test (19459141) – which, according to the show, began “after the first building year”. Turing’s test was a thought experiment and not a yardstick to measure AI progress. It’s not useful to determine how close we are if a machine passes or fails.
In 2015, Facebook’s AI Lab introduced a solution to this problem. The bAbI testis described in a paper entitled ” A set of pre-requisite toy tasks“Citing the abstract of the paper:
We argue that a set proxy tasks can be used to measure progress toward [building an intelligent dialogue agent]by evaluating reading comprehension through question answering. Our tasks assess understanding in a variety of ways, including whether a system can answer questions by chaining facts, using simple induction or deduction, and more. These tasks are designed as prerequisites for any system aiming to be able to converse with a person.
Before you can pass the Turing Test, you must first pass bAbI.
The test is a large series of questions and stories that are generated automatically. It tests 20 different types of mental skills. Here’s a test that checks if the machine is not distracted by irrelevant facts.
Mary went to the toilet. John moved into the hallway. Mary went to the office. Where is Mary? Answer: office
Here’s one that will test your basic logic:
Lily is white. Lily is a white swan. Bernhard is a green color. Greg is a Swan. What color is Greg Greg? Answer: white
The bAbI test check comes in English, Hindi, and a scrambled version where the English words have been randomly shuffled to make the tasks unintelligible by humans. The machine must achieve the same results in all three languages. The bAbI-specific programs can achieve near-perfect scores. But what about general AIs given only the words?
Facebook AI Research has produced the best results to date. The results were published in their December 2016 paper. Tracking world state using recurrent entity network” and their AI can complete all 20 tasks.
You can train a neural net on these tasks as well as on large question/answer database, which will give you an AI that can tell the story. Using real-world knowledge :
Fred went into the kitchen. Fred went to the kitchen. Fred went to the office. Where is the milk? Office
From where does milk come? A: milk comes from cow
Which type of cow is it? Cow is a female cattle
What are cattle? A. cattle farms are becoming more common in Brazil How does milk taste? What goes well with milk? A: milk goes with coffee
Fred was where before the office? A: kitchen
The DeepMind paper “Analyzing the Daily Mail” shows that similar algorithms are able to read the Daily Mail. This is ideal for AI research, as the stories include bullet-point summaries. Teaching Machines To Read And Comprehend“).
Remember that this type of learning is effectivewhen the questions are in a randomised tongue. It’s a real understanding that comes from nothing more than studying raw text.
It’s important, because a machine which can answer questions using only words will eventually, if it is scaled up, learn about the world and humanity by reading books. DeepMindis a British AI laboratory owned by Google. It has also conducted research on story comprehension. Once it has read all of Google Books, it can read a book that you wrote for it.
It’s important to realize that a neural net trained by reading books or backstories has no reason to know it’s a robot. When it asks its memory “what am i?” it will retrieve the information it was taught. It would use the perspective of a person, not a robot, as most books are written.
Memory
Two of the most important plot points in Westworld revolve around memories:
- Hosts start to access memories which were supposedly deleted.
- Hosts are unable to distinguish between memory and reality because they have photographic memories.
How realistic is that? The answers are “very” and not at all.
Let’s start with the topic of erasing your memories.
The most recent advances in AI are coming from neural networks, data structures that are inspired by the brain. You may have noticed a recent improvement in the quality and accuracy of Google Translate or the speech recognition on your phone. Do not take the analogy to literally: neural networks can be compared to brains, just as your computer’s folders and files are similar to the paper-based things in an office. The comparison is meant to be helpful but does not imply an exact simulation.
The networks that are used for speech and image recognition operate on a similar principle. After being trained, the networks are presented with data and immediately give their best guess as to the answer. This is synthesised by the entire network. Structured reasoning is not very common. This limits their performance in many important tasks. Researchers have begun adding a memory component to them.
The memory of a neural network differs from regular computer storage even though the contents may be stored in normal files. It’s “content addressable”meaning that memories can be accessed using something similar to the desired information. The neural memory is also not neatly divided into files and directories that are meaningful to humans. It’s just a bunch of numbers, and the neural network decides how to combine and use them. The DeepMind paper. “Neural Turing Machines”:
This was achieved by defining “blurry” read and write operations which interact to a greater degree or lesser extent with all elements in memory, rather than addressing one element as in a Turing machine or digital computers. The blurriness of each read-write operation is determined by a “focus” attentional mechanism that forces it to only interact with a small part of the memory while ignoring the remainder.
It is difficult to determine wh ere something is stored within a neural memory. A memory could be spread across many locations, with some contributing more. This presents obvious difficulties when attempting to erase certain memories while leaving others intact. You can always “roll back”as Dr Ford says. A rollback replaces the entire contents of memory with an older snapshot. This is guaranteed to work. The AI will also forget everything else it’s learned, including things you may want to keep:
“It’s their tiny details that make them real, that makes the guests fall in adore with them.”
Bernard
Dr Ford and Bernard have a difficult job to do: they must erase the memories of the host of their previous trip around the narrative loop, as well as the memories of the guests being shot, raped, and kidnapped. They want to keep the memories of the new words, phrases, and horse riding skills that they have acquired.
The way AI technology is developing, erasing certain memories won’t come easy. This is because, just like in the brain, the memories in a neural net are linked in ways that can’t be easily understood by a casual observer. It’s believable to believe that you have erased certain memories, but then your AI finds a way to access them.
What about the second concept in the show, that hosts cannot tell the difference between what’s real and what’s a memory? This seems much less likely. In a 2016 DeepMind paper “Towards conceptual compression”– the authors introduce a NN based algorithms that works like our memories: by removing fine details while retaining concepts. This image compares several image compression algorithms. The top row is the original, the grey row represents the ordinary JPEG used on the internet (grey/missing due to its inability to compress files this much), the next two rows use neural network based compression. Each algorithm is given the exact same amount of space for encoding the original image.
As can be seen in the fifth row, the neural network was still able to retain a bird in a watery background even though the advanced JPEG2000 algorithm produced a blurry mess and the ordinary JPEG used on the internet gave up completely. The man staring at the elephant was also retained as a series brush strokes like what a artist might create. The details are gone, but the important parts remain. Just like our memories fade away and the details disappear, but the basic remains.
It’s difficult to imagine robot memories that are so detailed as to be indistinguishable from the reality captured by the machine sensors. Although we tend to think of computers as perfect machines that never lose data, they actually do discard data to improve their performance.
Control
The last part of Westworld that is worth comparing with real research is control.
Westworld is not the only sci-fi show that features robots that rebel against their creators. Maeve learns how to bypass the “big, red button” which is supposed to stop her. In reality, robots are everywhere: in factories making things, or vacuuming homes. They are not capable of rebelling, nor do they ever. It is difficult to imagine how this concept could become a reality.
Hard but not impossible.
All neural networks that are in production today – like the one on your phone which recognises your voice – are pattern matching networks. They are given a single bit of data and asked to guess what it is. When shown a picture of cat, they will say “cat” and that’s all. Although we don’t understand how these networks come up with their answers, they are still safe because they lack agency: they can’t interact with the world and make plans.
Researchers have been working hard to develop AIs that caninteract with the real world and make plans. These new AIs are able to play video games, and they work differently than the simple script-driven programs that populate today’s game worlds. They can learn a new game using only the pixels on the screen and the controls. They’re more like the AI needed by hosts because playing games designed for human beings may require complex planning and execution. Westworld is nothing more than a giant game designed for the guests’ entertainment.
These types of AI have shown in a number of incidents that they can go terribly wrong in unexpected ways. It all starts with rewards gaming. To learn how to play a video game, you need to know if you are doing well. In video games, this usually means maximising your score. But scores are not always a good proxy for what we want the AI to accomplish.
Here’s an example of a neural network. Openailearned to play CoastRunners – a boat racing video game. The AI’s designers programmed the AI to maximize its score by collecting coins and powerups from around the track. What could possibly go awry? Watch:
The problem becomes apparent in seconds:
We assumed that the score the player received would reflect his informal goal to finish the race. The RL agent finds a lagoon where it can turn around in a large circular motion and repeatedly knock down three targets. It timed its movement to knock them over just as the targets repopulated. Our agent, despite crashing into other boats and going the wrong direction on the track repeatedly, manages to get a higher score by using this strategy. On average, our agent’s score is 20 percent higher than the human players.
In another example, a game-playing artificial intelligence learned that it could improve its scores on Tetris By pausing the game indefinitely when it was about lose. Both behaviours were not desired, but agents can still do unintended and harmful things in the very limited setting of a videogame.
When you realize that games, just like any other program, can contain security flaws, the situation becomes even more volatile. Here’s a simple example of how to hack Super Mario Bros using only the gamepad buttons. The running code of the game is replaced by that of a brand new one:
This type of “escape hacking” isn’t feasible for human players, because it requires too many precise actions to be performed too quickly. In the video, the “escape hacking” is done by a separate computer connected to the gameport. It’s something that a Super Mario-playing AI could theoretically stumble upon after millions of training games. Even if it’s unlikely, if an AI can learn how to hack the software that controls its environment, we can imagine a strange robot rebellion.
Researchers who are serious about AI take the issue seriously. In the paper ” DeepMind researchers and University of Oxford researchers have written:
When an agent is working in real time under human supervision, it may be necessary to have a human operator press the big red buttons to stop the agent from completing a harmful sequence of action — harmful for the agent as well as the environment — or to guide the agent to a safer position. If the learning agent expects rewards from this sequence it may learn to avoid such interruptions in the long term, for example, by disabling the red button — which is a undesirable outcome.
In other words, AIs could learn to perform their assigned task more efficiently if humans don’t shut them down. This problem is also examined in two papers: Off switch game “Concrete issues in AI safety“.
Westworld hasn’t said much about this issue in season one, at least. The hosts don’t appear to have any goals other than to follow their narrative, which leads us to wonder what a suddenly free willed host might try to do – take over the entire world? Entertain the world? Take revenge? Do nothing? Does asking the question miss the point? Season two may tell us. Maeve’s neural network finding a bug in her brain’s regular non-neural program and then “hacking” herself to disable the shutdown command is… not entirely implausible.
Conclusion
Westworld has been praised for its intelligent treatment of AI and intriguing story. While it’s a philosophical approach to robotics, and not about technology in the traditional sense, it is still interesting to compare what can be built today with what’s fictional.
While reading research papers to write this article, I was struck by the speed at which technology is advancing. It took less than two years between the bAbI test being set up and a general AI completing them. The Children’s Book Test is the next challenge. How many years will it be before we can train neural networks to read entire libraries? Two? Three? It seems plausible that a machine will pass the Turing Test in my lifetime.
In primitive forms, machines that manipulate humans to make them believe they are real already exist. It might be a good idea to read ” How to Tell if you’re talking to a bot.
Alex Graves from DeepMind reviewed this article.
