Import AI: #95: Learning to predict and avoid internet arguments, White House announces Select Committee on AI, and BMW trains cars to safely change lanes
by Jack Clark
Cornell, Google, and Wikimedia researchers train AI to predict when we’ll get angry on the internet:
…Spoiler: Very blunt comments with little attempt made at being polite tend to lead to aggressive conversations…
Have you ever read a comment addressed to you on the internet and decided not to reply because your intuition tells you the person is looking to start a fight? Is it possible to train AI systems to have a similar predictive ability and thereby create automated systems that can flag conversations as having a likelihood of spiraling into aggression? That’s the idea behind new research from Cornell University, Jigsaw, and the Wikimedia Foundation. The research tries to predict troublesome conversations based on a dataset taken from the discussion sections of ‘Wikipedia Talk’ pages.
Dataset: To carry out the experiment, the researchers gathered a total of 1,270 conversations, half consisting of ones which became aggressive following the initial comments, and half consisting of ones which remained civil. (Categorizing civil versus on-track was done via a combination of the use of Jigsaw’s “Perspective” API, and gathering labels from humans via CrowdFlower.) These conversations had an average length of 4.6 comments.
How it works: Armed with this dataset, the researchers characterized conversations via what they call “pragmatic devices signalling politeness”. This is a set of features that correspond to whether the conversation includes attempts to be friendly (liberal use of ‘thanks’, ‘please’, and so on), along with words used to indicate a position that welcomes debate (eg, by clarifying statements with phrases like “I believe” or “I think”). They then study the initial comment and see if their system can learn to predict whether it will yield negative comments in the future.
Results: Humans are successful about 72% of the time at predicting nasty conversations from this dataset. The system designed by these researchers (which relies on logistic regression – nothing too fancy) is about 61.6% accurate, and baselines (bag of words and sentiment lexicon) get around ~56%. (One variant of the proposed technique gets accuracy of 64.9%, but this is a little dubious as it is trained on way more data and it’s unclear whether it is overfitting, as it is also trained on the same data corpus.) The researchers also derive some statistical correlations that could help humans as well as machines better spot comments that are prone to spiral into aggresion. “We find a rough correspondence between linguistic directness and the likelihood of future personal attacks. In particular, comments which contain direct questions, or exhibit sentence initial you (i.e., “2nd person start”), tend to start awry-turning conversations significantly more often than ones that stay on track,” they write. “This effect coheres with our intuition that directness signals some latent hostility from the conversation’s initiator, and perhaps reinforces the forcefulness of contentious impositions.”
Why it matters: Systems like this show how with a relatively small amount of data it is possible to build classification systems that can, if paired with the right features, effectively categorize subtle human interactions online. While here such a system is used to do something that seems to be for the purpose of social good (figuring out how to identify and potentially avoid aggressive conversations), it’s worth remembering that a very similar approach could be used to, for instance, identify conversations where initial comments could correlate to conversations that have a high chance of displaying political views that are contrary to those views of the people building such systems, and so on. It would be nice to see an acknowledgement of this in the paper itself.
Read more: Conversations Gone Awry: Detecting Early Signs of Conversational Failure (Arxiv).
Chinese researchers tackle Dota-like game King of Glory with RL + MCTS:
…Tencent researchers take inspiration from AlphaGo Zero to tackle Chinese MOBA King of Glory…
Modern multiplayer strategy games are becoming a testbed for reinforcement learning and multi-agent algorithms. Following work by Facebook and DeepMind on StarCraft 1 and 2, and work by OpenAI on Dota, researchers with the University of Pittsburgh and Tencent AI Lab have published details on an AI technique which they evaluate on King of Glory, a Tencent-made massively multiplayer online battle arena (MOBA) game. The proposed system uses Monte Carlo Tree Search (MCTS – a technique also crucial to DeepMind’s work on tackling the board game Go) and incorporates techniques from AlphaGo Zero to “to produce a stronger tree search using previous tree results”. “Our proposed algorithm is a provably near-optimal variant (and in some respects, generalization) of the AlphaGo Zero algorithm” they write.
Results: The researchers test out their technique within King of Glory by evaluating agents trained with their technique against other agents controlled by the in-game AI. They also test it against four variants of their proposed technique which, respectively: have no rollouts; use direct policy iteration; implement approximate value iteration; and one trained via supervised learning on 100,000 state-action pairs of human gameplay data. (This also functions as a basic ablation study of the proposed technique, also). Their system beats all of these approaches, with the closest competitor being the variant with no rollouts (this one also looks most similar to AlphaGo Zero).
Things that make you go hmmm: Researchers still tend to attack problems like this by training the AI systems over a multitude of hand-selected features, so it’s not like these algorithms are automatically inferring optimal inputs from which to learn from. “The state variable of the system is taken to be a 41-dimensional vector containing information obtained directly from the game engine, including hero locations, hero health, minion health, hero skill state, and relative locations to various structures,” they write. A lot of human ingenuity goes into selecting these inputs and likely adjusting hyperparameters to denote the importance of any particular input, so there’s a significant unacknowledged human component to this work.
Why it matters: This paper provides more evidence that AI researchers are going to use increasingly modern, sophisticated games to test and evaluate AI systems. It’s also quite interesting that this work comes from a Chinese AI lab, indicating that these research organizations are pursuing similarly large-scale problems to some labs in the West – there’s more commonality here than I think people presume, and it’d be interesting to see the various researchers come together and discuss ideas in the future about how to tackle even more advanced games.
Read more: Feedback-Based Tree Search for Reinforcement Learning (Arxiv).
Today’s AI amounts to little more than curve-fitting, says Turing Award winner:
…Judea Pearl is impressed by deep learning success, but worries researchers have become complacent about inability to deal with causality…
Turing Award-winner Judea Pearl is concerned that the AI industry’s current obsession with deep learning is causing it to ignore harder problems, like developing machines that can build causal models of the world. He discusses some of these concerns in an interview with Quanta Magazine to discuss his new book “The Book of Why: The New Science of Cause and Effect“.
– “Mathematics has not developed the asymmetric language required to capture our understanding that if X causes Y that does not mean that Y causes X.”
– “As much as I look into what’s being done with deep learning, I see they’re all stuck there on the level of associations. Curve fitting.”
– “We did not expect that so many problems could be solved by pure curve fitting. It turns out they can. But I’m asking about the future — what next? Can you have a robot scientist that would plan an experiment and find new answers to pending scientific questions? That’s the next step.”
– “The first step, one that will take place in maybe 10 years, is that conceptual models of reality will be programmed by humans..the next step will be that machines will postulate such models on their own and will verify and refine them based on empirical evidence.
Read more: To Build Truly Intelligent Machines, Teach Them Cause and Effect (Arxiv).
Google prepares auto-email service “Smart Compose”:
…Surprisingly simple components lead to powerful things, given enough data…Google researchers have outlined the technology they’ve used to create ‘Smart Compose’, a new service within Gmail that will automatically compose emails for people as they type them. The main ingredients are a Bag of Words model and a Recurrent Neural Network Language Model. This combination of technologies leads to a system that is “faster than the seq2seq models with only a slight sacrificed to model prediction quality”. These components are also surprisingly simple, indicating just how much can be achieved when you’ve got access to a scalable technique and a truly massive dataset. Google says that by offloading most of the computation onto TPUs it was able to reduce the average latency to tens of milliseconds – earlier experiments showed it that latencies higher than 100 milliseconds or so led to user dissatisfaction.
Read more: Smart Compose: Using Neural Networks to Help Write Emails (Google Blog).
White House plans Select Committee on AI:
…Hosts summit between AI and industry experts, reinforces regulatory-light approach to tech…
The White House recently hosted a “Summit on AI for American Industry”, bringing together industry, academia, and government, to discuss how to support and further artificial intelligence in America. A published summary of the event from the Office of Science and Technology Policy highlights some of the steps this administration has taken with regard to AI – much of the actions include the elevation of AI in White House communications as a strategic area, with more mentions of it in documents ranging from the National Defense and National Security Strategy, to guidance from the Office of Management and Budget (OMB) given to agencies.
Select Committee on AI: The White House will create a “Select Committee on Artificial Intelligence”, which will primarily be comprised of “the most senior R&D officials in the Federal government”. This committee will advise the White House, facilitate partnerships with industry and academia, enhance coordination across the Federal government on AI R&D, and identify ways to use government data and compute resources to support AI. The committee will feature staff from OSTP, the National Science Foundation, the Defense Advanced Research Projects Agency, the director of IARPA, and others. The committee may call upon the private sector as well, according to its charter.
Regulation: In prepared remarks OSTP Deputy US Chief Technology Officer Michael Kratsios said “Our Administration is not in the business of conquering imaginary beasts. We will not try to “solve” problems that don’t exist. To the greatest degree possible, we will allow scientists and technologists to freely develop their next great inventions right here in the United States. Command-control policies will never be able to keep up. Nor will we limit ourselves with international commitments rooted in fear of worst-case scenarios.”
Why it matters: Around the world, countries are enacting broad national strategies relating to artificial intelligence. France has committed substantially far more funding relative to its existing funding amount to AI than other countries, and China (which by virtue of its governance structure will tend to out-spend Western countries on broad science and technology developments) has committed many additional billions of dollars of funding to AI. It remains to be seen whether the US’s strategy of leaving the vast amount of AI development to the private sector is the optimal decision, given the immense opportunities the technology holds and its demonstrable responsiveness to additional infusions of money. America also has some problems with its AI ecosystem that aren’t being dealt with today, like the fact that many of academia’s most creative professors are being drawn into industry at the same time as class sizes for undergraduate and graduate AI courses are booming and PHD applications are spiking, reducing the quality of US education in AI. It’d be interesting to see what kinds of recommendations the Select Committee makes and how effective it will be at confronting the contemporary challenges and opportunities faced by the administration with regard to US AI competitiveness.
Read more: Summary of the 2018 White House Summit on Artificial Intelligence for American Industry (White House OSTP, PDF)
Democrat Representative calls for National AI Strategy:
…Points to European, French, Chinese efforts as justification for US action…
Congressman John Delaney (Maryland) has written an op-ed in Wired calling for a National AI Strategy for the US. Delaney has himself co-sponsored a bill (along with Republican and Democrat congresspeople and senators) calling for the creation of a commission to device such a strategy, called the FUTURE of AI Act (Fundamentally Understanding the Usability and Realistic Evolution of Artificial Intelligence Act).
– “The United States needs a full assessment of the state of American research and technology, and what the short and long-term problems and opportunities are.”
– “Whether you are a conservative or a progressive, this future is coming. As I look at where the world is headed, I believe that we need to expand public investment in research, encourage collaboration between the public and private sector, and make sure that AI is deployed in a way that is wholly consistent with our values and with existing laws.”
– ” If the US doesn’t act, we’re in danger of falling behind.”
Why it matters: Societies across the world are changing as a consequence of the deployment of artificial intelligence, whether through unparallelled opportunities for providing better healthcare and accessibility services to citizens, to being able to utilize the same technologies for surveillance and various national security purposes. It seems to intuitively make sense to survey the whole AI field and look for ways that a country can implement a holistic plan. It seems likely that there will be a bunch of complementary initiatives in the US, ranging from targeted actions like those espoused by the OSTP, to broader analyses performed by other parties, like the Senate, or government agencies.
Read more: France, China, and the EU all have an AI strategy, shouldn’t the US? (Wired Opinion).
Learning to lane change with recurrent neural networks:
…BMW researchers try to teach safe driving via seq2seq learning…
Researchers with car company BMW and the technical university of Munich in Germany have trained simulated self-driving car AI agents in a crude simulation to learn how to lane change safely. They achieve this by implementing a bidirectional RNN with long short-term memory, which learns to predict the velocity of a car and its surrounding neighbors at any point in time, then uses this prediction to work out if it will be safe for the vehicle to change into another lane.
Results: The system is evaluated against the NGSIM dataset, a detailed traffic dataset taken from monitoring real traffic in LA in the mid-2000s. It outperforms other baselines but, given the restricted nature of the domain, the lack of an ability to compare performance against (secret) systems developed by other automotive experts, and the absence of integration with a deeper car simulation, it’s not clear how well this result will transfer to real domains.
Why it matters: All cars are becoming significantly more automated, regardless of the overall maturity of full self-driving car technology. Papers like this give us a view into the development of increasingly automated vehicular systems that use components developed by the rest of the AI community.
Read more: Situation Assessment for Planning Lane Changes: Combining Recurrent Models and Prediction (Arxiv).
I guess we should have expected them, these billionaire cities. They started sprouting up after the price of basic space travel came down enough for billionaires to build their own launchpads, letting them mesh their business and life enough to create miniature cities to tend to their numerous inter-locking businesses. Many of these cities were built in places far above sea level, in preparation for an expected dire climate future.
These cities always had a few common components: a datacenter to host secure data and compute services, frequently also running local artificial intelligence services; automated transit systems to ferry people around; fleets of drones providing constant logistics and surveillance abilities; goods robots for heavy lifting; robotic chefs; and even a few teams of humans, which tended to these machines or spoke to other humans or worked in some other manner for the billionaire.
These cities grew as the billionaires (and eventually trillionairs) competed with eachother to build ever more sophisticated and ever more automated systems. Soon after this competition began, we heard the first rumors of the brain-interface projects.
Teams of people were said to be hired by these billionaires to work within these by-now almost entirely automated gleaming cities. The people were paid gigantic sums of money to sign themselves away for contracts of two to three years, and to be discrete about it. Then the billionaire would fly-in teams of surgeons and have them perform brain surgery on the people, giving them interfaces that let them plug in to the data feeds of the city, intuitively sensing them and being able to eventually learn to understand them. It was said that arrangements of this kind, with the digital AI of the city and the augmented human brains interlinked, led to superior performance and flexibility to other systems.
We have recently heard rumors of other things – longer contracts, more elaborate surgeries, but those are as yet unsubstantiated.
Things that inspired this story: Brain-machine interfaces, Gini coefficient, spaceships with VTOL capability, cybernetics.