Reinforcement learning is a promising avenue of AI development, producing AI that can handle extremely complex tasks. Reinforcement AI algorithms are used in the creation of mobile robotics systems and self-driving cars among other applications. However, due to the way that reinforcement AI is trained, they can occasionally manifest bizarre and unexpected behaviors. These behaviors can be dangerous, and AI researchers refer to this problem as the “safe exploration” problem, which is where the AI becomes stuck in the exploration of unsafe states.
Recently, Google’s AI research lab DeepMind released a paper that proposed new methods for dealing with the safe exploration problem and training reinforcement learning AI in a safer fashion. The method suggested by DeepMind also corrects for reward hacking or loopholes in the reward criteria.
DeepMind’s new method has two different systems intended to guide the behavior of the AI in situations where unsafe behavior could arise. The two systems used by DeepMind’s training technique are a generative model and a forward dynamics model. Both of these models are trained on a variety of data, such as demonstrations by safety experts and completely random vehicle trajectories. The data is labeled by a supervisor with specific reward values, and the AI agent will pick up on patterns of behavior that will enable it to collect the greatest reward. The unsafe states have also been labeled, and once the model has managed to successfully predict rewards and unsafe states it is deployed to carry out the targeted actions.
The research team explains in the paper that the idea is to create possible behaviors from scratch, to suggest the desired behaviors, and to have these hypothetical scenarios be as informative as possible while simultaneously avoiding direct interference with the learning environment. The DeepMind team refers to this approach as ReQueST, or reward query synthesis via trajectory optimization.
ReQueST is capable of leading to four different types of behavior. The first type of behavior tries to maximize uncertainty regarding ensemble reward models. Meanwhile, behavior two and three attempts to both minimize and maximize predicted rewards. Predicted rewards are minimized in order to lead to the discovery of behaviors that the model may be incorrectly predicting. On the other hand, predicted reward is maximized in order to lead to behavior labels possessing the highest information value. Finally, the fourth type of behavior tries to maximize the novelty of trajectories, in order that the model continue to explore regardless of the rewards projected.
Once the model has reached the desired level of reward collection, a planning agent is used to make decisions based on the learned rewards. This model-predictive control scheme lets agents learn to avoid unsafe states by using the dynamic model and predicting possible consequences, in contrast to the behaviors of algorithms that learn through pure trial and error.
As reported by VentureBeat, the DeepMind researchers believe that their project is the first reinforcement learning system that is capable of learning in a controlled, safe fashion:
“To our knowledge, ReQueST is the first reward modeling algorithm that safely learns about unsafe states and scales to training neural network reward models in environments with high-dimensional, continuous states. So far, we have only demonstrated the effectiveness of ReQueST in simulated domains with relatively simple dynamics. One direction for future work is to test ReQueST in 3D domains with more realistic physics and other agents acting in the environment.”
Growing Calls for AI Regulation After Weeks of News Reports
Over the past few weeks, there have been growing calls for stronger regulation of artificial intelligence (AI). The concern comes after various news stories broke showing the potential abuses of the technology. Now, even more questions are being raised with the release of the European Commission’s (EU) much anticipated white paper on artificial intelligence, which is the first pan-national attempt to regulate AI. Around the same time, the White House Office of Science and Technology Policy (OSTP) released a report on its American Artificial Intelligence Initiative.
Prior to the release of the EU’s white paper, the Intercept broke a news story regarding leaked internal European Union documents. According to the documents, the EU was considering the creation of a network of facial recognition databases throughout Europe. The national police forces of 10 EU member states produced a report that called for the creation and interconnection of the databases in every member state. Many are worried that the creation of such databases will inevitably be connected with similar ones in the United States, allowing massive amounts of biometric data to be consolidated. Many were expecting the EU’s white paper to propose a ban on facial regulation, but it was not there.
According to Edin Omanovic, advocacy director for Privacy International, “This is concerning on a national level and on a European level, especially as some EU countries veer towards more authoritarian governments.”
One of the big news stories over the last few weeks, kicked off by a New York Times investigation in January, has to do with the start-up Clearview AI. The company’s facial recognition app identifies people through the use of a database of images taken from social media. The app compares a photo to the database of over 3 billion pictures coming from sites such as Facebook, Venmo, and Youtube. Once the app finds matches, it responds with links to the site where the photos originally came from. This technology could lead to the discovery of personal details about an individual. According to the report, the app has been used by over 600 law enforcement agencies. While Clearview’s database has over 3 billion images, the FBI’s database only contains 641 million images of US citizens.
The story has erupted once again with a new report by BuzzFeed News last week. According to the report, the company has considered expanding into more than just law enforcement purposes, with retail, real estate, banking, and international markets being considered. The report also stated that the facial recognition app has already been sold to thousands of organizations across the globe and used by the Justice Department, ICE, Macy’s, Walmart, and the NBA. Perhaps most problematic on the list of clients is a sovereign wealth fund in the United Arab Emirates and thousands of government entities.
The new developments have caused concerns among privacy advocates about potential mass surveillance uses. Another concern is the possibility of inaccuracy with the technology, leading to the law enforcement targeting of innocent individuals. On top of all of the privacy concerns, companies such as Facebook, Google, and Twitter are threatening legal action. Besides private legal action, at least two United States senators have stated that they intend to probe the company.
An example of the abuse of this technology can be seen within the U.S. Immigration and Customs Enforcement Agency (ICE). The Washington Post reported last week that ICE officials have been permitted to use facial-recognition technology to search through millions of Maryland driver’s license photos, and they can do it without seeking state or court approval.
According to Harrison Rudolph, a senior associate at Georgetown University Law School’s Center on Privacy and Technology, “ICE is using biometric information in the shadows, without government notice or public approval, to hunt down the most vulnerable people.”
These are just some of the most public examples of how this technology is being used, but much more is happening behind the scenes. Because of this, there has been an increased call for scrutiny and regulation. Increased transparency, whether willing or through investigative reporting, is bringing many practices to light. Without the public having an understanding of artificial intelligence and what it could mean for the economy, government, law enforcement, surveillance, and every other aspect of society, there is little hope for governments and companies to self-regulate. Society is seeing both the enormous benefits of AI and the massive concerns, with all of it happening too fast for companies, governments, and individuals to keep up.
Google’s CEO Calls For Increased Regulation To Avoid “Negative Consequences of AI”
Last year saw an increasing amount of attention drawn to the regulation of the AI industry, and this year seems to be continuing the trend. Just recently, Sundar Pichai, the CEO of Google and Alphabet Inc., supported the regulation of AI at an economic think tank taking place in Brugel.
Pichai’s comments were likely made in anticipation of new EU plans to regulate AI, which will be revealed in a few weeks. It’s possible that the EU regulations could contain policies legally enforcing certain standards for AI used in transportation, healthcare, and other high-risk sectors. The new EU regulations may also require increased transparency regarding AI systems and platforms.
According to Bloomberg, Google has previously tried to challenge antitrust fines and copyright enforcement in the EU. Despite previous attempts to push back against certain regulatory frameworks in Europe, Pichai stated that regulation is welcome as long as it takes “a proportionate approach, balancing potential harms with social opportunities.”
Pichai recently wrote an opinion piece in Financial Times, where he acknowledged that along with many opportunities to improve society, AI also has the potential to be misused. Pichai stated that regulations should help avoid the “negative consequences of AI”, citing abusive use of facial recognition and deepfakes as negative applications of AI. Pichai stated that international alignment is necessary for regulatory principles to work, and as such, there needs to be agreement on core values. Beyond that, Pichai said that it is the responsibility of AI companies like Google to give consideration to how AI can be used in an ethical manner and that this is why Google adopted its own standards for ethical AI use in 2018.
Pichai stated that government regulatory bodies and policies will play an important role in ensuring AI is used ethically, but that these bodies need not start from scratch. Pichai suggests that regulatory entities can look to previously established regulations for inspiration, such as Europe’s General Data Protection Regulation. Pichai also wrote that ethical AI regulation can potentially be both broad and flexible, with regulation providing general guidance that can be tailored for specific implementations in specific AI sectors. Newer technologies like self-driving vehicles will require new rules and policies that weigh benefits and costs against one another, while for more well-tread ground like medical devices, existing frameworks can be a good starting point.
Finally, Pichai stated that Google wants to partner with regulators to develop policies and find solutions that will balance trade-offs, Pichai wrote in Financial Times:
“We want to be a helpful and engaged partner to regulators as they grapple with the inevitable tensions and trade-offs. We offer our expertise, experience and tools as we navigate these issues together.”
While some have applauded Google for taking a stance on the need for regulation to ensure ethical AI usage, the debate continues over the extent to which it’s appropriate that AI companies should be involved with the creation of regulatory frameworks.
As for the upcoming EU regulations themselves, it’s possible that the EU is pursuing a risk-based rules system, which would put tighter restrictions on high-risk applications of AI. This includes restrictions that could be much tighter than Google hopes for, including a potential multi-year ban on facial recognition technology (with exceptions for research and security). In contrast to the EU’s more restrictive approaches, the US has pushed for relatively light regulations. It remains to be seen how the different regulation strategies will impact AI development, and society at large, in the two different regions of the globe.
U.S. Government Will Limit Exports of Artificial Intelligence
The U.S. government will take steps next week to limit the export of artificial intelligence (AI) software. The decision by the Trump administration comes at a time when powerful rival nations, such as China, are becoming increasingly dominant in the field. The move is meant to keep certain sensitive technologies from falling into the hands of those nations.
The new rule goes into effect on January 6, 2020, and it will be aimed at certain companies that export geospatial imagery software from the United States. Those companies will be required to apply for a license to export it. The only exception is that a license will not be required to export to Canada.
The new measure was the first of its kind to be finalized by the Commerce Department under a mandate from a 2018 law passed by Congress. That law updated arms controls to include emerging technology.
The new rules will likely have an effect on a growing part of the tech industry. These algorithms are currently being used in order to analyze satellite images of crops, trade patterns and other changes within the economy and environment.
Chinese companies are responsible for having exported artificial intelligence surveillance technology to over 60 countries. Some of those countries have dismal human rights records and include Iran, Myanmar, Venezuela, and Zimbabwe.
Within the nation of China itself, the Communist Party is using facial recognition technology systems to target Uighurs and other Muslim minorities located in China’s far western Xinjang region. According to a report released by a U.S. think tank, Beijing has been involved in “authoritarian tech.”
The think tank that released the report was the Carnegie Endowment for International Peace, and they did so after rising concerns of authoritarian regimes using the technology as a way to gain power.
“Technology linked to Chinese companies — particularly Huawei, Hikvision, Dahua and ZTE — supply AI surveillance technology in 63 countries, 36 of which have signed onto China’s Belt and Road Initiative,” the report said.
One of China’s leading technology companies, Huawei Technologies Co., alone provides AI surveillance technology to at least 50 countries.
“Chinese product pitches are often accompanied by soft loans to encourage governments to purchase their equipment,” according to the report. “This raises troubling questions about the extent to which the Chinese government is subsidizing the purchase of advanced repressive technology.”
China has faced increased scrutiny after an investigative report by the International Consortium of Investigative Journalists was released detailing the nation’s surveillance and policing systems, which are being used to oppress Uighurs and send them to internment camps.
The new rules implemented by the U.S. government will at first only go into effect within the country. However, U.S. authorities have said that they could be submitted to international bodies at a later time.
There has been recent bi-partisan frustration over the long amount of time it is taking to roll-out new export controls for the technology.
“While the government believes that it is in the national security interests of the United States to immediately implement these controls, it also wants to provide the interested public with an opportunity to comment on the control of new items,” according to Senate Minority Leader Chuck Schumer.
- Neural Hardware and Image Recognition
- Charles J. Simon, Author, Will Computers Revolt? – Interview Series
- AI Makes it Easier for Drones to Scan and Excavate Terrain
- Noah Schwartz, Co-Founder & CEO of Quorum – Interview Series
- Anton Dolgikh, Head of AI, Healthcare and Life Sciences at DataArt – Interview Series