The utility of learning a dynamics/world model of the environment in reinforcement learning has been shown in a many ways. When using neural networks, however, these models suffer catastrophic forgetting when learned in a lifelong or continual fashion. Current solutions to the continual learning problem require experience to be segmented and labeled as discrete tasks, however, in continuous experience it is generally unclear what a sufficient segmentation of tasks would be. Here we propose a method to continually learn these internal world models through the interleaving of internally generated rollouts from past experiences (i.e., pseudo-rehearsal). We show this method can sequentially learn unsupervised temporal prediction, without task labels, in a disparate set of Atari games. Empirically, this interleaving of the internally generated rollouts with the external environment's observations leads to an average 4.5x reduction in temporal prediction loss compared to non-interleaved learning. Similarly, we show that the representations of this internal model remain stable across learned environments. Here, an agent trained using an initial version of the internal model can perform equally well when using a subsequent version that has successfully incorporated experience from multiple new environments.

That was among the takeaways from the Tuesday morning keynote sessions here at RSA 2019. "AI is the new foundation for our entire industry, it will enable us to better defend ourselves, to better detect threats, to out-innovate our adversaries, to solve other key issues," said Steve Grobman, CTO of McAfee. "But we have to ask, are we looking at AI objectively? We cannot only focus on the potential, we must also understand the limitations and how it will be used against us." Grobman continued with an example of work McAfee did in taking public safety data sets about crime and with 50 lines of python and machine learning to predict whether a crime would be committed in a specific region of the city based on certain parameters.

Most autonomous vehicle tech ventures such as Waymo, GM Cruise and Nvidia rack up miles of deep learning experience to build reliably safe systems for self-driving cars. Cortica and Renesas Electronics' entirely different approach focuses on helping cars learn on their own. Cortica, an Israeli company with roots in predictive artificial intelligence based on visual perception, is embedding its latest "Autonomous A.I." solution on the Renesas R-Car V3H system-on-chip (SoC) solution for self-driving cars. Referred to by the companies as "unsupervised learning," Cortica's autonomous A.I. enables a vehicle to make predictions based on visual data received from forward-facing cameras. According to Cortica, the system uses "'unsupervised learning' methodology to mimic the way humans experience and incorporate the world around them."

Defense Advanced Research Projects Agency (DARPA) officials will include a panel discussion on ethics and legal issues at the Artificial Intelligence (AI) Colloquium being held March 6-7 in Alexandria, Virginia. "We're looking at the ethical, legal and social implications of our technologies, particularly as they become powerful and democratized in a way," reveals John Everett, deputy director of DARPA's Information Integration Office. Questions abound regarding the ethics and legal implications of AI, such as who is responsible if an self-driving automobile runs over a pedestrian, or whether military weapon systems should have a "human in the loop" controlling unmanned systems to prevent mistakes on the battlefield. Those questions become more acute as AI becomes more prevalent. "A lot of the technology of the 20th century was not widely accessible to people. You have high school students editing genes," Everett notes.

Our flagship product, Viz LVO, leverages advanced deep learning to communicate time-sensitive information about stroke patients straight to a specialist who can intervene and treat. In February 2018, the U.S. Food and Drug Administration (FDA) granted a De Novo clearance for Viz LVO, the first-ever computer-aided triage and notification platform to identify LVO strokes in CTA imaging. Most recently, Viz.ai announced its second FDA clearance for Viz CTP through the 510(k) pathway, offering healthcare providers an important tool for automated cerebral image analysis.

Tom is an analyst at the US Department of Defense (DoD).1 All day long, he and his team collect and process massive amounts of data from a variety of sources--weather data from the National Weather Service, traffic information from the US Department of Transportation, military troop movements, public website comments, and social media posts--to assess potential threats and inform mission planning. While some of the information Tom's group collects is structured and can be categorized easily (such as tropical storms in progress or active military engagements), the vast majority is simply unstructured text, including social media conversations, comments on public websites, and narrative reports filed by field agents. Because the data is unstructured, it's difficult to find patterns and draw meaningful conclusions. Tom and his team spend much of their day poring over paper and digital documents to detect trends, patterns, and activity that could raise red flags. In response to these kinds of challenges, DoD's Defense Advanced Research Projects Agency (DARPA) recently created the Deep Exploration and Filtering of Text (DEFT) program, which uses natural language processing (NLP), a form of artificial intelligence, to automatically extract relevant information and help analysts derive actionable insights from it.2 Across government, whether in defense, transportation, human services, public safety, or health care, agencies struggle with a similar problem--making sense out of huge volumes of unstructured text to inform decisions, improve services, and save lives.

Huawei has launched a special "transparency centre" as it attempts to stop people being afraid of it spying on them. The new building in Brussels is an attempt to win back the government leaders and cyber security experts that are worried about its equipment, which powers much of the data connections of the world. Numerous governments have accused the company of using that infrastructure to intercept communications and send them back to the government in China, where it is based. We'll tell you what's true. You can form your own view.

Earlier this year, the UK Education Secretary called for the IT industry to work with educators to make "smarter use" of technologies, such as artificial intelligence (AI), to cut teachers' burgeoning workloads. Speaking at the 2019 BETT educational IT show in London, Damian Hinds said: "Education is one of the few sectors where technology has been associated with an increase in workload rather than the reverse." But he added that, if used wisely, it could also reduce the amount of time educators had to spend on non-teaching tasks, such as lesson planning, marking and general admin. Hinds cited the instance of Bolton College, which has deployed IBM's Watson AI programme as a virtual clerk called Ada. Ada, which can answer natural language questions, provides about 14,000 students with personalised learning assessments and handles queries about the curriculum and attendance issues, both of which teachers would previously have had to do in their own time.

Andreas Schleicher is a German data scientist--tall and precise with a grey mustache and a steely gaze. The head of the education division at the Organisation for Economic Cooperation and Development (OECD), he gives off an impression of determined focus. That's useful, considering that he's on a mission to change the way countries around the world teach their children. Society, according to Schleicher, is preparing for the future of work all wrong. We're scared that human jobs will be replaced by robots. But we're still teaching kids to think like machines. "What we know is that the kinds of things that are easy to teach, and maybe easy to test, are precisely the kinds of things that are easy to digitize and to automate," Schleicher said at the LearnIt conference in London in January. It's fairly easy to teach and test math, for example--but robots happen to be pretty good at math, too.

Jason Baker was studying political science at UC Davis when he got his hands on "SimCity." He took a careful approach to the computer game. "I was not one of the players who enjoyed Godzilla running through your city and destroying it. I enjoyed making my city run well." This conscientious approach gave him a boost in a class on local government.