Commonsense knowledge acquisition and reasoning have long been a core artificial intelligence problem. However, in the past, there has been a lack of scalable methods to collect commonsense knowledge. In this paper, we propose to develop principles for collecting commonsense knowledge based on selectional preference. We generalize the definition of selectional preference from one-hop linguistic syntactic relations to higher-order relations over linguistic graphs. Unlike previous commonsense knowledge definition (e.g., ConceptNet), the selectional preference (SP) knowledge only relies on statistical distribution over linguistic graphs, which can be efficiently and accurately acquired from the unlabeled corpus with modern tools. Following this principle, we develop a large-scale eventuality (a linguistic term covering activity, state, and event)-based knowledge graph ASER, where each eventuality is represented as a dependency graph, and the relation between them is a discourse relation defined in shallow discourse parsing. The higher-order selectional preference over collected linguistic graphs reflects various kinds of commonsense knowledge. Moreover, motivated by the observation that humans understand events by abstracting the observed events to a higher level and can thus transferring their knowledge to new events, we propose a conceptualization module to significantly boost the coverage of ASER. In total, ASER contains 438 million eventualities and 648 million edges between eventualities. After conceptualization with Probase, a selectional preference based concept-instance relational knowledge base, our concept graph contains 15 million conceptualized eventualities and 224 million edges between them. Detailed analysis is provided to demonstrate its quality. All the collected data, APIs, and tools are available at https://github.com/HKUST-KnowComp/ASER.

For many real-world classification problems, e.g., sentiment classification, most existing machine learning methods are biased towards the majority class when the Imbalance Ratio (IR) is high. To address this problem, we propose a set convolution (SetConv) operation and an episodic training strategy to extract a single representative for each class, so that classifiers can later be trained on a balanced class distribution. We prove that our proposed algorithm is permutation-invariant despite the order of inputs, and experiments on multiple large-scale benchmark text datasets show the superiority of our proposed framework when compared to other SOTA methods.

This article is part of the VB Lab / NVIDIA GTC insight series. "The story of GTC is in many ways the story of NVIDIA, and it's also the story of what's happening in technology," says Greg Estes, VP of corporate marketing and developer programs at NVIDIA. Twelve years ago, GTC began as a conference focused squarely on GPUs, and at that time, that meant primarily graphics and gaming. "But then people figured out that GPUs are the perfect architecture for AI," says Estes. GTC is now billed as the conference for AI innovators, developers, technologists, startups and creatives, and this year it will offer over 1,500 sessions covering breakthroughs in AI, data center, accelerated computing, autonomous vehicles, health care, intelligent networking, game development, and more.

Since 2012, in a case-study in Bucaramanga-Colombia, 179 pedestrians died in car accidents, and another 2873 pedestrians were injured. Each day, at least one passerby is involved in a tragedy. Knowing the causes to decrease accidents is crucial, and using system-dynamics to reproduce the collisions' events is critical to prevent further accidents. This work implements simulations to save lives by reducing the city's accidental rate and suggesting new safety policies to implement. Simulation's inputs are video recordings in some areas of the city. Deep Learning analysis of the images results in the segmentation of the different objects in the scene, and an interaction model identifies the primary reasons which prevail in the pedestrians or vehicles' behaviours. The first and most efficient safety policy to implement - validated by our simulations - would be to build speed bumps in specific places before the crossings reducing the accident rate by 80%.

Abstract: Recent work has explored how complementary strengths of humans and artificial intelligence (AI) systems might be productively combined. However, successful forms of human-AI partnership have rarely been demonstrated in real-world settings. We present the iterative design and evaluation of Lumilo, smart glasses that help teachers help their students in AI-supported classrooms by presenting real-time analytics about students' learning, metacognition, and behavior. Results from a field study conducted in K-12 classrooms indicate that students learn more when teachers and AI tutors work together during class. We discuss implications for the design of human-AI partnerships, arguing for participatory approaches to research in this area, and for principled approaches to studying human-AI decision-making in real-world contexts. Artificial intelligence (AI) systems are increasingly used to support human work in deeply social contexts such as education, healthcare, social work, and criminal justice. In these contexts, AI can automate routine parts of practitioners' work, while freeing up their time for activities they find more meaningful (17, 28, 39).

Most of state of the art methods applied on time series consist of deep learning methods that are too complex to be interpreted. This lack of interpretability is a major drawback, as several applications in the real world are critical tasks, such as the medical field or the autonomous driving field. The explainability of models applied on time series has not gather much attention compared to the computer vision or the natural language processing fields. In this paper, we present an overview of existing explainable AI (XAI) methods applied on time series and illustrate the type of explanations they produce. We also provide a reflection on the impact of these explanation methods to provide confidence and trust in the AI systems.

SCI COMMUN### Astrophysics The team that in 2019 used a global network of radio telescopes to reveal the first image of a black hole has offered a new twist on that iconic view: the same black hole in polarized light. The thin lines spiraling in toward the black hole's shadow (above) show areas of light that differ in their polarization—the direction in which the light waves vibrate. The light, from plasma near the black hole's edge, was polarized by magnetic fields, and so the new image, described last week in The Astrophysical Journal by the Event Horizon Telescope team, indicates their structure. Researchers hope to learn how the fields help accreting black holes funnel matter and energy into jets emanating from their poles. 69% —Percentage of postdoctoral researchers surveyed in October 2020 by the U.S. National Institutes of Health who anticipate the COVID-19 pandemic will negatively affect their careers. For researchers at all levels, the figure was 55%. ### Conservation Despite the antienvironmental policies of its current leadership, Brazil has become the 130th country to ratify the Nagoya Protocol, a part of the Convention on Biological Diversity that lays out measures to protect countries' biodiversity claims, the CBD announced last week. The ratification, first proposed by a previous administration in 2012, had languished until 2019, when rampant deforestation led pro-environment leaders to push for approval. The current government is seen as having consented because the protocol allows nations to impose rules on the international trade in its plant and animal products; by legitimizing the sales, the regulations are expected to increase exports and tax revenues. For example, money from sales of native plants such as açai ( Euterpe oleracea ) and Brazil nut ( Bertholletia excelsa ) could be returned to help Indigenous communities that use and harvest them. Observers question whether the ratification alone will protect Brazil's biodiversity, perhaps the world's greatest—but hailed the step as helpful. ### Public health The United States and 13 other countries this week criticized a report by a World Health Organization panel that had visited China to investigate how the COVID-19 pandemic started. The 300-page document says the most likely cause was a bat coronavirus that infected another, unidentified animal and then moved to humans, but it recommends further research. The report's most definitive conclusion is also its most controversial: that it is “extremely unlikely” that SARS-CoV-2 came out of a Chinese laboratory. Scientists from China made up half of the 34-member international panel. A joint statement by other countries complained that the investigation was “significantly delayed and lacked access to complete, original data, and samples.” It called for a transparent, “rapid, independent, expert-led, and unimpeded evaluation of the origins.” ### Funding The science committee in the U.S. House of Representatives wants to more than double the budget of the National Science Foundation (NSF) in the next 5 years, from $8.5 billion to $18.3 billion. A sizable chunk of the extra money—$5 billion by 2026—would go to a new directorate, Science and Engineering Solutions, that would accelerate the conversion of basic research into new technologies and products. Last year, Senate Majority Leader Chuck Schumer (D–NY) proposed growing NSF to $100 billion over 5 years, with roughly one-third of that money going to a new technology directorate. Schumer's vision for NSF is part of still-evolving draft legislation affecting many federal agencies that pinpoints key technologies needed to address economic and security threats posed by China's growing technological prowess. In contrast, the House bill is limited to NSF's programs and is aimed at strengthening basic research across all disciplines that NSF supports. The House and Senate would need to agree on a vision for NSF, and other legislation would be needed to appropriate the money. ### Astronomy Light pollution from space junk and satellites may have already robbed the entire Earth of the dark skies best for sensitive astronomical observations, an analysis has found. Researchers estimated the size and shininess of tens of thousands of objects in orbit as of 2020, before an onslaught of thousands more satellites that companies plan to launch in the coming years. Even at Earth's darkest sites, the sky glows from natural sources such as ionized particles; but the existing orbiting objects reflect and scatter about 10% more of this diffuse light back into the atmosphere, the research team calculates in a paper accepted this week by the Monthly Notices of the Royal Astronomical Society . That extra amount violates an International Astronomical Union standard for observing sites and could compromise observations of the dimmest galaxies, which scientists study for clues about the physics of galaxy formation and the nature of dark matter. To gather such data, astronomers already need long exposures on the biggest telescopes at the darkest available sites. ### Ethics Harvard University last week penalized quantitative biologist Martin Nowak for his connections with disgraced financier Jeffrey Epstein. Epstein had donated $6.5 million for Nowak's research in 2003; after being convicted in 2008 of soliciting prostitution from a minor, Epstein introduced Nowak to donors who provided an additional $7.5 million. Nowak's actions after 2008—repeatedly hosting Epstein on campus, promoting Epstein on his program's web page, and providing false information about Epstein's support in a grant application—violated Harvard policies, and other actions showed “blameworthy negligence and unprofessional behavior,” Claudine Gay, dean of arts and sciences, wrote in an email last week to faculty members. Nowak will continue at Harvard as a math professor, but his Program for Evolutionary Dynamics will be shut down and he will be barred for at least 2 years from serving as a principal investigator on grants. “I regret the connection I was part of fostering between Harvard and Jeffrey Epstein,” Nowak said in a statement last week. Epstein died by suicide in 2019. ### Archaeology Chinese archaeologists last week reported unearthing more than 500 artifacts, including gold ornaments, bronze heads, ivory and jade tools, and a gold mask dating back about 3000 years at the Sanxingdui archaeological site in southwestern Sichuan province. Sanxingdui, then ruled by the Shu kingdom, has already yielded thousands of bronze relics unlike anything found elsewhere in China, including at sites of the contemporaneous Shang dynasty in the Yellow River region. The new finds, retrieved from what are thought to be sacrificial pits, may shed light on how the Shu kingdom contributed to Chinese civilization. VACCINE LEADER FIRED Moncef Slaoui, who headed COVID-19 vaccine development during the Trump administration, has been fired as chairman of a medical research firm controlled by manufacturer GlaxoSmithKline after he was accused of sexual harassment. The company said an outside investigation substantiated the allegation by a female employee about Slaoui's behavior several years ago when he worked there. Slaoui also stepped down from leadership roles at two other pharmaceutical companies and issued a statement in which he apologized to the woman and his family. RETURNING LOOTED ART Museums in Germany have pledged to return hundreds of artifacts, including bronze statues, looted during the colonial era from the kingdom of Benin in what is now Nigeria. The British Museum and others face growing pressure to join them. PARDON SOUGHT The Australian Academy of Science issued a statement saying a court ignored new genetic evidence when it denied last week an appeal by a woman convicted of killing her four young children. Tests point to a natural cause of the deaths: Two of the children carried a mutation in the CALM2 gene that is associated with sudden death by cardiac failure in infants and children. Prosecutors had accused Kathleen Folbigg of smothering the children but have not presented medical evidence that supports that position. Academy members have signed a petition asking New South Wales's governor to pardon her. AI IN MEDICINE The Broad Institute has received $300 million to study how machine learning can improve the prevention and treatment of disease. Half the sum is coming from a foundation of Wendy and Eric Schmidt, a member of Broad's board and former CEO of Google, and the rest from the Broad Foundation. R&D SPENDING RISE The United States spent more than 3% of gross domestic product on R&D in 2019 for the first time. The 3.07% share is a record and met a goal set by former President Barack Obama a decade ago. Israel led globally with 4.9%, the Organisation for Economic Co-operation and Development said. Total U.S. spending was more than any other country's.

BERLIN – A turning point for Rafael Yuste, a neuroscientist at New York's Columbia University, came when his lab discovered it could activate a few neurons in a mouse's visual cortex and make it hallucinate. The mouse had been trained to lick at a water spout every time it saw two vertical bars, and researchers were able to prompt it to drink even with no bars in sight, said Yuste, whose team published a study on the experiment in 2019. "We could make the animal see something it didn't see, as if it were a puppet," he said in a phone interview. "If we can do this today with an animal, we can do it tomorrow with a human for sure." Yuste is part of a group of scientists and lawmakers, stretching from Switzerland to Chile, who are working to rein in the potential abuses of neuroscience by companies from tech giants to wearable startups.

Out of a hundred trials, how many errors does your speaker verifier make? For the user this is an important, practical question, but researchers and vendors typically sidestep it and supply instead the conditional error-rates that are given by the ROC/DET curve. We posit that the user's question is answered by the Bayes error-rate. We present a tutorial to show how to compute the error-rate that results when making Bayes decisions with calibrated likelihood ratios, supplied by the verifier, and an hypothesis prior, supplied by the user. For perfect calibration, the Bayes error-rate is upper bounded by min(EER,P,1-P), where EER is the equal-error-rate and P, 1-P are the prior probabilities of the competing hypotheses. The EER represents the accuracy of the verifier, while min(P,1-P) represents the hardness of the classification problem. We further show how the Bayes error-rate can be computed also for non-perfect calibration and how to generalize from error-rate to expected cost. We offer some criticism of decisions made by direct score thresholding. Finally, we demonstrate by analyzing error-rates of the recently published DCA-PLDA speaker verifier.

The increasing availability of large-scale remote sensing labeled data has prompted researchers to develop increasingly precise and accurate data-driven models for land cover and crop classification (LC&CC). Moreover, with the introduction of self-attention and introspection mechanisms, deep learning approaches have shown promising results in processing long temporal sequences in the multi-spectral domain with a contained computational request. Nevertheless, most practical applications cannot rely on labeled data, and in the field, surveys are a time consuming solution that poses strict limitations to the number of collected samples. Moreover, atmospheric conditions and specific geographical region characteristics constitute a relevant domain gap that does not allow direct applicability of a trained model on the available dataset to the area of interest. In this paper, we investigate adversarial training of deep neural networks to bridge the domain discrepancy between distinct geographical zones. In particular, we perform a thorough analysis of domain adaptation applied to challenging multi-spectral, multi-temporal data, accurately highlighting the advantages of adapting state-of-the-art self-attention based models for LC&CC to different target zones where labeled data are not available. Extensive experimentation demonstrated significant performance and generalization gain in applying domain-adversarial training to source and target regions with marked dissimilarities between the distribution of extracted features.