Amy McGovern is a Lloyd G. and Joyce Austin Presidential Professor in the School of Computer Science and an Adjunct Professor in the School of Meteorology at the University of Oklahoma. She has been leading the development of AI/ML for weather applications for 15 years. As climate change affects weather patterns and sea levels rise, the world's need for accurate, usable predictions of weather and ocean and their impacts has never been greater. At the same time, the quantity and quality of Earth observation and modeling systems are increasing dramatically, offering a deluge of data so rich that only automated intelligent systems can fully exploit it. In this talk, I will discuss our approach to developing trustworthy AI methods for environmental science.
> Science's COVID-19 coverage is supported by the Pulitzer Center. In a normal summer, Appledore Island, a 39-hectare outcrop 12 kilometers off the coast of Maine and New Hampshire, becomes a classroom. Students from high school to graduate level live in close quarters, eat in a communal dining hall, and work shoulder to shoulder to explore the biology of the shore and waters in 18 courses organized by the Shoals Marine Laboratory. But this summer, with the pandemic surging, students have stayed home. Instead, a skeleton staff on Appledore is streaming field trips and dissections of fish and invertebrates and setting up cameras to gather data for students. Rather than leading students around the island, coastal restoration ecologist Gregg Moore from the University of New Hampshire (UNH), Durham, hauls a backpack full of equipment: “a dual modem with two different cellular carriers, a signal-boosting directional antenna, and a large DC power source,” he says. The equipment allows him to teach 12 remote students—twice the course's usual enrollment—basic techniques of coastal ecology. Moore's is just one of hundreds of lab and field courses forced online by COVID-19—“a seismic shift for those who were not already involved in distance or online education,” says Martin Storksdieck, a science education researcher at Oregon State University, Corvallis. Some researchers worry students will miss out on certain practical and problem-solving skills and won't be able to judge whether the hands-on work of a scientist is a good fit for them. But instructors are developing high-tech ways to simulate the field and lab experiences. “I would say [these courses] are not virtual,” says Jennifer Seavey, director of the Shoals lab. “They are real.” And some advantages are emerging. By lowering geographical and financial barriers, Seavey says, “Virtual field courses are democratizing fieldwork.” The shift has taken ingenuity. “Professors must get creative and use a combination of what is available,” including online videos and free or commercially available online labs, says Mildred Pointer, a physiologist at Howard University who is working on a fall course in general biology. No single tool meets all their needs, Pointer says. As the pandemic gained momentum, emails flew among the leaders of the National Association of Geoscience Teachers. Many U.S. geology majors must take a “capstone” field course to graduate. The cancellation of more than three-quarters of these courses jeopardized graduation for many majors. So the association invited instructors to develop learning objectives that did not depend on students doing fieldwork. It also compiled online exercises to help the 29 field courses that have moved online this summer. Lessons range from “Orienteering in Minecraft” to “Geology of Yosemite Valley,” which includes a 43-stop Google Earth tour with photos and embedded text. Like Moore, geoscientist Jim Handschy wanted to give remote students “as close to the real experience as possible.” He runs Indiana University's Judson Mead Geologic Field Station in Montana, which had enrolled 60 students before classes were canceled in March. He and a few instructors visited each outcrop in their course plan, filmed the rocks and landscape, and captured magnified views of samples. Each week, the class delves deeper into the rock layers and their history. For their final project, students digitally map a 3100-hectare landscape. Shannon Dulin, a geologist at the University of Oklahoma, Norman, who just finished teaching a field course, sees the value of learning how to survey a landscape without setting foot on it. On their class evaluations, her students said they gained unexpected skills. “And these are skills they are going to need on the job,” she adds, as geologists are increasingly being asked to evaluate sites they don't visit. In other fields, hands-on learning takes place in labs. Typically, students work in pairs and share equipment, “so there are a lot of issues about virus transmission,” says Heather Lewandowski, a physicist at the University of Colorado (CU), Boulder. At her university this fall, lab exercises as diverse as building an electrical circuit or analyzing solar flare data will most likely be completely remote. Luckily, physics already had a foot in the virtual lab world—especially at CU. There, back in 2002, Nobel laureate Carl Wieman developed the Physics Education Technology (PhET) Interactive Simulations project to provide “games” that teach students basic physics concepts. The PhET web portal now has 106 physics-based simulations and another 50 or so for other disciplines. It became a go-to place this spring for faculty shifting to online teaching; traffic increased fivefold, says Director Katherine Perkins. In addition, several universities have adopted a handheld device called the iOLab that rents for $50 a semester. With it, students can measure magnetism, light intensity, acceleration, temperature, gravity, and atmospheric pressure, and do basic physics experiments at home. “They like that we trust them and are not just giving them instructions,” says iOLab inventor and physicist Mats Selen at the University of Illinois, Urbana-Champaign. Lewandowski and her colleagues surveyed physics instructors and students about their experiences and posted their findings on arXiv, the physics preprint server, on 2 July. Respondents said online labs work best when projects are open-ended, and online class meetings are kept small. They complained about technical difficulties, students having unequal access to the internet and materials, and longer prep times for both students and instructors. But they reported they could meet most key learning objectives, Lewandowski says, even though “there are lots of things we can't replicate in remote experiments,” such as such as building vacuum chambers or troubleshooting equipment. Some institutions decided this spring that virtual just wouldn't do. The Marine Biological Laboratory (MBL) in Woods Hole, Massachusetts, simply canceled its summer courses. “MBL courses are world-renowned for the intensity of the hands-on nature of the lab work,” says Director Nipam Patel. Students spend long hours with famous faculty and do their own projects using organisms collected locally. “We felt that it would be exceedingly difficult to replicate these experiences as a virtual lab course.” Other institutions will try for a mix of in-person and virtual labs. Suely Black, chemistry chair at Norfolk State University, expects only half of his students will be in lab each week this fall, while the other half will be in online classes analyzing data and writing reports. “The crisis has caused us to more critically evaluate what activities students must experience in the lab setting,” he says. Similarly, this fall, organic chemistry students at the University of Michigan (UM), Ann Arbor, will rotate into the lab in small groups, giving each a taste of the hands-on experience. Personal protection equipment is standard for this course and all the work is done in hoods with excellent air exchange, so “they are really fully protected,” says UM biochemist Kathleen Nolta. Storksdieck, an advocate of online learning, questions the value of smelling fumes or using a pipette. “We have to ask whether all the hands-on taught so far was all that great,” he says. Dominique Durand, a biomedical engineer at Case Western Reserve University, says after he put a master's program in biomedical engineering completely online 5 years ago, he concluded that solving problems was more important than hands-on experience. And University of California, Santa Cruz, ecologist Erika Zavaleta thinks virtual courses will open fieldwork to far more students. “There are things you can do online that you can't do in person,” she adds, such as visiting more places than possible by driving. Even so, Handschy laments that his geology students will not have the 12-hour-a-day immersive interactions with each other and faculty that past classes have had. Natalie White, a rising junior at UNH who took Moore's course on Appledore last year, agrees: “You don't have all the time in between when you walk around the island and can ask impromptu questions.” Appledore Island is the source of some her fondest memories. “I think they are missing out on the community.”
Two mayors discussed how they are using artificial intelligence and machine learning to improve their cities and prepare for the workforce of the future at a conference held April 23 in Chicago. The event was hosted by news organization Axios and the United States Conference of Mayors and led by Axios Executive Editor Mike Allen. Also joining the discussion was Imir Arifi, head of artificial intelligence and machine learning at Health Care Service Corporation. According to Arifi, the main use of AI and machine learning is through historical data to predict future events. In a city, for example, Arifi said AI can be used to predict how many potholes the city will need to fill in a year based on data from previous years.
Imitation learning has been commonly applied to solve different tasks in isolation. This usually requires either careful feature engineering, or a significant number of samples. This is far from what we desire: ideally, robots should be able to learn from very few demonstrations of any given task, and instantly generalize to new situations of the same task, without requiring task-specific engineering. In this paper, we propose a meta-learning framework for achieving such capability, which we call one-shot imitation learning. Specifically, we consider the setting where there is a very large (maybe infinite) set of tasks, and each task has many instantiations. For example, a task could be to stack all blocks on a table into a single tower, another task could be to place all blocks on a table into two-block towers, etc. In each case, different instances of the task would consist of different sets of blocks with different initial states. At training time, our algorithm is presented with pairs of demonstrations for a subset of all tasks. A neural net is trained that takes as input one demonstration and the current state (which initially is the initial state of the other demonstration of the pair), and outputs an action with the goal that the resulting sequence of states and actions matches as closely as possible with the second demonstration. At test time, a demonstration of a single instance of a new task is presented, and the neural net is expected to perform well on new instances of this new task. Our experiments show that the use of soft attention allows the model to generalize to conditions and tasks unseen in the training data. We anticipate that by training this model on a much greater variety of tasks and settings, we will obtain a general system that can turn any demonstrations into robust policies that can accomplish an overwhelming variety of tasks.
Summary: I learn best with toy code that I can play with. This tutorial teaches DeepMind's Neural Stack machine via a very simple toy example, a short python implementation. I will also explain my thought process along the way for reading and implementing research papers from scratch, which I hope you will find useful.
Although most of the Kaggle competition winners use stack/ensemble of various models, one particular model that is part of most of the ensembles is some variant of Gradient Boosting (GBM) algorithm. Take for an example the winner of latest Kaggle competition: Michael Jahrer's solution with representation learning in Safe Driver Prediction. His solution was a blend of 6 models. 1 LightGBM (a variant of GBM) and 5 Neural Nets. Although his success is attributed to the new semi-supervised learning that he invented for the structured data, but gradient boosting model has done the useful part too. Even though GBM is being used widely, many practitioners still treat it as complex black-box algorithm and just run the models using pre-built libraries.
Imitation learning has been commonly applied to solve different tasks in isolation. This usually requires either careful feature engineering, or a significant number of samples. This is far from what we desire: ideally, robots should be able to learn from very few demonstrations of any given task, and instantly generalize to new situations of the same task, without requiring task-specific engineering. In this paper, we propose a meta-learning framework for achieving such capability, which we call one-shot imitation learning. Specifically, we consider the setting where there is a very large set of tasks, and each task has many instantiations. For example, a task could be to stack all blocks on a table into a single tower, another task could be to place all blocks on a table into two-block towers, etc. In each case, different instances of the task would consist of different sets of blocks with different initial states. At training time, our algorithm is presented with pairs of demonstrations for a subset of all tasks. A neural net is trained that takes as input one demonstration and the current state (which initially is the initial state of the other demonstration of the pair), and outputs an action with the goal that the resulting sequence of states and actions matches as closely as possible with the second demonstration. At test time, a demonstration of a single instance of a new task is presented, and the neural net is expected to perform well on new instances of this new task. The use of soft attention allows the model to generalize to conditions and tasks unseen in the training data. We anticipate that by training this model on a much greater variety of tasks and settings, we will obtain a general system that can turn any demonstrations into robust policies that can accomplish an overwhelming variety of tasks. Videos available at https://bit.ly/nips2017-oneshot .
At this time, blogging assists students to become more ICT literate which is a crucial 21st century skill. Through blogging, we're able to incidentally discuss many ICT skills such as keyboard shortcuts, Creative Commons, researching online and troubleshooting. Creating a class blog requires teamwork and collaboration. Students and teachers learn and share together. A real sense of classroom community can be developed through blogging and establishing a class identity Also, with this we can create an authentic audience, In the traditional classroom, the only audience of student work was the teacher and sometimes classmates and parents.