This work studies the (non)robustness of two-layer neural networks in various high-dimensional linearized regimes. We establish fundamental trade-offs between memorization and robustness, as measured by the Sobolev-seminorm of the model w.r.t the data distribution, i.e the square root of the average squared $L_2$-norm of the gradients of the model w.r.t the its input. More precisely, if $n$ is the number of training examples, $d$ is the input dimension, and $k$ is the number of hidden neurons in a two-layer neural network, we prove for a large class of activation functions that, if the model memorizes even a fraction of the training, then its Sobolev-seminorm is lower-bounded by (i) $\sqrt{n}$ in case of infinite-width random features (RF) or neural tangent kernel (NTK) with $d \gtrsim n$; (ii) $\sqrt{n}$ in case of finite-width RF with proportionate scaling of $d$ and $k$; and (iii) $\sqrt{n/k}$ in case of finite-width NTK with proportionate scaling of $d$ and $k$. Moreover, all of these lower-bounds are tight: they are attained by the min-norm / least-squares interpolator (when $n$, $d$, and $k$ are in the appropriate interpolating regime). All our results hold as soon as data is log-concave isotropic, and there is label-noise, i.e the target variable is not a deterministic function of the data / features. We empirically validate our theoretical results with experiments. Accidentally, these experiments also reveal for the first time, (iv) a multiple-descent phenomenon in the robustness of the min-norm interpolator.

Suzanne is a global expert and proven strategic marketing and communications lead on strategic initiatives for corporations, start-ups, non-profits, and the community. Suzanne's most recent project is Mindful Mornings Miami, the hottest new one-hour talk show on independent JoltRadio. With a reach of over 200,000, the show focuses on what it means to wake up and live in the world today. We have the pleasure of welcoming Suzanne Jewell to our interview series, I am Aishwarya Jain from the peopleHum team before we begin just a quick introduction of peopleHum, peopleHum is an end-to-end, one-view, integrated Human Capital Management automation platform, the winner of the 2019 global Codie Award for HCM that is specifically built for crafted employee experiences and the future of work with AI and automation technologies. We run the peopleHum blog and video channel which receives upwards of 200,000 visitors a year and publish around 2 interviews with well-known names globally, every month. We're thrilled to have you on our series. Awesome to be here and welcome to everyone, and my first wish is that everyone is awake, aware and well, today! Thank you so much for that. You know, my experience of living in one place and working on pretty much every other continent on the planet really made me aware that I was always in, kind of autopilot. The media tends to do that, both in the way that it bombards us. And I find that I personally was really being the same way, and so, in that regard for me, mindfulness became a way to really get present to what was going on wherever I happen to be on the planet for my GPS location, but also where my head and my mind and my body is located. The most recent Harvard study actually shows that 47% of the time that we're out of bed and awakened vertical, our mind and our body are actually not in the same place.

A journalist probes the tech companies racing to entice consumers—and investors—with futuristic foods. An outsider documents his ascent in academia. A policy expert proposes a human-centered approach to solving society's problems. From an ode to azure to a deep dive into data, this year's summer reading picks—reviewed by alumni of the AAAS Mass Media Science & Engineering Fellows program—offer readers fresh perspectives on timely scientific topics. Confront the biases that have long imperiled women's health, probe the mysteries of memory, celebrate a prescient economist, and more, with the books reviewed below. —Valerie Thompson Reviewed by Ming Ivory 1 Governance professor Beth Simone Noveck, who formerly served as the first White House deputy chief technology officer, believes that “public entrepreneurship” can counter the failures that have dominated public policy design in the United States since the 1960s. Her new book, Solving Public Problems , revisits the four stages of policy design—identifying problems, identifying solutions, designing for implementation, and evaluation and evolution—while identifying 20 crucial decisions that prioritize “human-centered public policies.” Experts often expend much effort on program design, but once these programs are created, there is usually little fine-tuning of the implementation and hardly any emphasis on measuring whether the desired outcomes are achieved. The US federal civil service, for example, first celebrated as a defense of the “public interest” for its structural insulation from shortsighted patronage and political corruption, has recently come to be viewed by some as a nonelected “deep state” that frustrates legitimate partisan power and private-sector freedom. Noveck fearlessly defends the existence of “public interests,” arguing that their complexity and ethical significance are distinct from academic theory, electoral politics, and private-sector capitalism. Noveck describes governance ideas that expand public policy designs in a variety of sectors, such as health care, transportation, housing, employment, justice, information, and education. She emphasizes participatory elements of the process that ensure that the communities most in need and those who will be most directly affected by proposed policies are consulted throughout the process, and she encourages training in quantitative and qualitative scientific techniques, such as data analysis, research design, artificial intelligence, survey construction, interviewing, crowdsourcing, budgeting, and program evaluation. She discusses what can be learned from administrative records and from the temporary suspension of regulations that encourage private-sector experimentation. Each chapter ends with exercises that, if conscientiously followed, could launch a community and its institutional partners on the path to practical policies that combat problems such as unemployment, information deficits, and housing discrimination. These exercises include checklists of tasks that must be accomplished to achieve the desired outcome, and they direct readers to an index full of organizations that could be sources of assistance. If there is a weakness to the guide, it is that there are an almost overwhelming number of examples, replete with management jargon, that must be waded through without a lot of information on their relative quality. A list synthesizing what Noveck considers the best of these programs—organized either by agency type or by the skill set supported—would have been useful. Overall, however, the book offers a wealth of information necessary to improve human-centered design in public policies. Reviewed by Anna Funk 2 The latest food tech to hit the mainstream may be plant-based burgers, but countless start-ups and research labs are gearing up to transform the way humans think about what we eat—or at least, so they hope. From giant tanks of protein-rich algae to petri dishes culturing animal cells, researchers are seeking ways to give consumers—and investors—products that will improve food's sustainability, healthiness, or preferably both (bonus points if they are accompanied by new, patentable technologies that will keep the cash coming). Some groups are trying to turn plants into meat, while others are trying to turn meat into more meat without killing more animals. Others still are trying to revolutionize parts of the production process, building artificial intelligence–laden greenhouses in cities, for example, or salvaging food waste and turning it into more food. In Technically Food , journalist Larissa Zimberoff explores eight of the latest tech trends in the food sector, giving readers an inside look at the progress that has been made in each, a thoughtful look at current shortcomings, and, whenever possible, a taste test. Zimberoff walks readers through the latest breakthroughs from groups working to turn algae, fungi, or peas into protein sources; the worlds of upcycling and vertical farming; laboratories culturing cell-based meat; and more-mainstream staples such as nondairy milks, nonchicken eggs, and plant-based burgers. The book wraps up with a surprisingly delightful medley of commentary from 19 experts on what they think will be on our plates in 20 years. (My personal favorite was from author and animal rights activist Paul Shapiro, who asks: What if local establishments could brew their own meat on-site like they would a craft IPA?) Unfortunately, interspersed throughout Zimberoff 's otherwise detailed reporting were more than a few technical flubs—mostly harmless in nature but certainly distracting to a careful reader. She mentions, for example, that ocean acidification occurs when pH levels rise (it is the opposite), references COVID-19 when she means SARS-CoV-2, and refers to yeast as bacteria. I also could have done without the occasional implication that science is boring or hard to understand (“Have your eyes glazed over yet?”) and her take on Expo West, a huge natural products show, where she tasted and spat out the free food samples. (The displays, she writes, were “enough to torture anyone's waistline.”) Still, the reporting behind this book is masterful. I was constantly pulled along by ideas about the food system that I had never considered, from secondary plant compounds that might be beneficial to human health—and are only produced if you lay off the pesticides and let a plant get nibbled a little—to what the median age of the United States's traditional farmers (57.5 years in 2017) portends about the future of farming. Even the title proved to be a wink I did not expect—not just “technically” as in technical, technological, but also “technically” as in “well, technically, it's food.” The overarching question of whether high-tech food is actually an improvement or not is not answered by Zimberoff, but she leaves readers with plenty of food for thought. Reviewed by Stephani Sutherland 3 “The history of mankind is a history of repeated injuries and usurpations on the part of man toward woman,” declared abolitionist and suffragist Elizabeth Cady Stanton in 1848 at a convention to address inequities faced by women. Although the sentiment may sound radical to some, cultural historian Elinor Cleghorn's new book suggests that Stanton's argument is not far off track. In Unwell Women , Cleghorn provides an extensive history of how feminine anatomy, physiology, and psychology have been studied and manipulated—mainly by men—and how they have often been used to oppress the female sex. The book is populated by meticulously researched and quoted historical figures—some famous, others simply captured in quotidian documentation. The cast of characters includes girls and women who suffered at the hands of men, as well as men who shaped history for pioneering medical techniques and theories under the frequently false guise of protecting, curing, and acting in women's best interests. Also present are the women who have been driving forces for change, pushing for the right to hold authority over one's own body and life. Unwell Women details a history in which women were tortured, burned, and hanged for “witchcraft”; enslaved for the purpose of gynecological experimentation; and clitoridectomized for the crime of masturbation. During Victorian times, we learn, women of certain social status were often prescribed a forced “rest cure” for hysteria, which entailed utter isolation and inactivity (with the exception of tooth cleaning) and a diet of four quarts of milk per day and raw beef soup. More recently, suffragists were physically assaulted, imprisoned, and force-fed, and many women have been sterilized on the grounds of “feeblemindedness” and “social inadequacy,” often with racial undercurrents. Despite the tremendous recent gains made in the rights of women—to vote, to work, to be educated, to control various facets of one's own life—still, the inequities are massive. Nowhere is that gap more evident and more harshly felt than in the medical realm, where, to this day, women are disbelieved, dismissed, and gaslighted by medical professionals, particularly when their conditions prove difficult to diagnose. One glaring example is the mysterious condition called myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS). When ME/CFS was “discovered” in the 1980s, it was largely dismissed as a psychosomatic illness in wealthy white women who were perhaps “bored” with their lives—a sentiment not far from those used to explain “hysteria” in years past. Another recent example is the alleged hysterectomies being performed without consent on women at a US Immigration and Customs Enforcement facility in 2020 in Georgia. Cleghorn brings her message home in the final chapter, aptly titled “Believe us.” She points out that the only path to change is for the medical profession to reckon with how it has used medicalization to control women for centuries and indeed still does. Although Unwell Women may not become required reading for medical professionals and students, Cleghorn's final message should be heard loud and clear: Believe women. Reviewed by Tamar L. Goulet 4 COVID-19 vaccines are a limited resource. Will some countries hoard them, exhibiting the selfish behavior that ecologist Garrett Hardin anticipated in his 1968 essay “The tragedy of the commons” ([ 1 ][1]), or will they share their extra doses with others who need them? The answer is currently unfolding, but promising signs suggest the latter, confirming predictions articulated by economist Elinor “Lin” Ostrom in her landmark 1990 book, Governing the Commons ([ 2 ][2]). Ostrom, who is profiled in Erik Nordman's new book, The Uncommon Knowledge of Elinor Ostrom , was the first woman to win the Nobel Prize in Economics, securing the award in 2009 for her work on governing the commons and, according to Nordman, for her integration of theoretical foundations with fieldwork, anchoring economic concepts in real-world research. At first, the book's chapters do not seem to be connected. Studies about volunteerism are sandwiched between chapters on global climate change and space, and Ostrom's “eight design principles for managing a commons” appear for the first time in chapter four. Several pages into each chapter, however, Nordman presents the links that tie these themes together: Although resource use is key to Ostrom's principles, resources are not really the foci. Rather, managing “any common-pool resource,” as Nordman states, “is really about managing people.” Ostrom studied a myriad of “commons,” including one that is currently in the societal spotlight: the organization and efficiency of police departments, which are often run through a centralized bureaucracy. Communities with smaller-sized local police departments tend to exert control by engaging in more formal and informal communication with the police, she and her colleagues found, thereby building trust between the two entities. Police forces in larger cities that employ a mixture of centralized and decentralized components may ultimately have better outcomes, they concluded. As Nordman summarizes, “Public safety is a service that is coproduced by police departments and citizens.” Ostrom's observations from the 1970s are worth revisiting half a century later as the United States questions police structure and the role of communities. The work of managing commons reveals much about the complexities of life, illustrating what happens when these entities fail to fit into neat bins, when consequences are not binary, when outcomes seemingly defy logic, and when interconnectedness is key and collaboration is necessary and a strength. As Nordman asserts, “Ostrom left us with the tools to address these global challenges, but the work is up to us.” Reviewed by Max Kozlov 5 Every cat GIF shared on social media, credit card swiped, video watched on a streaming platform, and website visited add more data to the mind-bending 2.5 quintillion bytes of information that humans produce every single day. All of that information has a cost: Data centers alone consume about 47 billion watts, equivalent to the resting metabolism of more than a tenth of all the humans on the planet. In The Ascent of Information: Books, Bits, Genes, Machines, and Life's Unending Algorithm , astrobiologist Caleb Scharf probes this deluge of data, which he terms the “dataome,” to examine how it is changing us just as quickly as we are changing it. Masterfully weaving together anecdotes and thought experiments from neuroscience, evolutionary biology, theoretical physics, astrobiology, and information theory, Scharf investigates how our relationship with the dataome has fundamentally altered our lives and how it will continue to do so. Scharf begins by invoking William Shakespeare, whose legacy permeates the public consciousness more than four centuries after his death, to show just how powerful the dataome can be. On the basis of the average physical weight of one of his plays, “it is possible that altogether the simple act of human arms raising and lowering copies of Shakespeare's writings has expended over 4 trillion joules of energy,” he writes. These calculations do not even account for the energy expended as the neurons in our brains fire to make sense of the Bard's language. Zooming out as the book progresses, Scharf weaves in his own area of expertise—exoplanets—to dissect the argument of whether life exists beyond the confines of our planet. As a result of the same thermodynamic imperatives that gave rise to living systems on this planet, other dataomes, he maintains, are all but an inevitability. The dataome has been around since long before us, and it will persist long after we are gone, Scharf writes, tracing the flow of information and energy back to the birth of the Universe. He compares the rise of information to the rise of oxygen on Earth; both, he argues, involve the reconfiguration of matter and energy flow in very specific ways. Unlike atmospheric oxygen, however, humans have contributed to growing the dataome at unparalleled rates. Some estimate that by 2040, the world's computer chips will demand more electricity than is expected to be produced globally. Scharf ends with a sharply worded warning: All of these data represent a vastly different reality than anything biology has equipped us to deal with. How then, he asks, can we simultaneously preserve and support both our dataome and our planet? We must treat information as a natural resource, Scharf argues, one that cannot be extracted, refined, or used without cost or repercussions. “Information really isn't ‘free,’ nor has it ever been so,” he concludes. Reviewed by Elizabeth Gamillo 6 A Quantum Life: My Unlikely Journey from the Street to the Stars is the autobiography of astrophysicist and science communicator Hakeem Oluseyi, told as a journey from the author's challenging youth to the beginning of his career as a scientist. The book is split into four sections, each documenting pivotal parts of his life, including vivid accounts of his early childhood in an unstable home and his struggles with racism and classism as a Black doctoral student at Stanford University. Oluseyi, born James Plummer Jr., draws readers in with his candid and personable writing style. As I read, it felt as if he was sitting in front of me telling me his life story. Oluseyi sought refuge in books early on in his life and recounts the transformative moment he learned about Albert Einstein and his theory of relativity as a child. “[R]ight away I felt connected to Einstein,” he writes. “I could tell from his photo with that wild hair that he was weird, like me.” As the chapters progress, Oluseyi describes his difficult journey to find acceptance; he is too academically inclined for most of his peers but has trouble fitting in easily in academia. As a graduate student, he is pressured to leave Stanford University after failing his qualifying exam and resents the “privileged snobs” who make up most of the student body, with whom he finds little common ground. Oluseyi's supervisor, solar physicist Art Walker—the only Black faculty member in the physics department—plays a critical role in helping Oluseyi find his footing, highlighting the critical role of scientific mentorship. “Congratulations, Doctor,” Walker tells his young mentee in the book's closing pages, acknowledging Oluseyi's successful dissertation defense. “Art's handshake, and the hug that followed, was all the affirmation I could ask for,” writes Oluseyi. In the book's epilogue, Oluseyi describes his efforts to inspire the next generation of research physicists and details the importance of having culturally relevant role models. Here, he reflects on his experience tutoring Black and Latino high school students in the US and the mentoring program he created for Black astronomy students in South Africa, revealing how he uses his own struggles to relate to and motivate them. I found Oluseyi's perseverance inspiring. His story serves as a reminder that barriers are meant to be broken and that there is no one right way to be a scientist. We need more such stories if we truly wish to increase diversity within the scientific enterprise. Reviewed by Daniel Ackerman 7 Pope Julius II spared no expense in commissioning the Sistine Chapel's ceiling frescoes. He fetched Italy's top artistic talent—Michelangelo—for the job and demanded that the artist render the sky a brilliant celestial blue with a particular pigment: ultramarine. Derived from lapis lazuli stone quarried in present-day Afghanistan, ultramarine was worth more than its weight in gold. Science journalist Kai Kupferschmidt shares Julius II's obsession with blue, and he indulges it in an ambitious new biography of the color. Blue: In Search of Nature's Rarest Color dutifully answers all the expected scientific questions—why is the sky blue?—but it shines brightest when Kupferschmidt blends the physical and the philosophical, asking, for example, is the sky a social phenomenon? Over 200 pages, Kupferschmidt sketches a comprehensive history of the color blue. He deftly bridges mineralogy, botany, and art history to explore humanity's quest for the perfect blue pigment. With equal ease, he describes Picasso's Blue Period palette and the microstructures that blue jays use to “cheat” their way to a dazzling cerulean. The book's most fascinating chapters, “Seeing” and “Speaking,” dwell on how we perceive and communicate color. “Blue light is not actually blue,” writes Kupferschmidt. Light is merely electromagnetic radiation—photons with particular wavelengths. It becomes “blue” only through a dance with the eye, the brain, and our shared understanding of the world. Here, readers learn about the evolution of the eye and follow along as Kupferschmidt ponders whether the ancient Greek poet Homer, who described both the ocean and oxen as “wine-dark,” might have perceived the color blue differently than we do today. Language structures our view of the colorful world, notes Kupferschmidt in this section, revealing that Russian speakers are faster than English speakers in distinguishing shades of blue. (The language splits light and dark blue into different categories, just as English separates green from yellow.) In his effort to see blue from every possible angle, Kupferschmidt's narrative thread occasionally frays—some sections read more like a collection of essays than a unified whole. Yet his lively writing and ability to wrangle disparate disciplines are more than enough to keep the curious reader aboard. And like the very best science writers, Kupferschmidt paints a radical vision of material that would feel mundane in the hands of a less-capable author. Reviewed by Barbara Gastel 8 Mysterious illnesses can serve as starting points for both medical science and popular science writing. They can lead physicians and scientists to identify previously unknown syndromes, better understand the body's functioning, and ultimately improve the prevention and treatment of diseases. For science writers, such cases supply scaffolding for narrative, allow easy integration of human interest, and offer chances to portray not only the products but also the process of science. Such cases are at the core of The Memory Thief and the Secrets Behind How We Remember by science journalist Lauren Aguirre. Early in the book, a young neurologist named Jed Barash views an MRI scan of the brain of a patient acting strangely after a drug overdose. Barash is taken aback: The patient's hippocampus—crucial to memory—is severely damaged, but the rest of his brain is intact. Upon examination, the patient shows profound memory difficulty, akin to the deficits seen in patients with Alzheimer's disease. Barash embarks on a search for other such cases, leading to the identification of what is now called opioid-associated amnestic syndrome. Along the way, he enlists other physicians and researchers to try to gain a sense of how common this syndrome might be, how it arises, what it might imply more broadly about the effects of opiate use, and whether it might offer insights into other memory impairments. Threaded throughout this narrative are accounts of well-known cases in which surgical injury or viral infection ravaged an individual's hippocampus, resulting in permanent memory impairment; descriptions of rodent studies that have helped researchers identify the roles of hippocampal neurons in memory formation; and more information about the effects of opioids on memory. Aguirre discusses the possible origins of Alzheimer's disease as well as factors that contribute to healthy aging of the human brain. Aguirre also recounts the story of Owen Rivers, a bright young man who has been all but unable to form new memories since overdosing on fentanyl in 2018. The book's prologue tells Rivers's story from shortly before to shortly after the overdose, and segments interspersed throughout the main text trace his history and follow his experiences and reflections since the incident. The epilogue includes an engaging essay in which Rivers presents his own perspective on his memory loss, offering readers a firsthand account of the experience. “Without Calendar notifications, task organization apps (huge shoutout to Trello), alarms, and meticulous preplanning each day, navigating everyday life on my own would be unfeasible,” he writes. The Memory Thief is extensively researched, and Aguirre writes clearly, concisely, and often cinematically. Some of the book's denser sections might bog down nonscientists, while experts might lose patience with some of the more informal storytelling. However, the book ultimately succeeds in providing an accessible yet substantive look at memory science and offering glimpses of the often-challenging process of biomedical investigation. 1. [↵][3]1. G. Hardin , Science 162, 1243 (1968). [OpenUrl][4][Abstract/FREE Full Text][5] 2. [↵][6]1. E. Ostrom , Governing the Commons: The Evolution of Institutions for Collective Action (Cambridge Univ. Press, 1990). [1]: #ref-1 [2]: #ref-2 [3]: #xref-ref-1-1 "View reference 1 in text" [4]: {openurl}?query=rft.jtitle%253DScience%26rft.stitle%253DScience%26rft.aulast%253DHardin%26rft.auinit1%253DG.%26rft.volume%253D162%26rft.issue%253D3859%26rft.spage%253D1243%26rft.epage%253D1248%26rft.atitle%253DThe%2BTragedy%2Bof%2Bthe%2BCommons%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.162.3859.1243%26rft_id%253Dinfo%253Apmid%252F5699198%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [5]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEzOiIxNjIvMzg1OS8xMjQzIjtzOjQ6ImF0b20iO3M6MjM6Ii9zY2kvMzcyLzY1NDYvMTAzMC5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30= [6]: #xref-ref-2-1 "View reference 2 in text"

A moment later, a missile fired from a drone slams into the vehicle, exploding in an orange flash, blowing the man off his feet and leaving the tank a smoldering wreck. The scene is one of dozens of aerial videos that were posted online in Azerbaijan last year showing off a new weapon. Over six weeks, it helped the nation regain territory in the Nagorno-Karabakh region that had been held by Russian-backed Armenian forces for more than two decades. The videos show attacks on tanks, trucks, command posts, mortar positions and radar installations. Smaller militaries around the world are deploying inexpensive missile-equipped drones against armored enemies, a new battlefield tactic that proved successful last year in regional conflicts, shifting the strategic balance around Turkey and Russia.

The Institute of Information Technology Professionals South Africa (IITPSA) has welcomed the plans by the South African Revenue Service to use big data analytics and artificial intelligence to improve revenue collection and locate tax evaders. IITPSA Board member Moira de Roche says the move is welcome, if overdue. "It should have been done years ago. The more efficiently SARS collects tax that is due, the better off we will be as a country. It has been relatively easy for people to dodge tax before now, so it is a move in the right direction to start using all the data that exists across various government departments and deploying AI to identify trends and potential fraud."

The European Commission recently published a proposal for a regulation on artificial intelligence (AI). This is the first document of its kind to attempt to tame the multi-tentacled beast that is artificial intelligence. "The sun is starting to set on the Wild West days of artificial intelligence," writes Jeremy Kahn. He may have a point. When this regulation comes into effect, it will change the way that we conduct AI research and development.

In this research paper, I will elaborate on a method to evaluate machine translation models based on their performance on underlying syntactical phenomena between English and Arabic languages. This method is especially important as such "neural" and "machine learning" are hard to fine-tune and change. Thus, finding a way to evaluate them easily and diversely would greatly help the task of bettering them.

We introduce RL-DARTS, one of the first applications of Differentiable Architecture Search (DARTS) in reinforcement learning (RL) to search for convolutional cells, applied to the Procgen benchmark. We outline the initial difficulties of applying neural architecture search techniques in RL, and demonstrate that by simply replacing the image encoder with a DARTS supernet, our search method is sample-efficient, requires minimal extra compute resources, and is also compatible with off-policy and on-policy RL algorithms, needing only minor changes in preexisting code. Surprisingly, we find that the supernet can be used as an actor for inference to generate replay data in standard RL training loops, and thus train end-to-end. Throughout this training process, we show that the supernet gradually learns better cells, leading to alternative architectures which can be highly competitive against manually designed policies, but also verify previous design choices for RL policies.

A good open-domain chatbot should avoid presenting contradictory responses about facts or opinions in a conversational session, known as its consistency capacity. However, evaluating the consistency capacity of a chatbot is still challenging. Employing human judges to interact with chatbots on purpose to check their capacities is costly and low-efficient, and difficult to get rid of subjective bias. In this paper, we propose the Addressing Inquiries about History (AIH), an efficient and practical framework for the consistency evaluation. At the conversation stage, AIH attempts to address appropriate inquiries about the dialogue history to induce the chatbot to redeclare the historical facts or opinions. We carry out the conversation between chatbots, which is more efficient than the human-bot interaction and can also alleviate the subjective bias. In this way, we manage to rapidly obtain a dialog session that contains responses with high contradiction possibilities. At the contradiction recognition stage, we can either employ human judges or a natural language inference (NLI) model to recognize whether the answers to the inquiries are contradictory with history. Finally, we are able to rank chatbots according to the contradiction statistics. Experiments on open-domain chatbots show that our approach can efficiently and reliably assess the consistency capacity of chatbots and achieve a high ranking correlation with the human evaluation. We release the framework and hope to help improve the consistency capacity of chatbots. \footnote{\url{https://github.com/ictnlp/AIH}}

Unlike the conventional wisdom in statistical learning theory, the test error of a deep neural network (DNN) often demonstrates double descent: as the model complexity increases, it first follows a classical U-shaped curve and then shows a second descent. Through bias-variance decomposition, recent studies revealed that the bell-shaped variance is the major cause of model-wise double descent (when the DNN is widened gradually). This paper investigates epoch-wise double descent, i.e., the test error of a DNN also shows double descent as the number of training epoches increases. By extending the bias-variance analysis to epoch-wise double descent of the zero-one loss, we surprisingly find that the variance itself, without the bias, varies consistently with the test error. Inspired by this result, we propose a novel metric, optimization variance (OV), to measure the diversity of model updates caused by the stochastic gradients of random training batches drawn in the same iteration. OV can be estimated using samples from the training set only but correlates well with the (unknown) \emph{test} error, and hence early stopping may be achieved without using a validation set.