High performance computing (HPC) is central to solving large problems in science and engineering through the deployment of massive amounts of computational power. During this period, the core functionality of the software is made more efficient, new features are added, and the software is ported across multiple platforms. Porting of software in general involves the change of compilers, optimization levels, arithmetic libraries, and many other aspects that determine the machine instructions that actually get executed. Unfortunately, such changes do affect the computed results to a significant (and often worrisome) extent. In a majority of cases, there are not easily definable a priori answers one can check against. A programmer ends up comparing the new answer against a trusted baseline previously established or checks for indirect confirmations such as whether physical properties such as energy are conserved. However, such non-systematic efforts might miss underlying issues, and the code may keep misbehaving until these are fixed. In this article, we present real-world evidence to show that ignoring numerical result changes can lead to misleading scientific conclusions. We present techniques and tools that can help computational scientists understand and analyze compiler effects on their scientific code. These techniques are applicable across a wide range of examples to narrow down the root-causes to single files, functions within files, and even computational expressions that affect specific variables. The developer may then rewrite the code selectively and/or suppress the application of certain optimizations to regain more predictable behavior.

Over the past decade, calls for better measures to protect sensitive, personally identifiable information have blossomed into what politicians like to call a "hot-button issue." Certainly, privacy violations have become rampant and people have grown keenly aware of just how vulnerable they are. When it comes to potential remedies, however, proposals have varied widely, leading to bitter, politically charged arguments. To date, what has chiefly come of that have been bureaucratic policies that satisfy almost no one--and infuriate many. Now, into this muddled picture comes differential privacy. First formalized in 2006, it's an approach based on a mathematically rigorous definition of privacy that allows formalization and proof of the guarantees against re-identification offered by a system. While differential privacy has been accepted by theorists for some time, its implementation has turned out to be subtle and tricky, with practical applications only now starting to become available. To date, differential privacy has been adopted by the U.S. Census Bureau, along with a number of technology companies, but what this means and how these organizations have implemented their systems remains a mystery to many. It's also unlikely that the emergence of differential privacy signals an end to all the difficult decisions and trade-offs, but it does signify that there now are measures of privacy that can be quantified and reasoned about--and then used to apply suitable privacy protections. A milestone in the effort to make this capability generally available came in September 2019 when Google released an open source version of the differential privacy library that the company has used with many of its core products. In the exchange that follows, two of the people at Google who were central to the effort to release the library as open source--Damien Desfontaines, privacy software engineer; and Miguel Guevara, who leads Google's differential privacy product development effort--reflect on the engineering challenges that lie ahead, as well as what remains to be done to achieve their ultimate goal of providing privacy protection by default.

In Moshe Vardi's September 2020 column, "Where Have All the Domestic Graduate Students Gone?," the short but woefully incomplete answer is that the wage premium for a Ph.D. in CS is simply too small to justify foregoing five years of industry-level salary. But why is that the case? Part of the answer may be due to government policy discussed back in 1989, when an NSF document addressed the "problem" of Ph.D. salaries being too high, and suggested as a remedy increasing the pool of international students (https://bit.ly/2IuFZl7). This would swell the labor market, holding down wage growth. "A growing influx of foreign Ph.D.'s into U.S. labor markets will hold down the level of Ph.D. salaries to the extent that foreign students are attracted to U.S. doctoral programs as a way of immigrating to the U.S." But the domestic students would find that the resulting wage suppression would make Ph.D. study a bad choice: "... a key issue [for the domestic students] is pay. The relatively modest salary premium for acquiring [a] Ph.D. may be too low to attract a number of able potential graduate students ... A number of them will select alternative career paths ... by choosing to acquire a'professional' degree in business or law ... For these baccalaureates, the effective premium for acquiring a Ph.D. may actually be negative."

Innovations in artificial intelligence (AI) have fundamentally changed the email security landscape in recent years, but it can often be hard to determine what makes one system different than the next. In reality, under that umbrella term significant differences exist in approaches that may determine whether the technology provides genuine protection or simply a perceived notion of defense. The Rise of Fearware When the global pandemic hit, and governments began enforcing travel bans and imposing stringent restrictions, there was undoubtedly a collective sense of fear and uncertainty. As explained in this blog, cybercriminals were quick to capitalize, taking advantage of people's desire for information to send out topical emails related to COVID-19 containing malware or credential-grabbing links. These emails often spoofed the Centers for Disease Control and Prevention (CDC) and, later on, as the economic impact of the pandemic began to take hold, the Small Business Administration (SBA).

Claim: fin(tech) offers lessons for many areas in law 163.

The U.S. government is tasked with protecting classified data and combating potential threats, an area of growing concern with the increasing use of web-based applications required for remote working. Due to high demands, the teams tasked with safeguarding data need a new way--or new capabilities--to scale cybersecurity efforts, especially as many government agencies also face the challenge of limited resources and massively growing data sets and feeds. Pushed by the pandemic, governments are accelerating digital transformation efforts to implement artificial intelligence for cybersecurity needs, as it brings capabilities beyond what manual human surveillance can provide. In fact, the Defense Department's investment in AI has increased from $600 million in fiscal 2016 to $2.5 billion in fiscal 2021. The security operations center is the "mothership" of security within government agencies.

We consider the problem of estimating a function from $n$ noisy samples whose discrete Total Variation (TV) is bounded by $C_n$. We reveal a deep connection to the seemingly disparate problem of Strongly Adaptive online learning (Daniely et al, 2015) and provide an $O(n \log n)$ time algorithm that attains the near minimax optimal rate of $\tilde O (n^{1/3}C_n^{2/3})$ under squared error loss. The resulting algorithm runs online and optimally adapts to the unknown smoothness parameter $C_n$. This leads to a new and more versatile alternative to wavelets-based methods for (1) adaptively estimating TV bounded functions; (2) online forecasting of TV bounded trends in time series.

We propose a framework for resilient autonomous navigation in perceptually challenging unknown environments with mobility-stressing elements such as uneven surfaces with rocks and boulders, steep slopes, negative obstacles like cliffs and holes, and narrow passages. Environments are GPS-denied and perceptually-degraded with variable lighting from dark to lit and obscurants (dust, fog, smoke). Lack of prior maps and degraded communication eliminates the possibility of prior or off-board computation or operator intervention. This necessitates real-time on-board computation using noisy sensor data. To address these challenges, we propose a resilient architecture that exploits redundancy and heterogeneity in sensing modalities. Further resilience is achieved by triggering recovery behaviors upon failure. We propose a fast settling algorithm to generate robust multi-fidelity traversability estimates in real-time. The proposed approach was deployed on multiple physical systems including skid-steer and tracked robots, a high-speed RC car and legged robots, as a part of Team CoSTAR's effort to the DARPA Subterranean Challenge, where the team won 2nd and 1st place in the Tunnel and Urban Circuits, respectively.

We propose a method for online news stream clustering that is a variant of the non-parametric streaming K-means algorithm. Our model uses a combination of sparse and dense document representations, aggregates document-cluster similarity along these multiple representations and makes the clustering decision using a neural classifier. The weighted document-cluster similarity model is learned using a novel adaptation of the triplet loss into a linear classification objective. We show that the use of a suitable fine-tuning objective and external knowledge in pre-trained transformer models yields significant improvements in the effectiveness of contextual embeddings for clustering. Our model achieves a new state-of-the-art on a standard stream clustering dataset of English documents.

We introduce the task of historical text summarisation, where documents in historical forms of a language are summarised in the corresponding modern language. This is a fundamentally important routine to historians and digital humanities researchers but has never been automated. We compile a high-quality gold-standard text summarisation dataset, which consists of historical German and Chinese news from hundreds of years ago summarised in modern German or Chinese. Based on cross-lingual transfer learning techniques, we propose a summarisation model that can be trained even with no cross-lingual (historical to modern) parallel data, and further benchmark it against state-of-the-art algorithms. We report automatic and human evaluations that distinguish the historic to modern language summarisation task from standard cross-lingual summarisation (i.e., modern to modern language), highlight the distinctness and value of our dataset, and demonstrate that our transfer learning approach outperforms standard cross-lingual benchmarks on this task.