AI Avatar Tiktok

AI Avatar Tiktok — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

• 

  Seccomp

  seccomp (short for secure computing) is a computer security facility in the Linux kernel. seccomp allows a process to make a one-way transition into a "secure" state where it cannot make any system calls except exit(), sigreturn(), read() and write() to already-open file descriptors. Should it attempt any other system calls, the kernel will either just log the event or terminate the process with SIGKILL or SIGSYS. In this sense, it does not virtualize the system's resources but isolates the process from them entirely. seccomp mode is enabled via the prctl(2) system call using the PR_SET_SECCOMP argument, or (since Linux kernel 3.17) via the seccomp(2) system call. seccomp mode used to be enabled by writing to a file, /proc/self/seccomp, but this method was removed in favor of prctl(). In some kernel versions, seccomp disables the RDTSC x86 instruction, which returns the number of elapsed processor cycles since power-on, used for high-precision timing. seccomp-bpf is an extension to seccomp that allows filtering of system calls using a configurable policy implemented using Berkeley Packet Filter rules. It is used by OpenSSH and vsftpd as well as the Google Chrome/Chromium web browsers on ChromeOS and Linux. (In this regard seccomp-bpf achieves similar functionality, but with more flexibility and higher performance, to the older systrace—which seems to be no longer supported for Linux.) Some consider seccomp comparable to OpenBSD pledge(2) and FreeBSD capsicum(4). == History == seccomp was first devised by Andrea Arcangeli in January 2005 for use in public grid computing and was originally intended as a means of safely running untrusted compute-bound programs. It was merged into the Linux kernel mainline in kernel version 2.6.12, which was released on March 8, 2005. == Software using seccomp or seccomp-bpf == Android uses a seccomp-bpf filter in the zygote since Android 8.0 Oreo. systemd's sandboxing options are based on seccomp. QEMU, the Quick Emulator, the core component to the modern virtualization together with KVM uses seccomp on the parameter --sandbox Docker – software that allows applications to run inside of isolated containers. Docker can associate a seccomp profile with the container using the --security-opt parameter. Arcangeli's CPUShare was the only known user of seccomp for a while. Writing in February 2009, Linus Torvalds expresses doubt whether seccomp is actually used by anyone. However, a Google engineer replied that Google is exploring using seccomp for sandboxing its Chrome web browser. Firejail is an open source Linux sandbox program that utilizes Linux namespaces, Seccomp, and other kernel-level security features to sandbox Linux and Wine applications. As of Chrome version 20, seccomp-bpf is used to sandbox Adobe Flash Player. As of Chrome version 23, seccomp-bpf is used to sandbox the renderers. Snap specify the shape of their application sandbox using "interfaces" which snapd translates to seccomp, AppArmor and other security constructs vsftpd uses seccomp-bpf sandboxing as of version 3.0.0. OpenSSH has supported seccomp-bpf since version 6.0. Mbox uses ptrace along with seccomp-bpf to create a secure sandbox with less overhead than ptrace alone. LXD, a Ubuntu "hypervisor" for containers Firefox and Firefox OS, which use seccomp-bpf Tor supports seccomp since 0.2.5.1-alpha Lepton, a JPEG compression tool developed by Dropbox uses seccomp Kafel is a configuration language, which converts readable policies into seccompb-bpf bytecode Subgraph OS uses seccomp-bpf Flatpak uses seccomp for process isolation Bubblewrap is a lightweight sandbox application developed from Flatpak minijail uses seccomp for process isolation SydBox uses seccomp-bpf to improve the runtime and security of the ptrace sandboxing used to sandbox package builds on Exherbo Linux distribution. File, a Unix program to determine filetypes, uses seccomp to restrict its runtime environment Zathura, a minimalistic document viewer, uses seccomp filter to implement different sandbox modes Tracker, a indexing and preview application for the GNOME desktop environment, uses seccomp to prevent automatic exploitation of parsing vulnerabilities in media files

  Read more →
• 

  DARPA Prize Competitions

  Over the years, the U.S. Defense Advanced Research Projects Agency (DARPA) has conducted numerous prize competitions to spur innovation. A prize competition allows DARPA to establish an ambitious goal, opening the door to novel approaches from the public that might otherwise appear too risky for experts in a particular field to pursue. == Statutory authorities == In 1999, Congress provided prize competition authority to DARPA in the National Defense Authorization Act for Fiscal Year 2000 (P.L. 106–65), 10 U.S.C. § 4025, formerly 10 U.S.C. §2374a. DARPA also conducts prize competitions under the America COMPETES Act, 15 U.S.C. § 3719. == Recent prize competitions == DARPA Grand Challenge (2004 and 2005) was a prize competition to spur the development of autonomous vehicle technologies. The $1 million prize went unclaimed as no vehicles could complete the challenging desert route from Barstow, CA, to Primm, NV, on March 13, 2004. A year later, on October 8, 2005, the Stanford Racing Team won the $2 million prize during the second competition of the Grand Challenge in the desert Southwest near the California/Nevada state line. DARPA Urban Challenge (2007) required the competitors to build an autonomous vehicle capable of driving in traffic and performing complex maneuvers such as merging, passing, parking, and negotiating intersections. On November 3, 2007, the Carnegie Mellon Team won the $2 million prize, and its vehicle became the first autonomous vehicle that interacted with both manned and unmanned vehicle traffic in an urban environment. DARPA Network Challenge (Red Balloon Challenge) (2009) explored the roles that the Internet and social networking play in solving broad-scope, time-critical problems. On December 5, 2009, the Massachusetts Institute of Technology team won $40,000 by locating the ten moored, eight-foot, red weather balloons at ten places in the United States within seven hours. DARPA Digital Manufacturing Analysis, Correlation and Estimation Challenge (DMACE) (2010) was a three-month contest to showcase the potential of digital manufacturing of advanced materials. The University of California at Santa Barbara team won a $50,000 prize for crushing 180 digitally manufactured (DM) titanium mesh spheres with the most accurate predictive model of the components’ properties. DARPA Shredder Challenge (2011) was to identify and assess potential capabilities and vulnerabilities to sensitive information in the national security community. Participating teams must download the images of the documents shredded into more than 10,000 pieces from the Challenge website, reconstruct the documents, and solve the five puzzles. Of almost 9,000 teams, the San Francisco-based All Your Shreds Are Belong to U.S team won the $50,000 prize. DARPA UAVForge Challenge (2011-2012) aimed to build and test a user-intuitive, backpack-portable unmanned aerial vehicle (UAV) that could quietly fly in and out of critical environments to conduct sustained surveillance for up to three hours. The $100,000 prize was not claimed because none of the 140 teams met the technical matrix. DARPA Cash for Locating & Identifying Quick Response Codes (CLIQR) Quest Challenge (2012) explored the role the Internet and social media played in the timely communication, wide-area team-building, and urgent mobilization required to solve broad scope, time-critical problems. The challenge offered $40,000 to the first individual or team that could locate seven posters appearing in U.S. cities bearing the DARPA logo and a quick response code (QR) within 15 days. No team found and submitted all seven codes. DARPA Fast Adaptable Next-Generation Ground Vehicle (FANG) Challenge (2012-2013) was to use three competitions for the design of an infantry fighting vehicle, culminating in prototypes. In April 2013, DARPA awarded US$1 million to a three-man team during the first competition. DARPA decided not to proceed with the second and third competitions as originally planned and transitioned the technologies to the defense and commercial industry through the Digital Manufacturing and Design Innovation Institute (DMDII). DARPA Spectrum Challenge (2013-2014) sought to demonstrate how a software-defined radio can use a given communication channel in the presence of other users and interfering signals. Three teams emerged as the overall winners, winning a total of $150,000 in prizes. DARPA Chikungunya (CHIKV) Challenge (2014-2015) was a health-related effort to develop the most accurate predictions of CHIKV cases for all Western Hemisphere countries and territories between September 2014 and March 2015. On May 12, 2015, DARPA awarded $500,000 in prizes to the 11 winners of the competition during a scientific review DARPA Robotics Challenge (DRC) (2013-2015) aimed to develop semi-autonomous ground robots that could do "complex tasks in dangerous, degraded, human-engineered environments." A South Korean team won the first prize of $2 million, and two U.S. teams won $1 million and $500,000 as second and third winners. DARPA Cyber Grand Challenge (CGC) (2014 - 2016) was to “create automatic defensive systems capable of reasoning about flaws, formulating patches and deploying them on a network in real time.” The top three winners were awarded prizes of $2 million, $1 million, and $750,000, respectively. DARPA Spectrum Collaboration Challenge (SC2) (2016-2019) aimed to encourage the development of AI-enabled wireless networks to “ensure that the exponentially growing number of military and civilian wireless devices would have full access to the increasingly crowded electromagnetic spectrum.” A team from the University of Florida won the overall top prize of US$2 million at the final SC2 competition. DARPA Subterranean (SubT) Challenge (2017-2021) was to develop robotic technologies to map, navigate, search and exploit complex underground environments. The first-place winners of the system final competition and of the virtual final competition were awarded $2 million and $750,000, respectively, with multiple prizes awarded to the second and third-place winners. DARPA Launch Challenge (2018-2020) was a $12 million satellite launch challenge to demonstrate responsive and flexible space launch capabilities from the small launch providers and was to culminate in two separate launch competitions where the competitors must launch a satellite to low Earth orbit (LEO) within days of each other at different locations in the United States. The competition ended without a winner. DARPA Forecasting Floats in Turbulence (FFT) Challenge (2021) was to spur technologies that could predict the location of sea drifters or floats within 10 days. DARPA awarded $25,000 for first place, with prizes of $15,000 and $10,000 for second place and third place. DARPA Artificial Intelligence Cyber Challenge (AIxCC) (2023–2025) was a two-year challenge and asks competitors to design novel AI systems to secure critical software code on which Americans rely. The total prize money is $29.5 million. In March 2024, the Advanced Research Projects Agency for Health (ARPA-H) partnered with DARPA, contributing an additional $20 million to the competition's prize pool to address software vulnerabilities in medical devices, hospital IT, and biotech equipment. AIxCC collaborates with Google, Microsoft, OpenAI, Anthropic, Linux Foundation, Open Source Security Foundation, Black Hat USA, and DEF CON, all of which provide AIxCC with access to large language models. In August 2024, AIxCC held the semifinal at DEF CON in Las Vegas. DARPA and ARPA-H tested all 42 submissions by running them through various open-source coding projects with deliberately injected vulnerabilities and scored the tools based on their effectiveness in identifying and fixing security flaws. Seven teams, each winning $2 million in the semifinals, competed in the final round of the AIxCC at the August 2025 DEF CON conference. Team Atlanta won first place with a $4 million prize for its cyber reasoning systems, which identified and patched vulnerabilities across 54 million lines of code. DARPA Triage Challenge (2023 – 2026) aims to spur the development of novel physiological features for medical triage, with a total prize money of $7 million. In October 2024, Challenge Event 1 was held in Perry, Georgia, featuring to-scale replicas of disaster sites such as an airplane crash and Hurricane Katrina, and teams competed based on how closely their data aligned with the agency’s official data and how quickly and accurately their autonomous systems could identify individuals most urgently in need of medical care. DARPA concluded the second year of competitions and, in November 2025, named the top performers in systems and data categories, which will advance to the final 2026 competition. The DARPA Lift Challenge (2025-2026) is for participants to design unmanned aerial systems capable of carrying up to four times their own weight, with a minimum payload of 110 pounds. Acco

  Read more →
• 

  Oxa

  Oxa (formerly Oxbotica) is an autonomous vehicle software company, headquartered in Oxfordshire, England, and founded by Paul Newman and Ingmar Posner. == History == In 2013, Newman and Posner led the RobotCar UK project as part of Oxford University's Department of Engineering Science Mobile Robotics Group. RobotCar became the first autonomous vehicle on UK roads. In 2014, the pair used the newly developed technology to found Oxbotica. Oxbotica has raised over $18 million to date and is backed by the IP Group, Parkwalk Advisors and AXA XL. In 2018, Uber's former EMEA business head, Fraser Robinson, was appointed to the board of directors. In May 2019, Ozgur Tohumcu replaced Dr Graeme Smith as Oxbotica's CEO. Also in 2019, the company opened an office in Toronto, Canada. In January 2021, Oxbotica announced it had raised $47 million in a Series B round. In August 2021, the company achieved a safety landmark as the first company to have its autonomy safety case assessed by BSI (British Standards Institution) against the requirements of the UK Code of Practice 2019, PAS 1881:2020 and PAS 1883:2020, certifying the safety conformity of its autonomous vehicle trials and testing. The assessment was completed as part of Project Endeavour, the UK's first multi-city demonstration of autonomous vehicle services and capability. In December 2021, Gavin Jackson was named CEO. In January 2023, the company raised $140 million in a Series C round. In May 2023, the company changed its name to Oxa. Oxa raised $103 million (£77 million) in March 2026, including $50 million from the UK National Wealth Fund. Nvidia's venture capital division, NVentures, also invested in the Series D funding round, along with existing Oxa shareholders IP Group, Australian pension fund Hostplus, and BP Ventures, a division of the UK oil company. == Technology == Oxa designs software and hardware for the conversion of industrial vehicles into autonomous ones. Its full stack, end-to-end Universal Autonomy software is both vehicle and platform-agnostic, with no dependence on external infrastructure such as GPS. It can be deployed in any environment and on any terrain. In addition to underground uses, the technology is also useful in natural canyons and forests, where GPS signals are weak or non-existent, but also in "urban canyons" — cities with tall buildings that obstruct GPS signals for proper navigation. == Public deployments == The LUTZ Pathfinder pod had its first public demonstration in February 2015 in Milton Keynes. The Government-funded project was designed to ensure that autonomous vehicles would comply with the Highway Code. The pod featured autonomous control software from Oxbotica, including 19 sensors, cameras, radar and Lidar. As part of the GATEway Project in 2017, Oxbotica trialled seven autonomous shuttle buses in Greenwich, navigating a two-mile riverside path near London's O2 Arena on a route that is also used by pedestrians and cyclists. Oxbotica ran the UK's first trial of autonomous grocery deliveries that year, with British online supermarket Ocado in London, as the next step in the GATEway Project. In 2018, Oxbotica deployed its autonomous vehicle software at London's Gatwick Airport, which subsequently became the first airport in the world to trial an autonomous shuttle service. The electric-powered vehicles transported staff via airside roads between the airport's North and South terminals. An airside trial of Oxbotica's autonomous driving technology was then successfully completed at Heathrow Airport in partnership with IAG Cargo, the first airside trial of an autonomous vehicle at a UK airport. The Oxbotica-designed CargoPod ran autonomously along a cargo route around the airside perimeter for three weeks. As part of the UK Centre for Connected and Autonomous Vehicles-funded DRIVEN project, Oxbotica is developing and deploying a fleet of Ford Fusion autonomous vehicles running in both London and Oxford on public roads, and in conjunction with its consortium partners, running real-time insurance. AXA XL is partnering with Oxbotica on the development of smart insurance products using Oxbotica's autonomy technology to improve road safety. In 2018, Oxbotica announced a partnership with London private taxi firm Addison Lee to develop and deploy autonomous taxis in the city of London by 2021. A 3D street mapping exercise was conducted in London's Canary Wharf. In 2019, Oxbotica deployed a fleet of their autonomous technology within Ford Mondeo cars on public roads in Stratford, London to test their use in city environments. The £13.2 million project is in collaboration with The DRIVEN Project to develop self-driving cars. == Awards == 2019 Royal Academy of Engineering Silver Medal - Paul Newman 2017 Financial Times ArcelorMittal Boldness in Business Award Barclays Award for Innovation 2016 Frost & Sullivan Award, Technology Leadership for Autonomous Driving Software

  Read more →
• 

  Thompson sampling

  Thompson sampling, named after William R. Thompson, is a heuristic for choosing actions that address the exploration–exploitation dilemma in the multi-armed bandit problem. It consists of choosing the action that maximizes the expected reward with respect to a randomly drawn belief. == Description == Consider a set of contexts X {\displaystyle {\mathcal {X}}} , a set of actions A {\displaystyle {\mathcal {A}}} , and rewards in R {\displaystyle \mathbb {R} } . The aim of the player is to play actions under the various contexts, such as to maximize the cumulative rewards. Specifically, in each round, the player obtains a context x ∈ X {\displaystyle x\in {\mathcal {X}}} , plays an action a ∈ A {\displaystyle a\in {\mathcal {A}}} and receives a reward r ∈ R {\displaystyle r\in \mathbb {R} } following a distribution that depends on the context and the issued action. The elements of Thompson sampling are as follows: a likelihood function P ( r | θ , a , x ) {\displaystyle P(r|\theta ,a,x)} ; a set Θ {\displaystyle \Theta } of parameters θ {\displaystyle \theta } of the distribution of r {\displaystyle r} ; a prior distribution P ( θ ) {\displaystyle P(\theta )} on these parameters; past observations triplets D = { ( x ; a ; r ) } {\displaystyle {\mathcal {D}}=\{(x;a;r)\}} ; a posterior distribution P ( θ | D ) ∝ P ( D | θ ) P ( θ ) {\displaystyle P(\theta |{\mathcal {D}})\propto P({\mathcal {D}}|\theta )P(\theta )} , where P ( D | θ ) {\displaystyle P({\mathcal {D}}|\theta )} is the likelihood function. Thompson sampling consists of playing the action a ∗ ∈ A {\displaystyle a^{\ast }\in {\mathcal {A}}} according to the probability that it maximizes the expected reward; action a ∗ {\displaystyle a^{\ast }} is chosen with probability ∫ I [ E ( r | a ∗ , x , θ ) = max a ′ E ( r | a ′ , x , θ ) ] P ( θ | D ) d θ , {\displaystyle \int \mathbb {I} \left[\mathbb {E} (r|a^{\ast },x,\theta )=\max _{a'}\mathbb {E} (r|a',x,\theta )\right]P(\theta |{\mathcal {D}})d\theta ,} where I {\displaystyle \mathbb {I} } is the indicator function. In practice, the rule is implemented by sampling. In each round, parameters θ ∗ {\displaystyle \theta ^{\ast }} are sampled from the posterior P ( θ | D ) {\displaystyle P(\theta |{\mathcal {D}})} , and an action a ∗ {\displaystyle a^{\ast }} chosen that maximizes E [ r | θ ∗ , a ∗ , x ] {\displaystyle \mathbb {E} [r|\theta ^{\ast },a^{\ast },x]} , i.e. the expected reward given the sampled parameters, the action, and the current context. Conceptually, this means that the player instantiates their beliefs randomly in each round according to the posterior distribution, and then acts optimally according to them. In most practical applications, it is computationally onerous to maintain and sample from a posterior distribution over models. As such, Thompson sampling is often used in conjunction with approximate sampling techniques. == History == Thompson sampling was originally described by Thompson in 1933. It was subsequently rediscovered numerous times independently in the context of multi-armed bandit problems. A first proof of convergence for the bandit case has been shown in 1997. The first application to Markov decision processes was in 2000. A related approach (see Bayesian control rule) was published in 2010. In 2010 it was also shown that Thompson sampling is instantaneously self-correcting. Asymptotic convergence results for contextual bandits were published in 2011. Thompson Sampling has been widely used in many online learning problems including A/B testing in website design and online advertising, and accelerated learning in decentralized decision making. A Double Thompson Sampling (D-TS) algorithm has been proposed for dueling bandits, a variant of traditional MAB, where feedback comes in the form of pairwise comparison. == Relationship to other approaches == === Probability matching === Probability matching is a decision strategy in which predictions of class membership are proportional to the class base rates. Thus, if in the training set positive examples are observed 60% of the time, and negative examples are observed 40% of the time, the observer using a probability-matching strategy will predict (for unlabeled examples) a class label of "positive" on 60% of instances, and a class label of "negative" on 40% of instances. === Bayesian control rule === A generalization of Thompson sampling to arbitrary dynamical environments and causal structures, known as Bayesian control rule, has been shown to be the optimal solution to the adaptive coding problem with actions and observations. In this formulation, an agent is conceptualized as a mixture over a set of behaviours. As the agent interacts with its environment, it learns the causal properties and adopts the behaviour that minimizes the relative entropy to the behaviour with the best prediction of the environment's behaviour. If these behaviours have been chosen according to the maximum expected utility principle, then the asymptotic behaviour of the Bayesian control rule matches the asymptotic behaviour of the perfectly rational agent. The setup is as follows. Let a 1 , a 2 , … , a T {\displaystyle a_{1},a_{2},\ldots ,a_{T}} be the actions issued by an agent up to time T {\displaystyle T} , and let o 1 , o 2 , … , o T {\displaystyle o_{1},o_{2},\ldots ,o_{T}} be the observations gathered by the agent up to time T {\displaystyle T} . Then, the agent issues the action a T + 1 {\displaystyle a_{T+1}} with probability: P ( a T + 1 | a ^ 1 : T , o 1 : T ) , {\displaystyle P(a_{T+1}|{\hat {a}}_{1:T},o_{1:T}),} where the "hat"-notation a ^ t {\displaystyle {\hat {a}}_{t}} denotes the fact that a t {\displaystyle a_{t}} is a causal intervention (see Causality), and not an ordinary observation. If the agent holds beliefs θ ∈ Θ {\displaystyle \theta \in \Theta } over its behaviors, then the Bayesian control rule becomes P ( a T + 1 | a ^ 1 : T , o 1 : T ) = ∫ Θ P ( a T + 1 | θ , a ^ 1 : T , o 1 : T ) P ( θ | a ^ 1 : T , o 1 : T ) d θ {\displaystyle P(a_{T+1}|{\hat {a}}_{1:T},o_{1:T})=\int _{\Theta }P(a_{T+1}|\theta ,{\hat {a}}_{1:T},o_{1:T})P(\theta |{\hat {a}}_{1:T},o_{1:T})\,d\theta } , where P ( θ | a ^ 1 : T , o 1 : T ) {\displaystyle P(\theta |{\hat {a}}_{1:T},o_{1:T})} is the posterior distribution over the parameter θ {\displaystyle \theta } given actions a 1 : T {\displaystyle a_{1:T}} and observations o 1 : T {\displaystyle o_{1:T}} . In practice, the Bayesian control amounts to sampling, at each time step, a parameter θ ∗ {\displaystyle \theta ^{\ast }} from the posterior distribution P ( θ | a ^ 1 : T , o 1 : T ) {\displaystyle P(\theta |{\hat {a}}_{1:T},o_{1:T})} , where the posterior distribution is computed using Bayes' rule by only considering the (causal) likelihoods of the observations o 1 , o 2 , … , o T {\displaystyle o_{1},o_{2},\ldots ,o_{T}} and ignoring the (causal) likelihoods of the actions a 1 , a 2 , … , a T {\displaystyle a_{1},a_{2},\ldots ,a_{T}} , and then by sampling the action a T + 1 ∗ {\displaystyle a_{T+1}^{\ast }} from the action distribution P ( a T + 1 | θ ∗ , a ^ 1 : T , o 1 : T ) {\displaystyle P(a_{T+1}|\theta ^{\ast },{\hat {a}}_{1:T},o_{1:T})} . === Upper-confidence-bound (UCB) algorithms === Thompson sampling and upper-confidence bound algorithms share a fundamental property that underlies many of their theoretical guarantees. Roughly speaking, both algorithms allocate exploratory effort to actions that might be optimal and are in this sense "optimistic". Leveraging this property, one can translate regret bounds established for UCB algorithms to Bayesian regret bounds for Thompson sampling or unify regret analysis across both these algorithms and many classes of problems.

  Read more →
• 

  Concurrency control

  In information technology and computer science, especially in the fields of computer programming, operating systems, multiprocessors, and databases, concurrency control ensures that correct results for concurrent operations are generated, while getting those results as quickly as possible. Computer systems, both software and hardware, consist of modules, or components. Each component is designed to operate correctly, i.e., to obey or to meet certain consistency rules. When components that operate concurrently interact by messaging or by sharing accessed data (in memory or storage), a certain component's consistency may be violated by another component. The general area of concurrency control provides rules, methods, design methodologies, and theories to maintain the consistency of components operating concurrently while interacting, and thus the consistency and correctness of the whole system. Introducing concurrency control into a system means applying operation constraints which typically result in some performance reduction. Operation consistency and correctness should be achieved with as good as possible efficiency, without reducing performance below reasonable levels. Concurrency control can require significant additional complexity and overhead in a concurrent algorithm compared to the simpler sequential algorithm. For example, a failure in concurrency control can result in data corruption from torn read or write operations. == Concurrency control in databases == Comments: This section is applicable to all transactional systems, i.e., to all systems that use database transactions (atomic transactions; e.g., transactional objects in Systems management and in networks of smartphones which typically implement private, dedicated database systems), not only general-purpose database management systems (DBMSs). DBMSs need to deal also with concurrency control issues not typical just to database transactions but rather to operating systems in general. These issues (e.g., see Concurrency control in operating systems below) are out of the scope of this section. Concurrency control in Database management systems (DBMS; e.g., Bernstein et al. 1987, Weikum and Vossen 2001), other transactional objects, and related distributed applications (e.g., Grid computing and Cloud computing) ensures that database transactions are performed concurrently without violating the data integrity of the respective databases. Thus concurrency control is an essential element for correctness in any system where two database transactions or more, executed with time overlap, can access the same data, e.g., virtually in any general-purpose database system. Consequently, a vast body of related research has been accumulated since database systems emerged in the early 1970s. A well established concurrency control theory for database systems is outlined in the references mentioned above: serializability theory, which allows to effectively design and analyze concurrency control methods and mechanisms. An alternative theory for concurrency control of atomic transactions over abstract data types is presented in (Lynch et al. 1993), and not utilized below. This theory is more refined, complex, with a wider scope, and has been less utilized in the Database literature than the classical theory above. Each theory has its pros and cons, emphasis and insight. To some extent they are complementary, and their merging may be useful. To ensure correctness, a DBMS usually guarantees that only serializable transaction schedules are generated, unless serializability is intentionally relaxed to increase performance, but only in cases where application correctness is not harmed. For maintaining correctness in cases of failed (aborted) transactions (which can always happen for many reasons) schedules also need to have the recoverability (from abort) property. A DBMS also guarantees that no effect of committed transactions is lost, and no effect of aborted (rolled back) transactions remains in the related database. Overall transaction characterization is usually summarized by the ACID rules below. As databases have become distributed, or needed to cooperate in distributed environments (e.g., Federated databases in the early 1990, and Cloud computing currently), the effective distribution of concurrency control mechanisms has received special attention. === Database transaction and the ACID rules === The concept of a database transaction (or atomic transaction) has evolved in order to enable both a well understood database system behavior in a faulty environment where crashes can happen any time, and recovery from a crash to a well understood database state. A database transaction is a unit of work, typically encapsulating a number of operations over a database (e.g., reading a database object, writing, acquiring lock, etc.), an abstraction supported in database and also other systems. Each transaction has well defined boundaries in terms of which program/code executions are included in that transaction (determined by the transaction's programmer via special transaction commands). Every database transaction obeys the following rules (by support in the database system; i.e., a database system is designed to guarantee them for the transactions it runs): Atomicity - Either the effects of all or none of its operations remain ("all or nothing" semantics) when a transaction is completed (committed or aborted respectively). In other words, to the outside world a committed transaction appears (by its effects on the database) to be indivisible (atomic), and an aborted transaction does not affect the database at all. Either all the operations are done or none of them are. Consistency - Every transaction must leave the database in a consistent (correct) state, i.e., maintain the predetermined integrity rules of the database (constraints upon and among the database's objects). A transaction must transform a database from one consistent state to another consistent state (however, it is the responsibility of the transaction's programmer to make sure that the transaction itself is correct, i.e., performs correctly what it intends to perform (from the application's point of view) while the predefined integrity rules are enforced by the DBMS). Thus since a database can be normally changed only by transactions, all the database's states are consistent. Isolation - Transactions cannot interfere with each other (as an end result of their executions). Moreover, usually (depending on concurrency control method) the effects of an incomplete transaction are not even visible to another transaction. Providing isolation is the main goal of concurrency control. Durability - Effects of successful (committed) transactions must persist through crashes (typically by recording the transaction's effects and its commit event in a non-volatile memory). The concept of atomic transaction has been extended during the years to what has become Business transactions which actually implement types of Workflow and are not atomic. However also such enhanced transactions typically utilize atomic transactions as components. === Why is concurrency control needed? === If transactions are executed serially, i.e., sequentially with no overlap in time, no transaction concurrency exists. However, if concurrent transactions with interleaving operations are allowed in an uncontrolled manner, some unexpected, undesirable results may occur, such as: The lost update problem: A second transaction writes a second value of a data-item (datum) on top of a first value written by a first concurrent transaction, and the first value is lost to other transactions running concurrently which need, by their precedence, to read the first value. The transactions that have read the wrong value end with incorrect results. The dirty read problem: Transactions read a value written by a transaction that has been later aborted. This value disappears from the database upon abort, and should not have been read by any transaction ("dirty read"). The reading transactions end with incorrect results. The incorrect summary problem: While one transaction takes a summary over the values of all the instances of a repeated data-item, a second transaction updates some instances of that data-item. The resulting summary does not reflect a correct result for any (usually needed for correctness) precedence order between the two transactions (if one is executed before the other), but rather some random result, depending on the timing of the updates, and whether certain update results have been included in the summary or not. Most high-performance transactional systems need to run transactions concurrently to meet their performance requirements. Thus, without concurrency control such systems can neither provide correct results nor maintain their databases consistently. === Concurrency control mechanisms === ==== Categories ==== The main categories of concurrency control mechanis

  Read more →
• 

  Oblivion (2013 film)

  Oblivion is a 2013 American epic post-apocalyptic science fiction action film produced and directed by Joseph Kosinski from a screenplay by Karl Gajdusek and Michael deBruyn, starring Tom Cruise in the main role alongside Morgan Freeman, Olga Kurylenko, Andrea Riseborough, Nikolaj Coster-Waldau, and Melissa Leo in supporting roles. Based on Kosinski's unpublished Radical Comics graphic novel of the same name, the film pays homage to 1970s sci-fi, and is a "love story" set in 2077 on an Earth desolated by an alien war; a maintenance technician on the verge of completing his mission finds a woman who survived from a space ship crash, leading him to question his purpose and discover the truth about the war. Oblivion premiered in Buenos Aires on March 26, 2013, and was released in theaters by Universal Pictures on April 19. The film grossed $286 million worldwide on a production budget of $120 million and received mixed reviews from critics. == Plot == In 2017, aliens known as Scavengers attack Earth and destroy the Moon, triggering global natural disasters. Although humanity wins the war using nuclear weapons, Earth is left uninhabitable. Sixty years later, the remnants of humanity have relocated to a colony on Saturn's moon Titan, except for Unit 49—technician Jack and his communications officer Victoria—who are scheduled to join them in two weeks. The pair oversee hydro rigs that convert seawater into fusion energy for the Tet, the last remaining human colony ship in orbit. Though Jack and Victoria are romantically involved and have had their memories erased for security reasons, Jack experiences recurring dreams of an unknown woman. He also secretly visits a hidden, verdant valley where he has built a lakeside cabin and collects relics of Earth's past. While investigating a missing drone—autonomous, highly advanced, and heavily armed machines—Jack is nearly captured by Scavengers. Later, he discovers the Scavengers are transmitting a signal into space. A NASA pod crash-lands at the signal's coordinates, carrying five humans in suspended animation, including the woman from Jack's dreams. A drone arrives and destroys four of the pods, but Jack rescues the remaining one and brings the unconscious woman to Unit 49's base. After reviving her, Jack and Victoria learn that the woman, Julia, has been in stasis aboard the Odyssey spaceship since 2017. Julia insists on recovering the ship's flight recorder. However, she and Jack are captured by Scavengers and brought to the Raven Rock Mountain Complex. Their leader, Malcolm, reveals that the Scavengers are actually surviving humans. Malcolm needs Jack to reprogram a captured drone to deliver a nuclear bomb, built from Odyssey's reactor, to the Tet. Jack refuses, so Malcolm releases him and Julia, urging him to seek the truth in the radiation zone, which is supposedly deadly and off-limits. Julia helps Jack recall that she is his wife, and fragments of his memories begin to return. When they arrive back at Unit 49, a devastated Victoria informs Sally, the Tet's mission controller, that she and Jack are no longer an "effective team." A drone activates and kills Victoria. Jack and Julia destroy the drone, but crash their aircraft inside the radiation zone. There, they encounter another version of Jack—"Jack-52"—who arrives to repair the drone. Jack subdues him, but Julia is seriously injured in the fight. Jack impersonates his clone to infiltrate Unit 52, meets Victoria-52, and steals medical supplies for Julia. They rest at his cabin. At Raven Rock, Malcolm reveals the truth: humanity lost the war, and the Tet is an alien machine intelligence harvesting Earth's resources. After the Moon's destruction, the Tet deployed thousands of clones of astronaut Jack Harper—brainwashed into obedience—to exterminate the remaining humans. Malcolm had assumed these clones were inhuman until witnessing Jack show interest in a discarded book, hinting at lingering humanity. Jack reprograms the captured drone, but it is destroyed in a surprise attack by other drones, leaving Malcolm badly wounded. Jack and Julia resolve to deliver the bomb themselves; Julia enters a stasis pod. En route, Jack listens to the Odyssey's flight recorder, which reveals the original Jack Harper and Victoria were astronauts sent to explore Titan before being confronted by the Tet. The pair were captured, but not before Jack ejected the remaining crew—including Julia—in stasis pods to protect them. Jack gains access to the Tet by claiming he is delivering Julia, as previously instructed. However, the stasis pod contains a dying Malcolm. Jack and Malcolm detonate the bomb, destroying the Tet and themselves. Julia later awakens at the cabin. Three years later, Julia lives there and it is revealed she had a daughter with Jack. A group of Raven Rock survivors arrives, alongside Jack-52, who has begun regaining fragments of his own lost identity. == Cast == Tom Cruise as Jack Harper—Tech 49, a technician who works to repair drones on Earth and questions his mission. Originally, he was the American commander of a mission en route to Titan who was captured by the Tet and cloned to fight humanity. Cruise also plays Jack Harper—Tech 52, a clone who seeks out Julia after the destruction of the Tet. Morgan Freeman as Malcolm Beech, an American veteran soldier and leader of a large community of scavengers, the human survivors of the alien Tet's attacks. Olga Kurylenko as Julia Rusakova Harper, Jack's wife and a Russian crew member on the Odyssey, who was sent back towards Earth by her husband to protect her from the initial contact with the Tet. Andrea Riseborough as Victoria "Vika" Olsen, Jack's communications partner and housemate. Originally, she was the British co-pilot of Jack's mission to Titan who was captured and cloned to assist in the Tet's war on humanity. Riseborough also plays a clone of Vika who Jack misleads to obtain medical supplies. Nikolaj Coster-Waldau as Sergeant Sykes, the main military commander of Beech's community of scavengers who is skeptical of Jack at first. Melissa Leo as the Tet, an alien artificial intelligence seeking to acquire Earth's natural resources and wipe out humanity. Leo also plays Sally, the mission director of Jack and Julia's mission to Titan; her likeness was copied by the Tet to serve as its visual and auditory representation. Zoë Bell as Kara, a soldier and member of the scavengers. == Production == === Development === Joseph Kosinski started the movie process by beginning work on a graphic novel called Oblivion featuring his story. While the completion of this would be teased to the public and the concept was used to pitch the movie, it was never finished and Kosinski claims he never intended to, stating it was "just a stage in the project [of film development]". Arvid Nelson was billed as co-writer and Radical Comics was attached as publisher. The novel was never finished; Kosinski explaining: "the partnership with Radical Comics allowed me to continue working on the story by developing a series of images and continuing to refine the story more over a period of years. Then I basically used all that development as a pitch kit to the studio. So even though we really never released it as an illustrated novel the story is being told as a film, which was always the intention." Walt Disney Pictures, which produced Kosinski's previous film Tron: Legacy (2010), acquired the Oblivion film adaptation rights from Radical Comics and Kosinski after a heated auction in August 2010. The film was a directing vehicle for Kosinski, with Barry Levine producing, and Jesse Berger executive producing. Other studios that made bids on the film were Paramount Pictures, 20th Century Fox, and Universal Pictures. Disney subsequently released the rights after realizing the PG-rated film they envisioned, in line with their family-oriented reputation, would require too many story changes. Universal, which had also bid for the original rights, then bought them from Kosinski and Radical and authorized a PG-13 film version. The film's script was originally written by Kosinski and William Monahan and underwent a first rewrite by Karl Gajdusek. When the film passed into Universal's hands, a final rewrite was done by Michael Arndt, under the pen name "Michael deBruyn". Universal was particularly appreciative of the script, saying, "It's one of the most beautiful scripts we've ever come across." The Bubble Ship operated by Cruise's main character, Jack 49, was inspired by the Bell 47 helicopter (often colloquially referred to as a "bubble cockpit" helicopter), a utilitarian 1947 vehicle with a transparent round canopy that Kosinski saw in the lobby of the Museum of Modern Art in Manhattan, and which he likened to a dragonfly. Daniel Simon, who previously worked with Kosinski as the lead vehicle designer on Tron: Legacy, was tasked with creating the Bubble Ship from this basis, incorporating elements evocative of an advanced fighter

  Read more →
• 

  Whisper (speech recognition system)

  Whisper is a machine learning model for speech recognition and transcription, created by OpenAI and first released as open-source software in September 2022. It is capable of transcribing speech in English and multiple other languages, and can translate several non-English languages into English. Whisper is a weakly-supervised deep learning acoustic model, made using an encoder-decoder transformer architecture. OpenAI claims that the combination of different training data and post-training filtering used in its development has led to improved recognition of accents, background noise, and jargon compared to previous approaches. While the model does not outperform larger, more specialized models and still experiences AI hallucination, it has been showed to be useful for general sound recognition and has many applications across different industries. == Background == Speech recognition has had a long history in research; the first approaches made use of statistical methods, such as dynamic time warping, and later hidden Markov models. At around the 2010s, deep neural network approaches became more common for speech recognition models, which were enabled by the availability of large datasets ("big data") and increased computational performance. Early approaches to deep learning in speech recognition included convolutional neural networks, which were limited due to their inability to capture sequential data, which later led to developments of Seq2seq approaches, which include recurrent neural networks, which made use of long short-term memory. Transformers, introduced in 2017 by Google, displaced many prior state-of-the-art approaches across a wide range in machine learning, and started becoming the core neural architecture in fields such as language modeling and computer vision. Weakly-supervised approaches to training acoustic models were recognized in the early 2020s as promising for speech recognition approaches using deep neural networks. According to a NYT report, in 2021 OpenAI believed they exhausted sources of higher-quality data to train their large language models and decided to complement scraped web text with transcriptions of YouTube videos and podcasts, and developed Whisper to solve this task. Whisper Large V2 was released on December 8, 2022, followed by Whisper Large V3 being released in November 2023, during the OpenAI Dev Day. In March 2025, OpenAI released new transcription models based on GPT-4o and GPT-4o mini, both of which have lower error rates than Whisper. == Architecture == The Whisper architecture is based on an encoder-decoder transformer. Input audio is resampled to 16,000 Hertz (Hz) and converted to an 80-channel Log-magnitude Mel spectrogram using 25 ms windows with a 10 ms stride. The spectrogram is then normalized to a [-1, 1] range with near-zero mean. The encoder takes this Mel spectrogram as input and processes it. It first passes through two convolutional layers. Sinusoidal positional embeddings are added. It is then processed by a series of Transformer encoder blocks (with pre-activation residual connections). The encoder's output is layer normalized. The decoder is a standard transformer decoder. It has the same width and Transformer blocks as the encoder. It uses learned positional embeddings and tied input-output token representations (using the same weight matrix for both the input and output embeddings). It uses a byte-pair encoding tokenizer, of the same kind as used in GPT-2. English-only models use the GPT-2 vocabulary, while multilingual models employ a re-trained multilingual vocabulary with the same number of words. Special tokens are used to allow the decoder to perform multiple tasks: Tokens that denote language (one unique token per language). Tokens that specify task (<|transcribe|> or <|translate|>). Tokens that specify if no timestamps are present (<|notimestamps|>). If the token is not present, then the decoder predicts timestamps relative to the segment, and quantized to 20 ms intervals. <|nospeech|> for voice activity detection. <|startoftranscript|>, and <|endoftranscript|> . Any text that appears before <|startoftranscript|> is not generated by the decoder, but given to the decoder as context. Loss is only computed over non-contextual parts of the sequence, i.e. tokens between these two special tokens. == Training data == The training dataset consists of 680,000 hours of labeled audio-transcript pairs sourced from the internet using semi-supervised learning. This includes 117,000 hours in 96 non-English languages and 125,000 hours of X→English translation data, where X stands for any non-English language. Preprocessing involved standardization of transcripts, filtering to remove machine-generated transcripts using heuristics (e.g., punctuation, capitalization), language identification and matching with transcripts, fuzzy deduplication, and deduplication with evaluation datasets to avoid data contamination. Speechless segments were also included to allow voice activity detection training. For the files still remaining after the filtering process, audio files were then broken into 30-second segments paired with the subset of the transcript that occurs within that time. If this predicted spoken language differed from the language of the text transcript associated with the audio, that audio-transcript pair was not used for training the speech recognition models, but instead for training translation. The model was trained using the AdamW optimizer with gradient norm clipping and a linear learning rate decay with warmup, with batch size 256 segments. Training proceeded for 1 million updates (approximately 2-3 epochs). No data augmentation or regularization, except for the Large V2 model, which used SpecAugment, Stochastic Depth, and BPE Dropout. The training used data parallelism with float16, dynamic loss scaling, and activation checkpointing. === Post-training filtering === After training the first model, researchers ran it on different subsets of the training data, each representing a distinct source. Data sources were ranked by a combination of their error rate and size. Manual inspection of the top-ranked sources (high error, large size) helped determine if the source was low quality (e.g., partial transcriptions, inaccurate alignment). After training, it was fine-tuned to suppress the prediction of speaker names and low-quality sources were then removed. == Capacity == While Whisper does not outperform models which specialize in the LibriSpeech dataset, when tested across many datasets, it is more robust and makes 55.2% fewer errors than other models. Whisper has a differing error rate with respect to transcribing different languages, with a higher word error rate in languages not well-represented in the training data. The authors found that multi-task learning improved overall performance compared to models specialized to one task. They conjectured that the best Whisper model trained is still underfitting the dataset, and larger models and longer training can result in better models. Third-party evaluations have found varying levels of AI hallucination. A study of transcripts of public meetings found hallucinations in eight out of every 10 transcripts, while an engineer discovered hallucinations in "about half" of 100 hours of transcriptions and a developer identified them in "nearly every one" of 26,000 transcripts. A study of 13,140 short audio segments (averaging 10 seconds) found 187 hallucinations (1.4%), 38% of which generated text that could be harmful because it inserted false references to things like race, non-existent medications, or violent events that were not in the audio. == Applications == The model has been used as the base for many applications, such as a unified model for speech recognition and more general sound recognition. Whisper has also been integrated into the workflow of biomedical research. In 2025, a study on Alzheimer's disease detection used the model to transcribe spontaneous speech recordings. The transcripts that were generated by the model were combined with LLM vector embeddings and traditional classifiers to help classify the patients' health. Another application is when OVALYTICS incorporated Whisper to transcribe YouTube videos and automate content moderation systems, which improved its detection of offensive content. The model has also been used in academic libraries and cultral heritage institutions to generate transcripts and captions for their digitized audiovisual collections. In a 2025 case study, Emory University Libraries found that Whisper reduced the labor used in transcription by around 30-35%, shifting work from text creation to text correction. However, human review is still necessary to make sure accuracy, formatting, and accessibility are all standard.

  Read more →
• 

  Fuzzy electronics

  Fuzzy electronics is an electronic technology that uses fuzzy logic, instead of the two-state Boolean logic more commonly used in digital electronics. Fuzzy electronics is fuzzy logic implemented on dedicated hardware. This is to be compared with fuzzy logic implemented in software running on a conventional processor. Fuzzy electronics has a wide range of applications, including control systems and artificial intelligence. == History == The first fuzzy electronic circuit was built by Takeshi Yamakawa et al. in 1980 using discrete bipolar transistors. The first industrial fuzzy application was in a cement kiln in Denmark in 1982. The first VLSI fuzzy electronics was by Masaki Togai and Hiroyuki Watanabe in 1984. In 1987, Yamakawa built the first analog fuzzy controller. The first digital fuzzy processors came in 1988 by Togai (Russo, pp. 2–6). In the early 1990s, the first fuzzy logic chips were presented to the public. Two companies which are Omron and NEC have announced the development of dedicated fuzzy electronic hardware in the year 1991. Two years later, the Japanese Omron Cooperation has shown a working fuzzy chip during a technical fair.

  Read more →
• 

  Security switch

  A security switch is a hardware device designed to protect computers, laptops, smartphones and similar devices from unauthorized access or operation, distinct from a virtual security switch which offers software protection. Security switches should be operated by an authorized user only; for this reason, it should be isolated from other devices, in order to prevent unauthorized access, and it should not be possible to bypass it, in order to prevent malicious manipulation. The primary purpose of a security switch is to provide protection against surveillance, eavesdropping, malware, spyware, and theft of digital devices. Unlike other protections or techniques, a security switch can provide protection even if security has already been breached, since it does not have any access from other components and is not accessible by software. It can additionally disconnect or block peripheral devices, and perform "man in the middle" operations. A security switch can be used for human presence detection since it can only be initiated by a human operator. It can also be used as a firewall. == Types == === Hardware kill switch === A hardware kill switch (HKS) is a physical switch that cuts the signal or power line to the device or disable the chip running them. == Examples == A cellphone is compromised by malicious software, and the device initiates video and audio recording. When the user activates the “prevent capture of audio/video” mode of the security switch, that either physically disconnects or cut the power to the microphone and the camera, which stops the recording. A laptop that has an embedded security switch is stolen. The security switch detects a lack of communication from a specific external source for 12 hours, and responds by disconnecting the screen, keyboard and other key components, rendering the laptop useless, with no possibility of recovery, even with a full format. A user wishes to prevent tracking of their location. The user then activates geolocation protection and the security switch disables all GPS communication, eliminating the possibility of tracking the device's location. A user desires to eliminate the possibility of their PIN being copied from their smartphone. They can activate the secure input function, causing the security switch to disconnect the touch screen from the operating system, so input signals are not available to any devices except the switch. A security switch performs scheduled monitoring and finds that a program is attempting to download malicious content from the internet. It then activates internet security function and disables internet access, interrupting the download. If laptop software is compromised by air-gap malware, the user may activate the security switch and disconnect the speaker and microphone, so it can not establish communication with the device. == History == Google started to work on a hardware kill switch for AI in 2016. In 2019, Apple, and Google, along with a handful of smaller players, are designing “kill switches” that cut the power to the microphones or cameras in their devices. Googles first product that implemented this is Nest Hub Max. Hardware kill switches are already available and widely tested on the PinePhone, Librem, Shiftphone, to cut power to the input peripherals (microphone, camera) but also the network connectivity modules (wifi, cellular network).

  Read more →
• 

  Whisper (speech recognition system)

  Whisper is a machine learning model for speech recognition and transcription, created by OpenAI and first released as open-source software in September 2022. It is capable of transcribing speech in English and multiple other languages, and can translate several non-English languages into English. Whisper is a weakly-supervised deep learning acoustic model, made using an encoder-decoder transformer architecture. OpenAI claims that the combination of different training data and post-training filtering used in its development has led to improved recognition of accents, background noise, and jargon compared to previous approaches. While the model does not outperform larger, more specialized models and still experiences AI hallucination, it has been showed to be useful for general sound recognition and has many applications across different industries. == Background == Speech recognition has had a long history in research; the first approaches made use of statistical methods, such as dynamic time warping, and later hidden Markov models. At around the 2010s, deep neural network approaches became more common for speech recognition models, which were enabled by the availability of large datasets ("big data") and increased computational performance. Early approaches to deep learning in speech recognition included convolutional neural networks, which were limited due to their inability to capture sequential data, which later led to developments of Seq2seq approaches, which include recurrent neural networks, which made use of long short-term memory. Transformers, introduced in 2017 by Google, displaced many prior state-of-the-art approaches across a wide range in machine learning, and started becoming the core neural architecture in fields such as language modeling and computer vision. Weakly-supervised approaches to training acoustic models were recognized in the early 2020s as promising for speech recognition approaches using deep neural networks. According to a NYT report, in 2021 OpenAI believed they exhausted sources of higher-quality data to train their large language models and decided to complement scraped web text with transcriptions of YouTube videos and podcasts, and developed Whisper to solve this task. Whisper Large V2 was released on December 8, 2022, followed by Whisper Large V3 being released in November 2023, during the OpenAI Dev Day. In March 2025, OpenAI released new transcription models based on GPT-4o and GPT-4o mini, both of which have lower error rates than Whisper. == Architecture == The Whisper architecture is based on an encoder-decoder transformer. Input audio is resampled to 16,000 Hertz (Hz) and converted to an 80-channel Log-magnitude Mel spectrogram using 25 ms windows with a 10 ms stride. The spectrogram is then normalized to a [-1, 1] range with near-zero mean. The encoder takes this Mel spectrogram as input and processes it. It first passes through two convolutional layers. Sinusoidal positional embeddings are added. It is then processed by a series of Transformer encoder blocks (with pre-activation residual connections). The encoder's output is layer normalized. The decoder is a standard transformer decoder. It has the same width and Transformer blocks as the encoder. It uses learned positional embeddings and tied input-output token representations (using the same weight matrix for both the input and output embeddings). It uses a byte-pair encoding tokenizer, of the same kind as used in GPT-2. English-only models use the GPT-2 vocabulary, while multilingual models employ a re-trained multilingual vocabulary with the same number of words. Special tokens are used to allow the decoder to perform multiple tasks: Tokens that denote language (one unique token per language). Tokens that specify task (<|transcribe|> or <|translate|>). Tokens that specify if no timestamps are present (<|notimestamps|>). If the token is not present, then the decoder predicts timestamps relative to the segment, and quantized to 20 ms intervals. <|nospeech|> for voice activity detection. <|startoftranscript|>, and <|endoftranscript|> . Any text that appears before <|startoftranscript|> is not generated by the decoder, but given to the decoder as context. Loss is only computed over non-contextual parts of the sequence, i.e. tokens between these two special tokens. == Training data == The training dataset consists of 680,000 hours of labeled audio-transcript pairs sourced from the internet using semi-supervised learning. This includes 117,000 hours in 96 non-English languages and 125,000 hours of X→English translation data, where X stands for any non-English language. Preprocessing involved standardization of transcripts, filtering to remove machine-generated transcripts using heuristics (e.g., punctuation, capitalization), language identification and matching with transcripts, fuzzy deduplication, and deduplication with evaluation datasets to avoid data contamination. Speechless segments were also included to allow voice activity detection training. For the files still remaining after the filtering process, audio files were then broken into 30-second segments paired with the subset of the transcript that occurs within that time. If this predicted spoken language differed from the language of the text transcript associated with the audio, that audio-transcript pair was not used for training the speech recognition models, but instead for training translation. The model was trained using the AdamW optimizer with gradient norm clipping and a linear learning rate decay with warmup, with batch size 256 segments. Training proceeded for 1 million updates (approximately 2-3 epochs). No data augmentation or regularization, except for the Large V2 model, which used SpecAugment, Stochastic Depth, and BPE Dropout. The training used data parallelism with float16, dynamic loss scaling, and activation checkpointing. === Post-training filtering === After training the first model, researchers ran it on different subsets of the training data, each representing a distinct source. Data sources were ranked by a combination of their error rate and size. Manual inspection of the top-ranked sources (high error, large size) helped determine if the source was low quality (e.g., partial transcriptions, inaccurate alignment). After training, it was fine-tuned to suppress the prediction of speaker names and low-quality sources were then removed. == Capacity == While Whisper does not outperform models which specialize in the LibriSpeech dataset, when tested across many datasets, it is more robust and makes 55.2% fewer errors than other models. Whisper has a differing error rate with respect to transcribing different languages, with a higher word error rate in languages not well-represented in the training data. The authors found that multi-task learning improved overall performance compared to models specialized to one task. They conjectured that the best Whisper model trained is still underfitting the dataset, and larger models and longer training can result in better models. Third-party evaluations have found varying levels of AI hallucination. A study of transcripts of public meetings found hallucinations in eight out of every 10 transcripts, while an engineer discovered hallucinations in "about half" of 100 hours of transcriptions and a developer identified them in "nearly every one" of 26,000 transcripts. A study of 13,140 short audio segments (averaging 10 seconds) found 187 hallucinations (1.4%), 38% of which generated text that could be harmful because it inserted false references to things like race, non-existent medications, or violent events that were not in the audio. == Applications == The model has been used as the base for many applications, such as a unified model for speech recognition and more general sound recognition. Whisper has also been integrated into the workflow of biomedical research. In 2025, a study on Alzheimer's disease detection used the model to transcribe spontaneous speech recordings. The transcripts that were generated by the model were combined with LLM vector embeddings and traditional classifiers to help classify the patients' health. Another application is when OVALYTICS incorporated Whisper to transcribe YouTube videos and automate content moderation systems, which improved its detection of offensive content. The model has also been used in academic libraries and cultral heritage institutions to generate transcripts and captions for their digitized audiovisual collections. In a 2025 case study, Emory University Libraries found that Whisper reduced the labor used in transcription by around 30-35%, shifting work from text creation to text correction. However, human review is still necessary to make sure accuracy, formatting, and accessibility are all standard.

  Read more →
• 

  Residuated lattice

  In abstract algebra, a residuated lattice is an algebraic structure that is simultaneously a lattice x ≤ y and a monoid x•y that admits operations x\z and z/y, loosely analogous to division or implication, when x•y is viewed as multiplication or conjunction, respectively. Called respectively right and left residuals, these operations coincide when the monoid is commutative. The general concept was introduced by Morgan Ward and Robert P. Dilworth in 1939. Examples, some of which existed prior to the general concept, include Boolean algebras, Heyting algebras, residuated Boolean algebras, relation algebras, and MV-algebras. Residuated semilattices omit the meet operation ∧, for example Kleene algebras and action algebras. == Definition == In mathematics, a residuated lattice is an algebraic structure L = (L, ≤, •, I) such that (i) (L, ≤) is a lattice. (ii) (L, •, I) is a monoid. (iii) For all z there exists for every x a greatest y, and for every y a greatest x, such that x•y ≤ z (the residuation properties). In (iii), the "greatest y", being a function of z and x, is denoted x\z and called the right residual of z by x. Think of it as what remains of z on the right after "dividing" z on the left by x. Dually, the "greatest x" is denoted z/y and called the left residual of z by y. An equivalent, more formal statement of (iii) that uses these operations to name these greatest values is (iii)' for all x, y, z in L, y ≤ x\z ⇔ x•y ≤ z ⇔ x ≤ z/y. As suggested by the notation, the residuals are a form of quotient. More precisely, for a given x in L, the unary operations x• and x\ are respectively the lower and upper adjoints of a Galois connection on L, and dually for the two functions •y and /y. By the same reasoning that applies to any Galois connection, we have yet another definition of the residuals, namely, x•(x\y) ≤ y ≤ x\(x•y), and (y/x)•x ≤ y ≤ (y•x)/x, together with the requirement that x•y be monotone in x and y. (When axiomatized using (iii) or (iii)' monotonicity becomes a theorem and hence not required in the axiomatization.) These give a sense in which the functions x• and x\ are pseudoinverses or adjoints of each other, and likewise for •x and /x. This last definition is purely in terms of inequalities, noting that monotonicity can be axiomatized as x • y ≤ (x∨z) • y and similarly for the other operations and their arguments. Moreover, any inequality x ≤ y can be expressed equivalently as an equation, either x∧y = x or x∨y = y. This along with the equations axiomatizing lattices and monoids then yields a purely equational definition of residuated lattices, provided the requisite operations are adjoined to the signature (L, ≤, •, I) thereby expanding it to (L, ∧, ∨, •, I, /, \). When thus organized, residuated lattices form an equational class or variety, whose homomorphisms respect the residuals as well as the lattice and monoid operations. Note that distributivity x • (y ∨ z) = (x • y) ∨ (x • z) and x•0 = 0 are consequences of these axioms and so do not need to be made part of the definition. This necessary distributivity of • over ∨ does not in general entail distributivity of ∧ over ∨, that is, a residuated lattice need not be a distributive lattice. However distributivity of ∧ over ∨ is entailed when • and ∧ are the same operation, a special case of residuated lattices called a Heyting algebra. Alternative notations for x•y include x◦y, x;y (relation algebra), and x⊗y (linear logic). Alternatives for I include e and 1'. Alternative notations for the residuals are x → y for x\y and y ← x for y/x, suggested by the similarity between residuation and implication in logic, with the multiplication of the monoid understood as a form of conjunction that need not be commutative. When the monoid is commutative the two residuals coincide. When not commutative, the intuitive meaning of the monoid as conjunction and the residuals as implications can be understood as having a temporal quality: x•y means x and then y, x → y means had x (in the past) then y (now), and y ← x means if-ever x (in the future) then y (at that time), as illustrated by the natural language example at the end of the examples. == Examples == One of the original motivations for the study of residuated lattices was the lattice of (two-sided) ideals of a ring. Given a ring R, the ideals of R, denoted Id(R), forms a complete lattice with set intersection acting as the meet operation and "ideal addition" acting as the join operation. The monoid operation • is given by "ideal multiplication", and the element R of Id(R) acts as the identity for this operation. Given two ideals A and B in Id(R), the residuals are given by A / B := { r ∈ R ∣ r B ⊆ A } {\displaystyle A/B:=\{r\in R\mid rB\subseteq A\}} B ∖ A := { r ∈ R ∣ B r ⊆ A } {\displaystyle B\setminus A:=\{r\in R\mid Br\subseteq A\}} It is worth noting that {0}/B and B\{0} are respectively the left and right annihilators of B. This residuation is related to the conductor (or transporter) in commutative algebra written as (A:B)=A/B. One difference in usage is that B need not be an ideal of R: it may just be a subset. Boolean algebras and Heyting algebras are commutative residuated lattices in which x•y = x∧y (whence the unit I is the top element 1 of the algebra) and both residuals x\y and y/x are the same operation, namely implication x → y. The second example is quite general since Heyting algebras include all finite distributive lattices, as well as all chains or total orders, for example the unit interval [0,1] in the real line, or the integers and ± ∞ {\displaystyle \pm \infty } . The structure (Z, min, max, +, 0, −, −) (the integers with subtraction for both residuals) is a commutative residuated lattice such that the unit of the monoid is not the greatest element (indeed there is no least or greatest integer), and the multiplication of the monoid is not the meet operation of the lattice. In this example the inequalities are equalities because − (subtraction) is not merely the adjoint or pseudoinverse of + but the true inverse. Any totally ordered group under addition such as the rationals or the reals can be substituted for the integers in this example. The nonnegative portion of any of these examples is an example provided min and max are interchanged and − is replaced by monus, defined (in this case) so that x-y = 0 when x ≤ y and otherwise is ordinary subtraction. A more general class of examples is given by the Boolean algebra of all binary relations on a set X, namely the power set of X2, made a residuated lattice by taking the monoid multiplication • to be composition of relations and the monoid unit to be the identity relation I on X consisting of all pairs (x,x) for x in X. Given two relations R and S on X, the right residual R\S of S by R is the binary relation such that x(R\S)y holds just when for all z in X, zRx implies zSy (notice the connection with implication). The left residual is the mirror image of this: y(S/R)x holds just when for all z in X, xRz implies ySz. This can be illustrated with the binary relations < and > on {0,1} in which 0 < 1 and 1 > 0 are the only relationships that hold. Then x(>\<)y holds just when x = 1, while x()y holds just when y = 0, showing that residuation of < by > is different depending on whether we residuate on the right or the left. This difference is a consequence of the difference between <•> and >•<, where the only relationships that hold are 0(<•>)0 (since 0<1>0) and 1(>•<)1 (since 1>0<1). Had we chosen ≤ and ≥ instead of < and >, ≥\≤ and ≤/≥ would have been the same because ≤•≥ = ≥•≤, both of which always hold between all x and y (since x≤1≥y and x≥0≤y). The Boolean algebra 2Σ of all formal languages over an alphabet (set) Σ forms a residuated lattice whose monoid multiplication is language concatenation LM and whose monoid unit I is the language {ε} consisting of just the empty string ε. The right residual M\L consists of all words w over Σ such that Mw ⊆ L. The left residual L/M is the same with wM in place of Mw. The residuated lattice of all binary relations on X is finite just when X is finite, and commutative just when X has at most one element. When X is empty the algebra is the degenerate Boolean algebra in which 0 = 1 = I. The residuated lattice of all languages on Σ is commutative just when Σ has at most one letter. It is finite just when Σ is empty, consisting of the two languages 0 (the empty language {}) and the monoid unit I = {ε} = 1. The examples forming a Boolean algebra have special properties treated in the article on residuated Boolean algebras. == Residuated semilattice == A residuated semilattice is defined almost identically for residuated lattices, omitting just the meet operation ∧. Thus it is an algebraic structure L = (L, ∨, •, 1, /, \) satisfying all the residuated lattice equations as specified above except those containing an occurrence of the symbol ∧. The option of defining x ≤ y as x∧y = x is then not available, leaving on

  Read more →
• 

  Artificial intelligence engineering

  Artificial intelligence engineering (AI engineering) is a technical discipline that focuses on the design, development, and deployment of AI systems. AI engineering involves applying engineering principles and methodologies to create scalable, efficient, and reliable AI-based solutions. It merges aspects of data engineering and software engineering to create real-world applications in diverse domains such as healthcare, finance, autonomous systems, and industrial automation. == Terminology ambiguity == According to Chip Huyen's book AI Engineering: Building Applications with Foundation Models, the term AI engineering refers to the process of building applications that use foundation models, which are typically models developed by a small number of research laboratories and made available as a service. Huyen distinguishes this from machine learning (ML) engineering, which involves building and deploying models developed in-house. She notes that most practical AI systems combine both approaches. For example, a customer-support chatbot may use a generative model to produce responses while also incorporating locally built components such as request classifiers or scoring mechanisms to assess response quality. As a result, the terms AI engineering and ML engineering are often used together or interchangeably in practice. The distinction and broader usage of the term have been discussed in industry publications and interviews, where AI engineering has been described as an emerging discipline focused on productionizing applications built with foundation models. == Key components == AI engineering integrates a variety of technical domains and practices, all of which are essential to building scalable, reliable, and ethical AI systems. === Data engineering and infrastructure === Data serves as the cornerstone of AI systems, necessitating careful engineering to ensure premium quality, wide spread availability, and usability. AI engineers gather large, diverse datasets from multiple sources such as databases, APIs, and real-time streams. This data undergoes cleaning, normalization, and preprocessing, often facilitated by automated data pipelines that manage extraction, transformation, and loading (ETL) processes. Efficient storage solutions, such as SQL (or NoSQL) databases and data lakes, must be selected based on data characteristics and use cases. Security measures, including encryption and access controls, are critical for protecting sensitive information and ensuring compliance with regulations like GDPR. Scalability is essential, frequently involving cloud services and distributed computing frameworks to handle growing data volumes effectively. === Algorithm selection and optimization === Selecting the appropriate algorithm is crucial for the success of any AI system. Engineers evaluate the problem (which could be classification or regression, for example) to determine the most suitable machine learning algorithm, including deep learning paradigms. Once an algorithm is chosen, optimizing it through hyperparameter tuning is essential to enhance efficiency and accuracy. Techniques such as grid search or Bayesian optimization are employed, and engineers often utilize parallelization to expedite training processes, particularly for large models and datasets. For existing models, techniques like transfer learning can be applied to adapt pre-trained models for specific tasks, reducing the time and resources needed for training. === Deep learning engineering === Deep learning is particularly important for tasks involving large and complex datasets. Engineers design neural network architectures tailored to specific applications, such as convolutional neural networks for visual tasks or recurrent neural networks for sequence-based tasks. Transfer learning, where pre-trained models are fine-tuned for specific use cases, helps streamline development and often enhances performance. Optimization for deployment in resource-constrained environments, such as mobile devices, involves techniques like pruning and quantization to minimize model size while maintaining performance. Engineers also mitigate data imbalance through augmentation and synthetic data generation, ensuring robust model performance across various classes. === Natural language processing === Natural language processing (NLP) is a crucial component of AI engineering, focused on enabling machines to understand and generate human language. The process begins with text preprocessing to prepare data for machine learning models. Recent advancements, particularly transformer-based models like BERT and GPT, have greatly improved the ability to understand context in language. AI engineers work on various NLP tasks, including sentiment analysis, machine translation, and information extraction. These tasks require sophisticated models that utilize attention mechanisms to enhance accuracy. Applications range from virtual assistants and chatbots to more specialized tasks like named-entity recognition (NER) and Part of speech (POS) tagging. === Reasoning and decision-making systems === Developing systems capable of reasoning and decision-making is a significant aspect of AI engineering. Whether starting from scratch or building on existing frameworks, engineers create solutions that operate on data or logical rules. Symbolic AI employs formal logic and predefined rules for inference, while probabilistic reasoning techniques like Bayesian networks help address uncertainty. These models are essential for applications in dynamic environments, such as autonomous vehicles, where real-time decision-making is critical. === Security === Security is a critical consideration in AI engineering, particularly as AI systems become increasingly integrated into sensitive and mission-critical applications. AI engineers implement robust security measures to protect models from adversarial attacks, such as evasion and poisoning, which can compromise system integrity and performance. Techniques such as adversarial training, where models are exposed to malicious inputs during development, help harden systems against these attacks. Additionally, securing the data used to train AI models is of paramount importance. Encryption, secure data storage, and access control mechanisms are employed to safeguard sensitive information from unauthorized access and breaches. AI systems also require constant monitoring to detect and mitigate vulnerabilities that may arise post-deployment. In high-stakes environments like autonomous systems and healthcare, engineers incorporate redundancy and fail-safe mechanisms to ensure that AI models continue to function correctly in the presence of security threats. === Ethics and compliance === As AI systems increasingly influence societal aspects, ethics and compliance are vital components of AI engineering. Engineers design models to mitigate risks such as data poisoning and ensure that AI systems adhere to legal frameworks, such as data protection regulations like GDPR. Privacy-preserving techniques, including data anonymization and differential privacy, are employed to safeguard personal information and ensure compliance with international standards. Ethical considerations focus on reducing bias in AI systems, preventing discrimination based on race, gender, or other protected characteristics. By developing fair and accountable AI solutions, engineers contribute to the creation of technologies that are both technically sound and socially responsible. == Workload == An AI engineer's workload revolves around the AI system's life cycle, which is a complex, multi-stage process. This process may involve building models from scratch or using pre-existing models through transfer learning, depending on the project's requirements. Each approach presents unique challenges and influences the time, resources, and technical decisions involved. === Problem definition and requirements analysis === Regardless of whether a model is built from scratch or based on a pre-existing model, the work begins with a clear understanding of the problem. The engineer must define the scope, understand the business context, and identify specific AI objectives that align with strategic goals. This stage includes consulting with stakeholders to establish key performance indicators (KPIs) and operational requirements. When developing a model from scratch, the engineer must also decide which algorithms are most suitable for the task. Conversely, when using a pre-trained model, the workload shifts toward evaluating existing models and selecting the one most aligned with the task. The use of pre-trained models often allows for a more targeted focus on fine-tuning, as opposed to designing an entirely new model architecture. === Data acquisition and preparation === Data acquisition and preparation are critical stages regardless of the development method chosen, as the performance of any AI system relies heavily on high-quality, re

  Read more →
• 

  Computer security compromised by hardware failure

  Computer security compromised by hardware failure is a branch of computer security applied to hardware. The objective of computer security includes protection of information and property from theft, corruption, or natural disaster, while allowing the information and property to remain accessible and productive to its intended users. Such secret information could be retrieved by different ways. This article focus on the retrieval of data thanks to misused hardware or hardware failure. Hardware could be misused or exploited to get secret data. This article collects main types of attack that can lead to data theft. Computer security can be compromised by devices, such as keyboards, monitors or printers (thanks to electromagnetic or acoustic emanation for example) or by components of the computer, such as the memory, the network card or the processor (thanks to time or temperature analysis for example). == Devices == === Monitor === The monitor is the main device used to access data on a computer. It has been shown that monitors radiate or reflect data on their environment, potentially giving attackers access to information displayed on the monitor. ==== Electromagnetic emanations ==== Video display units radiate: narrowband harmonics of the digital clock signals; broadband harmonics of the various 'random' digital signals such as the video signal. Known as compromising emanations or TEMPEST radiation, a code word for a U.S. government programme aimed at attacking the problem, the electromagnetic broadcast of data has been a significant concern in sensitive computer applications. Eavesdroppers can reconstruct video screen content from radio frequency emanations. Each (radiated) harmonic of the video signal shows a remarkable resemblance to a broadcast TV signal. It is therefore possible to reconstruct the picture displayed on the video display unit from the radiated emission by means of a normal television receiver. If no preventive measures are taken, eavesdropping on a video display unit is possible at distances up to several hundreds of meters, using only a normal black-and-white TV receiver, a directional antenna and an antenna amplifier. It is even possible to pick up information from some types of video display units at a distance of over 1 kilometer. If more sophisticated receiving and decoding equipment is used, the maximum distance can be much greater. ==== Compromising reflections ==== What is displayed by the monitor is reflected on the environment. The time-varying diffuse reflections of the light emitted by a CRT monitor can be exploited to recover the original monitor image. This is an eavesdropping technique for spying at a distance on data that is displayed on an arbitrary computer screen, including the currently prevalent LCD monitors. The technique exploits reflections of the screen's optical emanations in various objects that one commonly finds close to the screen and uses those reflections to recover the original screen content. Such objects include eyeglasses, tea pots, spoons, plastic bottles, and even the eye of the user. This attack can be successfully mounted to spy on even small fonts using inexpensive, off-the-shelf equipment (less than 1500 dollars) from a distance of up to 10 meters. Relying on more expensive equipment allowed to conduct this attack from over 30 meters away, demonstrating that similar attacks are feasible from the other side of the street or from a close by building. Many objects that may be found at a usual workplace can be exploited to retrieve information on a computer's display by an outsider. Particularly good results were obtained from reflections in a user's eyeglasses or a tea pot located on the desk next to the screen. Reflections that stem from the eye of the user also provide good results. However, eyes are harder to spy on at a distance because they are fast-moving objects and require high exposure times. Using more expensive equipment with lower exposure times helps to remedy this problem. The reflections gathered from curved surfaces on close by objects indeed pose a substantial threat to the confidentiality of data displayed on the screen. Fully invalidating this threat without at the same time hiding the screen from the legitimate user seems difficult, without using curtains on the windows or similar forms of strong optical shielding. Most users, however, will not be aware of this risk and may not be willing to close the curtains on a nice day. The reflection of an object, a computer display, in a curved mirror creates a virtual image that is located behind the reflecting surface. For a flat mirror this virtual image has the same size and is located behind the mirror at the same distance as the original object. For curved mirrors, however, the situation is more complex. === Keyboard === ==== Electromagnetic emanations ==== Computer keyboards are often used to transmit confidential data such as passwords. Since they contain electronic components, keyboards emit electromagnetic waves. These emanations could reveal sensitive information such as keystrokes. Electromagnetic emanations have turned out to constitute a security threat to computer equipment. The figure below presents how a keystroke is retrieved and what material is necessary. The approach is to acquire the raw signal directly from the antenna and to process the entire captured electromagnetic spectrum. Thanks to this method, four different kinds of compromising electromagnetic emanations have been detected, generated by wired and wireless keyboards. These emissions lead to a full or a partial recovery of the keystrokes. The best practical attack fully recovered 95% of the keystrokes of a PS/2 keyboard at a distance up to 20 meters, even through walls. Because each keyboard has a specific fingerprint based on the clock frequency inconsistencies, it can determine the source keyboard of a compromising emanation, even if multiple keyboards from the same model are used at the same time. The four different kinds way of compromising electromagnetic emanations are described below. ===== The Falling Edge Transition Technique ===== When a key is pressed, released or held down, the keyboard sends a packet of information known as a scan code to the computer. The protocol used to transmit these scan codes is a bidirectional serial communication, based on four wires: Vcc (5 volts), ground, data and clock. Clock and data signals are identically generated. Hence, the compromising emanation detected is the combination of both signals. However, the edges of the data and the clock lines are not superposed. Thus, they can be easily separated to obtain independent signals. ===== The Generalized Transition Technique ===== The Falling Edge Transition attack is limited to a partial recovery of the keystrokes. This is a significant limitation. The GTT is a falling edge transition attack improved, which recover almost all keystrokes. Indeed, between two traces, there is exactly one data rising edge. If attackers are able to detect this transition, they can fully recover the keystrokes. ===== The Modulation Technique ===== Harmonics compromising electromagnetic emissions come from unintentional emanations such as radiations emitted by the clock, non-linear elements, crosstalk, ground pollution, etc. Determining theoretically the reasons of these compromising radiations is a very complex task. These harmonics correspond to a carrier of approximately 4 MHz which is very likely the internal clock of the micro-controller inside the keyboard. These harmonics are correlated with both clock and data signals, which describe modulated signals (in amplitude and frequency) and the full state of both clock and data signals. This means that the scan code can be completely recovered from these harmonics. ===== The Matrix Scan Technique ===== Keyboard manufacturers arrange the keys in a matrix. The keyboard controller, often an 8-bit processor, parses columns one-by-one and recovers the state of 8 keys at once. This matrix scan process can be described as 192 keys (some keys may not be used, for instance modern keyboards use 104/105 keys) arranged in 24 columns and 8 rows. These columns are continuously pulsed one-by-one for at least 3μs. Thus, these leads may act as an antenna and generate electromagnetic emanations. If an attacker is able to capture these emanations, he can easily recover the column of the pressed key. Even if this signal does not fully describe the pressed key, it still gives partial information on the transmitted scan code, i.e. the column number. Note that the matrix scan routine loops continuously. When no key is pressed, we still have a signal composed of multiple equidistant peaks. These emanations may be used to remotely detect the presence of powered computers. Concerning wireless keyboards, the wireless data burst transmission can be used as an electromagnetic trigger to detect exactly when a key is pressed, while the matrix s

  Read more →
• 

  Dartmouth workshop

  The Dartmouth Summer Research Project on Artificial Intelligence was a 1956 summer workshop widely considered to be the founding event of artificial intelligence as a field. The workshop has been referred to as "the Constitutional Convention of AI". The project's four organizers, Claude Shannon, John McCarthy, Nathaniel Rochester and Marvin Minsky, are considered some of the "founding fathers" of AI. However it was not the first conference devoted to what would now be described as the question of artificial intelligence: it postdated meetings such as the 1951 Paris cybernetics conference and the Macy meetings. The project lasted approximately six to eight weeks and consisted largely of brainstorming sessions. Eleven mathematicians and scientists originally planned to attend; not all of them attended, but more than ten others came for short times. == Background == In the early 1950s, there were various names for the field of "thinking machines": cybernetics, automata theory, and complex information processing. The variety of names suggests the variety of conceptual orientations. In 1955, John McCarthy, then a young Assistant Professor of Mathematics at Dartmouth College, decided to organize a group to clarify and develop ideas about thinking machines. He picked the name 'Artificial Intelligence' for the new field. He chose the name partly for its neutrality; avoiding a focus on narrow automata theory, and avoiding cybernetics which was heavily focused on analog feedback, as well as him potentially having to accept the assertive Norbert Wiener as guru or having to argue with him. In early 1955, McCarthy approached the Rockefeller Foundation to request funding for a summer seminar at Dartmouth for about 10 participants. In June, he and Claude Shannon, a founder of information theory then at Bell Labs, met with Robert Morison, Director of Biological and Medical Research to discuss the idea and possible funding, though Morison was unsure whether money would be made available for such a visionary project. On September 2, 1955, the project was formally proposed by McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. The proposal is credited with introducing the term 'artificial intelligence'. The Proposal states: We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer. The proposal goes on to discuss computers, natural language processing, neural networks, theory of computation, abstraction and creativity (these areas within the field of artificial intelligence are considered still relevant to the work of the field). On May 26, 1956, McCarthy notified Robert Morison of the planned 11 attendees: For the full period: 1) Dr. Marvin Minsky 2) Dr. Julian Bigelow 3) Professor D.M. Mackay 4) Mr. Ray Solomonoff 5) Mr. John Holland 6) Dr. John McCarthy For four weeks: 7) Dr. Claude Shannon 8) Mr. Nathaniel Rochester 9) Mr. Oliver Selfridge For the first two weeks: 10) Dr. Allen Newell 11) Professor Herbert Simon He noted, "we will concentrate on a problem of devising a way of programming a calculator to form concepts and to form generalizations. This of course is subject to change when the group gets together." The actual participants came at different times, mostly for much shorter times. Trenchard More replaced Rochester for three weeks and MacKay and Holland did not attend—but the project was set to begin. Around June 18, 1956, the earliest participants (perhaps only Ray Solomonoff, maybe with Tom Etter) arrived at the Dartmouth campus in Hanover, N.H., to join John McCarthy who already had an apartment there. Solomonoff and Minsky stayed at Professors' apartments, but most would stay at the Hanover Inn. == Dates == The Dartmouth Workshop is usually said to have run for six weeks. Ray Solomonoff's notes taken during the workshop, however, indicate that it ran for roughly eight weeks, from about June 18 to August 17. Solomonoff's notes start on June 22; June 28 mentions Minsky, June 30 mentions Hanover, N.H., July 1 mentions Tom Etter. On August 17, Solomonoff gave a final talk. == Participants == Initially, McCarthy lost his list of attendees. Instead, after the workshop, McCarthy sent Solomonoff a preliminary list of participants and visitors plus those interested in the subject. 47 people were listed. Solomonoff, however, made a list of participants in his notes of the summer project: Ray Solomonoff Marvin Minsky John McCarthy Claude Shannon Trenchard More Nat Rochester Oliver Selfridge Julian Bigelow W. Ross Ashby W.S. McCulloch Abraham Robinson Tom Etter John Nash David Sayre Arthur Samuel Kenneth R. Shoulders Shoulders' friend Alex Bernstein Herbert Simon Allen Newell Shannon attended Solomonoff's talk on July 10 and Bigelow gave a talk on August 15. Solomonoff doesn't mention Bernard Widrow, but in 1994 Widrow said that he and an unidentified colleague from the same lab in MIT had attended for one week. In the same interview Widrow recalled that "I think [Wesley] Clark and [Belmont] Farley were there from Lincoln Lab." Trenchard mentions R. Culver and Solomonoff mentions Bill Shutz. Herb Gelernter didn't attend, but was influenced later by what Rochester learned. In an article in IEEE Spectrum, Grace Solomonoff additionally identifies Peter Milner in a photo taken by Nathaniel Rochester in front of Dartmouth Hall. Ray Solomonoff, Marvin Minsky, and John McCarthy were the only three who stayed for the full time. Trenchard took attendance during two weeks of his three-week visit. From three to about eight people would attend the daily sessions. == Event and aftermath == They had the entire top floor of the Dartmouth Math Department to themselves, and most weekdays they would meet at the main math classroom where someone might lead a discussion focusing on his ideas, or more frequently, a general discussion would be held. It was not a directed group research project; discussions covered many topics, but several directions are considered to have been initiated or encouraged by the Workshop: the rise of symbolic methods, systems focused on limited domains (early expert systems), and deductive systems versus inductive systems. One participant, Arthur Samuel, said, "It was very interesting, very stimulating, very exciting". Ray Solomonoff kept notes giving his impression of the talks and the ideas from various discussions. === McCarthy's 1956 AI distribution list === This is the list in the "People Interested in the Artificial Intelligence Problem" document which McCarthy produced in 1956, partly in lieu of a list of attendees at the Dartmouth workshop. According to McCarthy the list was "being sent to the people on the list and a few others", and its purpose was "to let those on it know who is interested in receiving documents on the problem" of artificial intelligence. McCarthy also promised to deliver them a report on the Dartmouth conference, and to send an updated list soon afterwards. It includes people who did not attend the conference and does not include everyone who did attend it.

  Read more →
• 

  Perceptual computing

  Perceptual computing is an application of Zadeh's theory of computing with words on the field of assisting people to make subjective judgments. == Perceptual computer == The perceptual computer – Per-C – an instantiation of perceptual computing – has the architecture that is depicted in Fig. 1 [2]–[6]. It consists of three components: encoder, CWW engine and decoder. Perceptions – words – activate the Per-C and are the Per-C output (along with data); so, it is possible for a human to interact with the Per-C using just a vocabulary. A vocabulary is application (context) dependent, and must be large enough so that it lets the end-user interact with the Per-C in a user-friendly manner. The encoder transforms words into fuzzy sets (FSs) and leads to a codebook – words with their associated FS models. The outputs of the encoder activate a Computing With Words (CWW) engine, whose output is one or more other FSs, which are then mapped by the decoder into a recommendation (subjective judgment) with supporting data. The recommendation may be in the form of a word, group of similar words, rank or class. Although many details are needed in order to implement the Per-C's three components – encoder, decoder and CWW engine – and they are covered in [5], it is when the Per-C is applied to specific applications, that the focus on the methodology becomes clear. Stepping back from those details, the methodology of perceptual computing is: Focus on an application (A). Establish a vocabulary (or vocabularies) for A. Collect interval end-point data from a group of subjects (representative of the subjects who will use the Per-C) for all of the words in the vocabulary. Map the collected word data into word-FOUs by using the Interval Approach [1], [5, Ch. 3]. The result of doing this is the codebook (or codebooks) for A, and completes the design of the encoder of the Per-C. Choose an appropriate CWW engine for A. It will map IT2 FSs into one or more IT2 FSs. Examples of CWW engines are: IF-THEN rules [5, Ch. 6] and Linguistic Weighted Averages [6], [5, Ch. 5]. If an existing CWW engine is available for A, then use its available mathematics to compute its output(s). Otherwise, develop such mathematics for the new kind of CWW engine. The new CWW engine should be constrained so that its output(s) resemble the FOUs in the codebook(s) for A. Map the IT2 FS outputs from the CWW engine into a recommendation at the output of the decoder. If the recommendation is a word, rank or class, then use existing mathematics to accomplish this mapping [5, Ch. 4]. Otherwise, develop such mathematics for the new kind of decoder. == Applications of Per-C == To-date a Per-C has been implemented for the following four applications: (1) investment decision-making, (2) social judgment making, (3) distributed decision making, and (4) hierarchical and distributed decision-making. A specific example of the fourth application is the so-called Journal Publication Judgment Advisor [5, Ch. 10] in which for the first time only words are used at every level of the following hierarchical and distributed decision making process: n reviewers have to provide a subjective recommendation about a journal article that has been sent to them by the Associate Editor, who then has to aggregate the independent recommendations into a final recommendation that is sent to the Editor-in-Chief of the journal. Because it is very problematic to ask reviewers to provide numerical scores for paper-evaluation sub-categories (the two major categories are Technical Merit and Presentation), such as importance, content, depth, style, organization, clarity, references, etc., each reviewer will only be asked to provide a linguistic score for each of these categories. They will not be asked for an overall recommendation about the paper because in the past it is quite common for reviewers who provide the same numerical scores for such categories to give very different publishing recommendations. By leaving a specific recommendation to the associate editor such inconsistencies can hope to be eliminated. How words can be aggregated to reflect each reviewer's recommendation as well as the expertise of each reviewer about the paper's subject matter is done using a linguistic weighted average. Although the journal publication judgment advisor uses reviewers and an associate editor, the word “reviewer” could be replaced by judge, expert, low-level manager, commander, referee, etc., and the term “associate editor” could be replaced by control center, command center, higher-level manager, etc. So, this application has potential wide applicability to many other applications. Recently, a new Per-C based Failure mode and effects analysis (FMEA) methodology was developed, with its application to edible bird's nest farming, in Borneo, has been reported. In addition, application of Per-C based method to educational assessment, for cooperative learning of students has been reported. In summary, the Per-C (whose development has taken more than a decade) is the first complete implementation of Zadeh's CWW paradigm, as applied to assisting people to make subjective judgments.

  Read more →