AI Analytics Usf

AI Analytics Usf — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

Normal distributions transform

The normal distributions transform (NDT) is a point cloud registration algorithm introduced by Peter Biber and Wolfgang Straßer in 2003, while working at University of Tübingen. The algorithm registers two point clouds by first associating a piecewise normal distribution to the first point cloud, that gives the probability of sampling a point belonging to the cloud at a given spatial coordinate, and then finding a transform that maps the second point cloud to the first by maximising the likelihood of the second point cloud on such distribution as a function of the transform parameters. Originally introduced for 2D point cloud map matching in simultaneous localization and mapping (SLAM) and relative position tracking, the algorithm was extended to 3D point clouds and has wide applications in computer vision and robotics. NDT is very fast and accurate, making it suitable for application to large scale data, but it is also sensitive to initialisation, requiring a sufficiently accurate initial guess, and for this reason it is typically used in a coarse-to-fine alignment strategy. == Formulation == The NDT function associated to a point cloud is constructed by partitioning the space in regular cells. For each cell, it is possible to define the mean q = 1 n ∑ i x i {\displaystyle \textstyle \mathbf {q} ={\frac {1}{n}}\sum _{i}\mathbf {x_{i}} } and covariance S = 1 n ∑ i ( x i − q ) ( x i − q ) ⊤ {\displaystyle \textstyle \mathbf {S} ={\frac {1}{n}}\sum _{i}\left(\mathbf {x} _{i}-\mathbf {q} \right)\left(\mathbf {x} _{i}-\mathbf {q} \right)^{\top }} of the n {\displaystyle n} points of the cloud x 1 , … , x n {\displaystyle \mathbf {x} _{1},\dots ,\mathbf {x} _{n}} that fall within the cell. The probability density of sampling a point at a given spatial location x {\displaystyle \mathbf {x} } within the cell is then given by the normal distribution e − 1 2 ( x − q ) ⊤ S − 1 ( x − q ) {\displaystyle e^{-{\frac {1}{2}}\left(\mathbf {x} -\mathbf {q} \right)^{\top }\mathbf {S} ^{-1}\left(\mathbf {x} -\mathbf {q} \right)}} . Two point clouds can be mapped by a Euclidean transformation f {\displaystyle f} with rotation matrix R {\displaystyle \mathbf {R} } and translation vector t {\displaystyle \mathbf {t} } f R , t ( x ) = R x + t {\displaystyle f_{\mathbf {R} ,\mathbf {t} }(\mathbf {x} )=\mathbf {R} \mathbf {x} +\mathbf {t} } that maps from the second cloud to the first, parametrised by the rotation angles and translation components. The algorithm registers the two point clouds by optimising the parameters of the transformation that maps the second cloud to the first, with respect to a loss function based on the NDT of the first point cloud, solving the following problem arg ⁡ min R , t { − ∑ i NDT ⁡ ( f R , t ( x i ) ) } {\displaystyle \arg \min _{\mathbf {R} ,\mathbf {t} }\left\{-\sum _{i}\operatorname {NDT} \left(f_{\mathbf {R} ,\mathbf {t} }\left(\mathbf {x_{i}} \right)\right)\right\}} where the loss function represents the negated likelihood, obtained by applying the transformation to all points in the second cloud and summing the value of the NDT at each transformed point f R , t ( x ) {\displaystyle f_{\mathbf {R} ,\mathbf {t} }(\mathbf {x} )} . The loss is piecewise continuous and differentiable, and can be optimised with gradient-based methods (in the original formulation, the authors use Newton's method). In order to reduce the effect of cell discretisation, a technique consists of partitioning the space into multiple overlapping grids, shifted by half cell size along the spatial directions, and computing the likelihood at a given location as the sum of the NDTs induced by each grid.
Read more →
Behavior selection algorithm

In artificial intelligence, a behavior selection algorithm, or action selection algorithm, is an algorithm that selects appropriate behaviors or actions for one or more intelligent agents. In game artificial intelligence, it selects behaviors or actions for one or more non-player characters. Common behavior selection algorithms include: Finite-state machines Hierarchical finite-state machines Decision trees Behavior trees Hierarchical task networks Hierarchical control systems Utility systems Dialogue tree (for selecting what to say) == Related concepts == In application programming, run-time selection of the behavior of a specific method is referred to as the strategy design pattern.
Read more →
List of artificial intelligence projects

The following is a list of current and past, non-classified notable artificial intelligence projects. == Specialized projects == === Brain-inspired === Blue Brain Project, an attempt to create a synthetic brain by reverse-engineering the mammalian brain down to the molecular level. Google Brain, a deep learning project part of Google X attempting to have intelligence similar or equal to human-level. Human Brain Project, ten-year scientific research project, based on exascale supercomputers. === Cognitive architectures === 4CAPS, developed at Carnegie Mellon University under Marcel A. Just ACT-R, developed at Carnegie Mellon University under John R. Anderson. AIXI, Universal Artificial Intelligence developed by Marcus Hutter at IDSIA and ANU. CALO, a DARPA-funded, 25-institution effort to integrate many artificial intelligence approaches (natural language processing, speech recognition, machine vision, probabilistic logic, planning, reasoning, many forms of machine learning) into an AI assistant that learns to help manage your office environment. CHREST, developed under Fernand Gobet at Brunel University and Peter C. Lane at the University of Hertfordshire. CLARION, developed under Ron Sun at Rensselaer Polytechnic Institute and University of Missouri. CoJACK, an ACT-R inspired extension to the JACK multi-agent system that adds a cognitive architecture to the agents for eliciting more realistic (human-like) behaviors in virtual environments. Copycat, by Douglas Hofstadter and Melanie Mitchell at the Indiana University. DUAL, developed at the New Bulgarian University under Boicho Kokinov. FORR developed by Susan L. Epstein at The City University of New York. IDA and LIDA, implementing Global Workspace Theory, developed under Stan Franklin at the University of Memphis. OpenCog Prime, developed using the OpenCog Framework. Procedural Reasoning System (PRS), developed by Michael Georgeff and Amy L. Lansky at SRI International. Psi-Theory developed under Dietrich Dörner at the Otto-Friedrich University in Bamberg, Germany. Soar, developed under Allen Newell and John Laird at Carnegie Mellon University and the University of Michigan. Society of Mind and its successor The Emotion Machine proposed by Marvin Minsky. Subsumption architectures, developed e.g. by Rodney Brooks (though it could be argued whether they are cognitive). === Games === AlphaGo, software developed by Google that plays the Chinese board game Go. Chinook, a computer program that plays English draughts; the first to win the world champion title in the competition against humans. Deep Blue, a chess-playing computer developed by IBM which beat Garry Kasparov in 1997. Halite, an artificial intelligence programming competition created by Two Sigma in 2016. Libratus, a poker AI that beat world-class poker players in 2017, intended to be generalisable to other applications. The Matchbox Educable Noughts and Crosses Engine (sometimes called the Machine Educable Noughts and Crosses Engine or MENACE) was a mechanical computer made from 304 matchboxes designed and built by artificial intelligence researcher Donald Michie in 1961. Quick, Draw!, an online game developed by Google that challenges players to draw a picture of an object or idea and then uses a neural network to guess what the drawing is. The Samuel Checkers-playing Program (1959) was among the world's first successful self-learning programs, and as such a very early demonstration of the fundamental concept of artificial intelligence (AI). Stockfish AI, an open source chess engine currently ranked the highest in many computer chess rankings. TD-Gammon, a program that learned to play world-class backgammon partly by playing against itself (temporal difference learning with neural networks). === Internet activism === Serenata de Amor, project for the analysis of public expenditures and detect discrepancies. === Knowledge and reasoning === Alice (Microsoft), a project from Microsoft Research Lab aimed at improving decision-making in Economics Braina, an intelligent personal assistant application with a voice interface for Windows OS. Cyc, an attempt to assemble an ontology and database of everyday knowledge, enabling human-like reasoning. Eurisko, a language by Douglas Lenat for solving problems which consists of heuristics, including some for how to use and change its heuristics. Google Now, an intelligent personal assistant with a voice interface in Google's Android and Apple Inc.'s iOS, as well as Google Chrome web browser on personal computers. Holmes a new AI created by Wipro. Microsoft Cortana, an intelligent personal assistant with a voice interface in Microsoft's various Windows 10 editions. MindsDB, is an AI automation platform for building AI/ML powered features and applications. Mycin, an early medical expert system. Open Mind Common Sense, a project based at the MIT Media Lab to build a large common sense knowledge base from online contributions. Siri, an intelligent personal assistant and knowledge navigator with a voice-interface in Apple Inc.'s iOS and macOS. SNePS, simultaneously a logic-based, frame-based, and network-based knowledge representation, reasoning, and acting system. Viv (software), a new AI by the creators of Siri. Wolfram Alpha, an online service that answers queries by computing the answer from structured data. === Motion and manipulation === AIBO, the robot pet for the home, grew out of Sony's Computer Science Laboratory (CSL). Cog, a robot developed by MIT to study theories of cognitive science and artificial intelligence, now discontinued. === Music === Melomics, a bioinspired technology for music composition and synthesization of music, where computers develop their own style, rather than mimic musicians. === Natural language processing === AIML, an XML dialect for creating natural language software agents. Apache Lucene, a high-performance, full-featured text search engine library written entirely in Java. Apache OpenNLP, a machine learning based toolkit for the processing of natural language text. It supports the most common NLP tasks, such as tokenization, sentence segmentation, part-of-speech tagging, named entity extraction, chunking and parsing. Artificial Linguistic Internet Computer Entity (A.L.I.C.E.), a natural language processing chatterbot. ChatGPT, a chatbot built on top of OpenAI's GPT-3.5 and GPT-4 family of large language models. Claude, a family of large language models developed by Anthropic and launched in 2023. Claude LLMs achieved high coding scores in several recognized LLM benchmarks. Cleverbot, successor to Jabberwacky, now with 170m lines of conversation, Deep Context, fuzziness and parallel processing. Cleverbot learns from around 2 million user interactions per month. DeepSeek: Chinese chatbot funded by hedge fund High-Flyer. DBRX, 136 billion parameter open sourced large language model developed by Mosaic ML and Databricks. ELIZA, a famous 1966 computer program by Joseph Weizenbaum, which parodied person-centered therapy. FreeHAL, a self-learning conversation simulator (chatterbot) which uses semantic nets to organize its knowledge to imitate a very close human behavior within conversations. Gemini, a family of multimodal large language model developed by Google's DeepMind. Drives the Gemini chatbot, formerly known as Bard. GigaChat, a chatbot by Russian Sberbank. GPT-3, a 2020 language model developed by OpenAI that can produce text difficult to distinguish from that written by a human. Jabberwacky, a chatbot by Rollo Carpenter, aiming to simulate natural human chat. LaMDA, a family of conversational neural language models developed by Google. LLaMA, a 2023 language model family developed by Meta that includes 7, 13, 33 and 65 billion parameter models.[1] Mycroft, a free and open-source intelligent personal assistant that uses a natural language user interface. PARRY, another early chatterbot, written in 1972 by Kenneth Colby, attempting to simulate a paranoid schizophrenic. SHRDLU, an early natural language processing computer program developed by Terry Winograd at MIT from 1968 to 1970. SYSTRAN, a machine translation technology by the company of the same name, used by Yahoo!, AltaVista and Google, among others. === Speech recognition === CMU Sphinx, a group of speech recognition systems developed at Carnegie Mellon University. DeepSpeech, an open-source Speech-To-Text engine based on Baidu's deep speech research paper. Whisper, an open-source speech recognition system developed at OpenAI. === Speech synthesis === 15.ai, a real-time artificial intelligence text-to-speech tool developed by an anonymous researcher from MIT. Amazon Polly, a speech synthesis software by Amazon. Festival Speech Synthesis System, a general multi-lingual speech synthesis system developed at the Centre for Speech Technology Research (CSTR) at the University of Edinburgh. WaveNet, a deep neural network for generating raw audio. === Video === CapCut is a video editor tool, developed
Read more →
Umbrella review

In medical research, an umbrella review is a review of systematic reviews or meta-analyses. They may also be called overviews of reviews, reviews of reviews, summaries of systematic reviews, or syntheses of reviews. Umbrella reviews are among the highest levels of evidence currently available in medicine. By summarizing information from multiple overview articles, umbrella reviews make it easier to review the evidence and allow for comparison of results between each of the individual reviews. Umbrella reviews may address a broader question than a typical review, such as discussing multiple different treatment comparisons instead of only one. They are especially useful for developing guidelines and clinical practice, and when comparing competing interventions.
Read more →
Kuaishou

Kuaishou Technology is a Chinese publicly traded partly state-owned holding company based in Haidian District, Beijing, that was founded in 2011 by Hua Su (Chinese: 宿华) and Cheng Yixiao (Chinese: 程一笑). The company, listed on the Hong Kong Stock Exchange, is known for developing a mobile app for sharing users' short videos, a social network, and video special effects editor. The app is known as Kwai in many countries outside of China. It is also known as Snack Video in India, Pakistan and Indonesia. == Ownership and governance == Kuaishou's overseas team is led by the former CEO of the application 99, and staff from Google, Facebook, Netflix, and TikTok were recruited to lead the company's international expansion. The China Internet Investment Fund, a state-owned enterprise controlled by the Cyberspace Administration of China, holds a golden share ownership stake in Kuaishou. == History == Kuaishou is China's first short video platform that was developed in 2011 by engineer Hua Su and Cheng Yixiao. Prior to co-founding Kuaishou, Su Hua had worked for both Google and Baidu as a software engineer. The company is headquartered in Haidian District, Beijing. Kuaishou's predecessor "GIF Kuaishou" was founded in March 2011. GIF Kuaishou was a mobile app with which users could make and share GIF pictures. In 2013, Kuaishou became a short-video social platform. By 2013, the app had reached 100 million daily users. By 2019, it had exceeded 200 million active daily users. In March 2017, Kuaishou closed a US$350 million investment round that was led by Tencent. In January 2018, Forbes estimated the company's valuation to be US$18 billion. In April 2018, Kuaishou's app was briefly banned from Chinese app stores after China Central Television (CCTV) reported on the platform popularizing videos of teenage mothers. In 2019, the company announced a partnership with the People's Daily, an official newspaper of the Central Committee of the Chinese Communist Party, to help it experiment with the use of artificial intelligence in news. In June 2020, following the start of the 2020–2021 China–India skirmishes, the Government of India banned Kwai along with 58 other apps, citing "data and privacy issues". In January 2021, Kuaishou announced it was planning an initial public offering (IPO) to raise approximately US$5 billion. Kuaishou's stock completed its first day of trading at $300 Hong Kong dollars (HKD) (US$38.70), more than doubling its initial offer price, and causing its market value to rise to over $1 trillion HKD (US$159 billion). In February 2021, Kuaishou made a debut on the Hong Kong Stock Exchange, with its shares soaring by 194% at the opening. The company subsequently encountered major setbacks as a result of heightened regulatory restrictions on Chinese internet firms, which contributed to its share price falling by nearly 80% from its post-IPO peak. By December 2021, Kuaishou announced a major reorganization, including the layoff of 30% of its staff, primarily targeting mid-level employees earning an annual salary of $157,000 or more. This restructuring aimed to cut costs and mitigate financial losses. In October 2022, state-owned Beijing Radio and Television Station took a minority ownership stake in Kuaishou. In April 2024, a Financial Times article citing current and former Kuaishou employees stated that the company has been running an ageist redundancy programme known internally as "Limestone", culling workers in their mid-30s. In June 2024, Kuaishou and the Sichuan international communication center launched a branch center in São Paulo, Brazil. In June 2024, Kuaishou released its diffusion transformer text-to-video model, Kling, which they claimed could generate two minutes of video at 30 frames per second and in 1080p resolution. The model has been compared to that of OpenAI's Sora text-to-video model. It is accessible to the public on Kuaishou's video editing app KwaiCut via signing up for a waitlist with a Chinese phone number. In December 2025, Kuaishou came under a cyberattack which led to a temporary influx of violent and pornographic content. == Popularity == As of 2019, it had a worldwide user base of over 200 million, leading the "Most Downloaded" lists of the Google Play and Apple App Store in eight countries, such as Brazil, where it was introduced in 2019. Its main short-video platform competitor was Douyin, which is known as TikTok outside China. Compared to Douyin, Kuaishou is more popular with older users living outside China's Tier 1 cities. Its initial popularity came from videos of Chinese rural life. The app is particularly well known for its "rustic" aesthetic and is popular among rural people. Kuaishou also relied more on e-commerce revenue than on advertising revenue compared to its main competitor. == Reception == Kwai (as the app is called outside of China) was banned in India in 2020 along with other short video apps like TikTok. Kuaishou then released the clone SnackVideo, which was subsequently also banned. The app is one of the most popular social media platforms in Brazil, where Kuaishou partnered with creators to make telenovela style content, and appeals to football fans by working with football teams CR Flamengo and Santos FC and sponsoring the tournament Copa América. Kwai was notable in Brazil for spreading information (and misinformation) about the COVID-19 vaccine and political misinformation. === Manjiao Wenhua === "Manjiao wenhua" (慢脚文化) is a sarcasm term on Chinese internet on the unethical or illegal contents on Kuaishou. State broadcaster China Central Television (CCTV) reported that many contents are about child pregnancy. "Dating, pregnancy, bearing a child...these are strictly prohibited in the real time by a minor, but these contents can easily shown to audiences here." In addition, many students from primary or secondary schools make a pose of smoking. Wang Zhenhui (王贞会) from CUPSL stated that these kinds of bad values will give negative effects to the minors.
Read more →
Agentic commerce

Agentic commerce (also referred to as agent-based commerce) describes an emerging form of e-commerce in which autonomous artificial intelligence (AI) agents independently execute purchasing and payment processes on behalf of users or organizations. Unlike conventional digital commerce systems, which require direct human interaction at key decision points, agentic commerce systems are designed to search for products or services, evaluate options, make purchasing decisions, and complete payments without real-time human involvement. An emerging development within the broader fields of e-commerce, fintech, and artificial intelligence; agentic commerce combines advances in generative AI, autonomous agents, application programming interfaces (APIs), and digital payment infrastructures to direct transactions with no direct human interaction. == Characteristics == A defining feature of agentic commerce is the delegation of end-to-end commercial activities to software agents. These agents typically operate according to predefined user preferences, rules, or constraints, such as price limits, quality criteria, delivery times, or preferred payment methods. Based on these parameters, an agent can autonomously perform tasks including product discovery, price comparison, contract selection, order placement, and payment execution. In contrast to decision-support systems, which provide recommendations to human users, agentic commerce systems are designed to act independently. Human involvement may be limited to initial configuration, periodic supervision, or exception handling. == Comparison with traditional and AI-assisted commerce == Traditional e-commerce requires users to manually browse products, select offers, and authorize payments. Generative AI systems used in commerce commonly assist users by answering questions or suggesting options, and do not complete transactions autonomously. Agentic commerce differs in that decision-making authority is partially or fully transferred to AI agents. As a result, the conventional customer journey, characterized by conscious decision points, may be replaced by continuous, automated micro-decisions performed by software. == Applications and business use cases == Potential applications of agentic commerce include recurring purchases, subscription management, business-to-business procurement, inventory replenishment, and price monitoring. In such contexts, transactions are often predictable and standardized, making them suitable for automation. From a business perspective, agentic commerce systems may be used to optimize supply chains, manage inventory levels, negotiate prices algorithmically, or execute transactions across multiple platforms. Enterprises adopting the new technology include retailers Walmart, Home Depot, Wayfair and Urban Outfitters, and ad tech DSPs, including Google Ads, Amazon, and Yahoo. Chinese tech firms are using apps to provide full-service shopping and payment tools. These includes Alibaba, Tencent, and ByteDance who are currently developing AI powered shopping apps. The Qwen AI chatbot allows users to complete transactions directly within its interface. US firms are still leading in developing AI models but integration is slower due to privacy restrictions. == Payments and technical infrastructure == Agentic commerce relies on digital payment systems capable of supporting automated, machine-initiated transactions, including API-based payment processing, tokenization, real-time authorization, and continuous risk monitoring. Typical user interfaces, such as shopping carts, may be replaced by backend integrations between AI agents, merchants, and payment service providers. For example, Iike 2025, Alibaba launched Alipay AI Pay, which grew and began operating as an application for different retailers. In December 2025, Alipay teamed up with Rokid to enable developers to integrate AI payments into AI agents on Rokid's Lingzhu platform. In January 2025, Alipay unveiled the Agentic Commerce Trust Protocol in partnership with Alibaba's consumer AI applications, such as the Qwen App and Taobao Instant Commerce. Qwen adopted the platform first, connecting it to Taobao Instant Commerce and Alipay AI Pay. Users could use Qwen's agentic feature to place food and drink orders within the application instead of having to click outside to an external browser. For merchants, participation in agentic commerce may require products and services to be presented in structured, machine-readable formats to ensure discoverability and interoperability with autonomous agents. == Universal Commerce Protocol (UCP) == In January 2026, Google announced the Universal Commerce Protocol (UCP), an open-source web standard intended to enable interoperability between AI agents and retail systems across the shopping journey, from discovery and checkout to post-purchase support. UCP makes use of REST, JSON-RPC transports, and support for Agent Payments Protocol (AP2), Agent2Agent (A2A), and Model Context Protocol (MCP). == Legal, regulatory, and security considerations == The use of autonomous agents in commerce raises legal and regulatory questions, particularly regarding authorization, liability, consumer protection, and fraud prevention. Existing payment and contract frameworks are generally based on human decision-makers, and their applicability to autonomous agents remains an area of active discussion. Open issues include responsibility for unauthorized or erroneous transactions, mechanisms for dispute resolution, standards for agent authentication, and compliance with data protection and financial regulations. Continuous, automated transaction patterns may also require new approaches to security and risk assessment. Traditional fraud models centered on identity verification may be insufficient for agentic commerce, and that merchants may need intent-based detection methods using machine learning and behavioral analysis to distinguish legitimate AI agents from malicious automation. === Governance frameworks === The deployment of autonomous AI agents in commercial environments has prompted the development of dedicated governance frameworks. These aim to define operational boundaries, decision authority, oversight mechanisms, and accountability structures for agentic systems. The Agentic Commerce Framework (ACF), created in 2025 by Vincent Dorange, is a governance standard that structures the deployment of autonomous AI agents around four founding principles (Decision Sovereignty, Governance by Design, Ultimate Human Control, Traceable Accountability), four operational layers, and 18 governance KPIs. In January 2026, Singapore's Infocomm Media Development Authority (IMDA) published the Model AI Governance Framework for Agentic AI, extending its existing AI governance guidelines to address agent-specific risks including delegation chains and multi-agent coordination. The Cloud Security Alliance (CSA) has also proposed an Agentic Trust Framework applying zero-trust principles to AI agent governance. == Ecosystem and implementation == The adoption of agentic commerce typically requires changes in commerce architecture, data modeling, identity and permissions, and API-based orchestration of checkout and post-purchase workflows. Management consultancies have identified agentic commerce as a structural evolution of digital commerce, emphasizing the role of AI-driven agents in automating discovery, decision-making, and transaction processes across commerce systems. McKinsey & Company has described agentic commerce as a significant shift in how consumers interact with brands and how enterprises design their commerce operating models. In Europe, this ecosystem also includes digital commerce consultancies specializing in the adoption of agentic commerce. Consulting firms such as Horrea support brands in understanding and implementing the technological and organizational shifts associated with agentic commerce. == Market development and outlook == Agentic commerce is generally regarded as an early-stage development. Industry analysts have projected that AI-driven agents could account for a small but growing share of digital payment transactions within the coming years. Due to the scale of global digital commerce, even limited adoption could represent substantial transaction volumes. Analysts expect that by 2029, AI agents could handle between 1% and 4% of all digital payment transactions. With a projected total transaction volume of over $36 trillion a year, even a small share translates into a market worth up to $1.47 trillion. According to a McKinsey study from October 2025, agentic commerce projects that by 2030, the U.S. business-to-consumer retail market alone could see up to $1 trillion in revenue orchestrated through agentic commerce. On a global scale, the opportunity could range from $3 trillion to $5 trillion. Early experiments and pilot projects have demonstrated both the potential and current limitations of the
Read more →
Storage area network

A storage area network (SAN) or storage network is a computer network which provides access to consolidated, block-level data storage. SANs are primarily used to access data storage devices, such as disk arrays and tape libraries from servers so that the devices appear to the operating system as direct-attached storage. A SAN typically is a dedicated network of storage devices not accessible through the local area network (LAN). Although a SAN provides only block-level access, file systems built on top of SANs do provide file-level access and are known as shared-disk file systems. Newer SAN configurations enable hybrid SAN and allow traditional block storage that appears as local storage but also object storage for web services through APIs. == Storage architectures == Storage area networks (SANs) are sometimes referred to as network behind the servers and historically developed out of a centralized data storage model, but with its own data network. A SAN is, at its simplest, a dedicated network for data storage. In addition to storing data, SANs allow for the automatic backup of data, and the monitoring of the storage as well as the backup process. A SAN is a combination of hardware and software. It grew out of data-centric mainframe architectures, where clients in a network can connect to several servers that store different types of data. To scale storage capacities as the volumes of data grew, direct-attached storage (DAS) was developed, where disk arrays or just a bunch of disks (JBODs) were attached to servers. In this architecture, storage devices can be added to increase storage capacity. However, the server through which the storage devices are accessed is a single point of failure, and a large part of the LAN network bandwidth is used for accessing, storing and backing up data. To solve the single point of failure issue, a direct-attached shared storage architecture was implemented, where several servers could access the same storage device. DAS was the first network storage system and is still widely used where data storage requirements are not very high. Out of it developed the network-attached storage (NAS) architecture, where one or more dedicated file server or storage devices are made available in a LAN. Therefore, the transfer of data, particularly for backup, still takes place over the existing LAN. If more than a terabyte of data was stored at any one time, LAN bandwidth became a bottleneck. Therefore, SANs were developed, where a dedicated storage network was attached to the LAN, and terabytes of data are transferred over a dedicated high speed and bandwidth network. Within the SAN, storage devices are interconnected. Transfer of data between storage devices, such as for backup, happens behind the servers and is meant to be transparent. In a NAS architecture data is transferred using the TCP and IP protocols over Ethernet. Distinct protocols were developed for SANs, such as Fibre Channel, iSCSI, Infiniband. Therefore, SANs often have their own network and storage devices, which have to be bought, installed, and configured. This makes SANs inherently more expensive than NAS architectures. == Components == SANs have their own networking devices, such as SAN switches. To access the SAN, so-called SAN servers are used, which in turn connect to SAN host adapters. Within the SAN, a range of data storage devices may be interconnected, such as SAN-capable disk arrays, JBODs and tape libraries. === Host layer === Servers that allow access to the SAN and its storage devices are said to form the host layer of the SAN. Such servers have host adapters, which are cards that attach to slots on the server motherboard (usually PCI slots) and run with a corresponding firmware and device driver. Through the host adapters the operating system of the server can communicate with the storage devices in the SAN. In Fibre channel deployments, a cable connects to the host adapter through the gigabit interface converter (GBIC). GBICs are also used on switches and storage devices within the SAN, and they convert digital bits into light impulses that can then be transmitted over the Fibre Channel cables. Conversely, the GBIC converts incoming light impulses back into digital bits. The predecessor of the GBIC was called gigabit link module (GLM). === Fabric layer === The fabric layer consists of SAN networking devices that include SAN switches, routers, protocol bridges, gateway devices, and cables. SAN network devices move data within the SAN, or between an initiator, such as an HBA port of a server, and a target, such as the port of a storage device. When SANs were first built, hubs were the only devices that were Fibre Channel capable, but Fibre Channel switches were developed and hubs are now rarely found in SANs. Switches have the advantage over hubs that they allow all attached devices to communicate simultaneously, as a switch provides a dedicated link to connect all its ports with one another. When SANs were first built, Fibre Channel had to be implemented over copper cables, these days multimode optical fibre cables are used in SANs. SANs are usually built with redundancy, so SAN switches are connected with redundant links. SAN switches connect the servers with the storage devices and are typically non-blocking allowing transmission of data across all attached wires at the same time. SAN switches are for redundancy purposes set up in a meshed topology. A single SAN switch can have as few as 8 ports and up to 32 ports with modular extensions. So-called director-class switches can have as many as 128 ports. In switched SANs, the Fibre Channel switched fabric protocol FC-SW-6 is used under which every device in the SAN has a hardcoded World Wide Name (WWN) address in the host bus adapter (HBA). If a device is connected to the SAN its WWN is registered in the SAN switch name server. In place of a WWN, or worldwide port name (WWPN), SAN Fibre Channel storage device vendors may also hardcode a worldwide node name (WWNN). The ports of storage devices often have a WWN starting with 5, while the bus adapters of servers start with 10 or 21. === Storage layer === The serialized Small Computer Systems Interface (SCSI) protocol is often used on top of the Fibre Channel switched fabric protocol in servers and SAN storage devices. The Internet Small Computer Systems Interface (iSCSI) over Ethernet and the Infiniband protocols may also be found implemented in SANs, but are often bridged into the Fibre Channel SAN. However, Infiniband and iSCSI storage devices, in particular, disk arrays, are available. The various storage devices in a SAN are said to form the storage layer. It can include a variety of hard disk and magnetic tape devices that store data. In SANs, disk arrays are joined through a RAID which makes a lot of hard disks look and perform like one big storage device. Every storage device, or even partition on that storage device, has a logical unit number (LUN) assigned to it. This is a unique number within the SAN. Every node in the SAN, be it a server or another storage device, can access the storage by referencing the LUN. The LUNs allow for the storage capacity of a SAN to be segmented and for the implementation of access controls. A particular server, or a group of servers, may, for example, be only given access to a particular part of the SAN storage layer, in the form of LUNs. When a storage device receives a request to read or write data, it will check its access list to establish whether the node, identified by its LUN, is allowed to access the storage area, also identified by a LUN. LUN masking is a technique whereby the host bus adapter and the SAN software of a server restrict the LUNs for which commands are accepted. In doing so LUNs that should never be accessed by the server are masked. Another method to restrict server access to particular SAN storage devices is fabric-based access control, or zoning, which is enforced by the SAN networking devices and servers. Under zoning, server access is restricted to storage devices that are in a particular SAN zone. == Network protocols == A mapping layer to other protocols is used to form a network: ATA over Ethernet (AoE), mapping of AT Attachment (ATA) over Ethernet Fibre Channel Protocol (FCP), a mapping of SCSI over Fibre Channel Fibre Channel over Ethernet (FCoE) ESCON over Fibre Channel (FICON), used by mainframe computers HyperSCSI, mapping of SCSI over Ethernet iFCP or SANoIP mapping of FCP over IP iSCSI, mapping of SCSI over TCP/IP iSCSI Extensions for RDMA (iSER), mapping of iSCSI over InfiniBand Network block device, mapping device node requests on UNIX-like systems over stream sockets like TCP/IP SCSI RDMA Protocol (SRP), another SCSI implementation for remote direct memory access (RDMA) transports Storage networks may also be built using Serial Attached SCSI (SAS) and Serial ATA (SATA) technologies. SAS evolved from SCSI direct-attached storage. SATA evolved from Para
Read more →
Information behavior

Information behavior is a field of information science research that seeks to understand the way people search for and use information in various contexts. It can include information seeking and information retrieval, but it also aims to understand why people seek information and how they use it. The term 'information behavior' was coined by Thomas D. Wilson in 1982 and sparked controversy upon its introduction. The term has now been adopted and Wilson's model of information behavior is widely cited in information behavior literature. In 2000, Wilson defined information behavior as "the totality of human behavior in relation to sources and channels of information". A variety of theories of information behavior seek to understand the processes that surround information seeking. An analysis of the most cited publications on information behavior during the early 21st century shows its theoretical nature. Information behavior research can employ various research methodologies grounded in broader research paradigms from psychology, sociology and education. In 2003, a framework for information-seeking studies was introduced that aims to guide the production of clear, structured descriptions of research objects and positions information-seeking as a concept within information behavior. == Concepts of information behavior == === Information need === Information need is a concept introduced by Wilson. Understanding the information need of an individual involved three elements: Why the individual decides to look for information, What purpose the information they find will serve, and How the information is used once it is retrieved === Information-seeking behavior === Information-seeking behavior is a more specific concept of information behavior. It specifically focuses on searching, finding, and retrieving information. Information-seeking behavior research can focus on improving information systems or, if it includes information need, can also focus on why the user behaves the way they do. A review study on information search behavior of users highlighted that behavioral factors, personal factors, product/service factors and situational factors affect information search behavior. Information-seeking behavior can be more or less explicit on the part of users: users might seek to solve some task or to establish some piece of knowledge which can be found in the data in question, or alternatively the search process itself is part of the objective of the user, in use cases for exploring visual content or for familiarising oneself with the content of an information service. In the general case, information-seeking needs to be understood and analysed as a session rather than as a one-off transaction with a search engine, and in a broader context which includes user high-level intentions in addition to the immediate information need. === Information use === An information need is the recognition that a gap exists in one’s knowledge, prompting a desire to seek information to fill that gap. It often arises when a person encounters a problem or question they cannot resolve with their current understanding. === Information poverty and barriers === Introduced by Elfreda Chatman in 1987, information poverty is informed by the understanding that information is not equally accessible to all people. Information poverty does not describe a lack of information, but rather a worldview in which one's own experiences inside their own small world may create a distrust in the information provided by those outside their own lived experiences. == Metatheories == In Library and Information Science (LIS), a metatheory is described "a set of assumptions that orient and direct theorizing about a given phenomenon". Library and information science researchers have adopted a number of different metatheories in their research. A common concern among LIS researchers, and a prominent discussion in the field, is the broad spectrum of theories that inform the study of information behavior, information users, or information use. This variation has been noted as a cause of concern because it makes individual studies difficult to compare or synthesize if they are not guided by the same theory. This sentiment has been expressed in studies of information behavior literature from the early 1980s and more recent literature reviews have declared it necessary to refine their reviews to specific contexts or situations due to the sheer breadth of information behavior research available. Below are descriptions of some, but not all, metatheories that have guided LIS research. === Cognitivist approach === A cognitive approach to understanding information behavior is grounded in psychology. It holds the assumption that a person's thinking influences how they seek, retrieve, and use information. Researchers that approach information behavior with the assumption that it is influenced by cognition, seek to understand what someone is thinking while they engage in information behavior and how those thoughts influence their behavior. Wilson's attempt to understand information-seeking behavior by defining information need includes a cognitive approach. Wilson theorizes that information behavior is influenced by the cognitive need of an individual. By understanding the cognitive information need of an individual, we may gain insight into their information behavior. Nigel Ford takes a cognitive approach to information-seeking, focusing on the intellectual processes of information-seeking. In 2004, Ford proposed an information-seeking model using a cognitive approach that focuses on how to improve information retrieval systems and serves to establish information-seeking and information behavior as concepts in and of themselves, rather than synonymous terms. === Constructionist approach === The constructionist approach to information behavior has roots in the humanities and social sciences. It relies on social constructionism, which assumes that a person's information behavior is influenced by their experiences in society. In order to understand information behavior, constructionist researchers must first understand the social discourse that surrounds the behavior. The most popular thinker referenced in constructionist information behavior research is Michel Foucault, who famously rejected the concept of a universal human nature. The constructionist approach to information behavior research creates space for contextualizing the behavior based on the social experiences of the individual. One study that approaches information behavior research through the social constructionist approach is a study of the information behavior of a public library knitting group. The authors use a collectivist theory to frame their research, which denies the universality of information behavior and focuses on "understanding the ways that discourse communities collectively construct information needs, seeking, sources, and uses". === Constructivist approach === The constructivist approach is born out of education and sociology in which, "individuals are seen as actively constructing an understanding of their worlds, heavily influenced by the social world(s) in which they are operating". Constructivist approaches to information behavior research generally treat the individual's reality as constructed within their own mind rather than built by the society in which they live. The constructivist metatheory makes space for the influence of society and culture with social constructivism, "which argues that, while the mind constructs reality in its relationship to the world, this mental process is significantly informed by influences received from societal conventions, history and interaction with significant others". == Theories == A common concern among LIS researchers, and a prominent discussion in the field, is the broad spectrum of theories that inform LIS research. This variation has been noted as a cause of concern because it makes individual studies difficult to compare if they are not guided by the same theory. Recent studies have shown that the impact of these theories and theoretical models is very limited. LIS researchers have applied concepts and theories from many disciplines, including sociology, psychology, communication, organizational behavior, and computer science. === Wilson's theory of information behavior (1981) === The term was coined by Thomas D. Wilson in his 1981 paper, on the grounds that the current term, 'information needs' was unhelpful since 'need' could not be directly observed, while how people behaved in seeking information could be observed and investigated. However, there is increasing work in the information-searching field that is relating behaviors to underlying needs. In 2000, Wilson described information behavior as the totality of human behavior in relation to sources and channels of information, including both active and passive information-seeking, and information use. He described info
Read more →
Pandas (software)

Pandas (styled as pandas) is a software library written for the Python programming language for data manipulation and analysis. In particular, it offers data structures and operations for manipulating numerical tables and time series. It is free software released under the three-clause BSD license. The name is derived from the term "panel data", an econometrics term for data sets that include observations over multiple time periods for the same individuals, as well as a play on the phrase "Python data analysis". Wes McKinney started building what would become Pandas at AQR Capital while he was a researcher there from 2007 to 2010. The development of Pandas introduced into Python many comparable features of working with DataFrames that were established in the R programming language. The library is built upon another library, NumPy. == History == Developer Wes McKinney started working on Pandas in 2008 while at AQR Capital Management out of the need for a high performance, flexible tool to perform quantitative analysis on financial data. Before leaving AQR, he was able to convince management to allow him to open source the library in 2009. Another AQR employee, Chang She, joined the effort in 2012 as the second major contributor to the library. In 2015, Pandas signed on as a fiscally sponsored project of NumFOCUS, a 501(c)(3) nonprofit charity in the United States. == Data model == Pandas is built around data structures called Series and DataFrames. Data for these collections can be imported from various file formats such as comma-separated values, JSON, Parquet, SQL database tables or queries, and Microsoft Excel. === Series === A Series is a one-dimensional array-like object that stores a sequence of values together with an associated set of labels, called an index. It is built on top of NumPy's array and affords many similar functionalities, but instead of using implicit integer positions, a Series allows explicit index labels of many data types. A Series can be created from Python lists, dictionaries, or NumPy arrays. If no index is provided, pandas automatically assigns a default integer index ranging from 0 to n-1, where n is the number of items in the Series. A simple example with customized labels is: To access a value or list of values from a Series, use its index or list of indices: Series can be used arithmetically, as in the statement series_3 = series_1 + series_2. This will align data points with corresponding index values in series_1 and series_2 (similar to a join in relational algebra), then add them together to produce new values in series_3. A Series has various attributes, such as name (Series name), dtype (data type of values), shape (number of rows), values, and index. They can be used in many of the same operations as NumPy arrays, with additional methods for reindexing, label-based selection, and handling missing data. === DataFrame === A DataFrame is a two-dimensional, tabular data structure with labeled rows and columns. Each column is stored internally as a Series and may hold a different data type (numeric, string, boolean, etc.). DataFrames can be created by a variety of means, including dictionaries of lists, NumPy arrays, and external files such as CSV or Excel spreadsheets: To retrieve a DataFrame column as a Series, use either 1) the index (dict-like notation) or 2) the name of column if the name is a valid Python identifier (attribute-like access). DataFrames support operations such as column assignment, row and column deletion, label-based indexing with loc, position-based indexing with iloc, reshaping, grouping, and joining. Merge operations implement a subset of relational algebra and allow one-to-one, many-to-one, and many-to-many joins. Some common attributes of a DataFrame include dtypes (data type of each column), shape (dimensions of the DataFrame returned as a tuple with form (number of rows, number of columns)), index/columns (labels of the DataFrame's rows/columns, respectively, returned as an Index object), values (data in the DataFrame returned as a 2D array), and empty (returns True if the DataFrame is empty). === Index === Index objects hold metadata for Series and Dataframe objects, such as axis labels and names, and are automatically created from input data. By default, a pandas index is a series of integers ascending from 0, similar to the indices of Python arrays. However, indices can also use any NumPy data type, including floating point, timestamps, or strings. Indices are also immutable, which allows them to be safely shared across multiple objects. pandas' syntax for mapping index values to relevant data is the same syntax Python uses to map dictionary keys to values. For example, if s is a Series, s['a'] will return the data point at index a. Unlike dictionary keys, index values are not guaranteed to be unique. If a Series uses the index value a for multiple data points, then s['a'] will instead return a new Series containing all matching values. A DataFrame's column names are stored and implemented identically to an index. As such, a DataFrame can be thought of as having two indices: one column-based and one row-based. Because column names are stored as an index, these are not required to be unique. If data is a Series, then data['a'] returns all values with the index value of a. However, if data is a DataFrame, then data['a'] returns all values in the column(s) named a. To avoid this ambiguity, Pandas supports the syntax data.loc['a'] as an alternative way to filter using the index. Pandas also supports the syntax data.iloc[n], which always takes an integer n and returns the nth value, counting from 0. This allows a user to act as though the index is an array-like sequence of integers, regardless of how it is actually defined. pandas also supports hierarchical indices with multiple values per data point through the "MultiIndex" class. MultiIndex objects allow a single DataFrame to represent multiple dimensions, similar to a pivot table in Microsoft Excel, where each level can optionally carry its own unique name. In practice, data with more than 2 dimensions is often represented using DataFrames with hierarchical indices, instead of the higher-dimension Panel and Panel4D data structures. == Functionality == pandas supports a variety of indexing and subsetting techniques, allowing data to be selected by label, index, or Boolean conditions. For example, df[df['col1'] > 5] will return all rows in the DataFrame df for which the value of the column col1 exceeds 5. The library also implements grouping operations based on the split-apply-combine approach, enabling users to aggregate, transform, or restructure data according to column values or functions applied to index labels. For example, df['col1'].groupby(df['col2']) groups the data in 'col1' by their values in 'col2', while df.groupby(lambda i: i % 2) groups all data in the whole DataFrame by whether their index is even. The library also provides extensive tools for transforming, filtering and summarizing data. Users may apply arbitrary functions to Series and DataFrames, and because the library is built on top of Numpy, most NumPy functions can be applied directly to pandas objects as well. The library also includes built-in operations for arithmetic operations, string processing, and descriptive statistics such as mean, median, and standard deviation. These built-in functions are designed to handle missing data, usually represented by the floating-point value NaN. In addition, pandas includes tools for reorganizing data into different structural formats, with methods that can reshape tabular data between "wide" and "long" formats and pivot values based on column labels. pandas also implements a flexible set of relational operations for combining datasets. For instance, merge() links row in DataFrames based on one or more shared keys or indices, supporting one-to-one, one-to-many, and many-to-many relationships in a manner analogous to join operations in relational databases like SQL. DataFrames can also be concatenated or stacked together along an axis through the concat() method, and overlapping data can be further spliced together using combine_first() to fill in missing values. Furthermore, the library includes specialized support for working with time-series data. Features include the ability to interpolate values and filter using a range of timestamps, such as data['1/1/2023':'2/2/2023'] , which will return all dates between January 1 and February 2. Missing values in time-series data are represented by a dedicated NaT (Not a Timestamp) object, instead of the NaN value it uses elsewhere. == Criticisms == Pandas has been criticized for its inefficiency. The entire dataset must be loaded in RAM, and the library does not optimize query plans or support parallel computing across multiple cores. Wes McKinney, the creator of Pandas, has recommended Apache Arrow as an alternative to address these performance concerns and ot
Read more →
Read–write conflict

In computer science, in the field of databases, read–write conflict, also known as unrepeatable reads, is a computational anomaly associated with interleaved execution of transactions. Specifically, a read–write conflict occurs when a "transaction requests to read an entity for which an unclosed transaction has already made a write request." Given a schedule S S = [ T 1 T 2 R ( A ) R ( A ) W ( A ) C o m . R ( A ) W ( A ) C o m . ] {\displaystyle S={\begin{bmatrix}T1&T2\\R(A)&\\&R(A)\\&W(A)\\&Com.\\R(A)&\\W(A)&\\Com.&\end{bmatrix}}} In this example, T1 has read the original value of A, and is waiting for T2 to finish. T2 also reads the original value of A, overwrites A, and commits. However, when T1 reads from A, it discovers two different versions of A, and T1 would be forced to abort, because T1 would not know what to do. This is an unrepeatable read. This could never occur in a serial schedule, in which each transaction executes in its entirety before another begins. Strict two-phase locking (Strict 2PL) or Serializable Snapshot Isolation (SSI) prevent this conflict. == Real-world example == Alice and Bob are using a website to book tickets for a specific show. Only one ticket is left for the specific show. Alice signs on first to see that only one ticket is left, and finds it expensive. Alice takes time to decide. Bob signs on and also finds one ticket left, and orders it instantly. Bob purchases and logs off. Alice decides to buy a ticket, to find there are no tickets. This is a typical read–write conflict situation.
Read more →
Visual Peer Review

== Development and history == Visual Peer Review was first described in a 2017 classroom study by Friedman and Rosen, which examined how students evaluate peer-produced data visualizations using structured rubrics. Developed within the broader fields of data visualization, information visualization, and educational technology, the system emphasized clear labeling, visual integrity, and reduction of chartjunk. Students assigned rubric scores and provided written explanations, aligning the activity with established principles of peer review. Follow-up research expanded both the methodological and analytic dimensions of the framework. Friedman and colleagues applied natural language processing (NLP) to peer-review text to analyze part-of-speech patterns, sentence complexity, and comment length. These analyses offered insight into how students expressed critique and engaged with core design principles. Later studies incorporated advanced statistical modeling to evaluate system-level behavior, including peer review networks and reviewer typologies. Between 2021 and 2024, the framework underwent iterative refinement through a series of studies that explored interface design, behavioral nudges, reviewer engagement, and social network dynamics. The system was influenced by earlier work in computer-supported peer review—particularly My Reviewers, a rubric-based writing assessment platform developed by Joe Moxley at the University of South Florida. While Moxley's platform focused on text-based feedback, Visual Peer Review adapted its core structure to support critique of DataVis and visual analytics. To guide structured analysis and feedback, Friedman and Rosen also drew on the “what, why, and how” framework introduced by Liu and Stasko (2010), which emphasizes understanding a visualization's purpose, task alignment, and encoding strategy. == Framework and components == Visual Peer Review is designed to support critique, reflection, and learning in courses focusing on data visualization, visual analytics, and related fields in educational technology. The system consists of interconnected component. Core components include: Visual Artifacts: Students generate original visualizations using software such as R (e.g., ggplot2), Tableau, Python, or Adobe Illustrator. These artifacts may include statistical graphics, dashboards, or design-oriented infographics. Rubric-Based Assessment: Peer reviewers evaluate submitted visualizations using structured rubrics grounded in visualization theory and design heuristics. Rubric dimensions typically include: Use of labeling and axis scales Minimalization of chartjunk and clutter (following Tufte's principles) Optimization of the data–ink ratio Preservation of visual integrity through accurate representation (lie factor) Written Peer Comments: In addition to scoring, reviewers provide narrative feedback explaining their reasoning. These comments aim to improve design literacy, strengthen visual reasoning, and support the learning process common to peer review across educational contexts. Instructor Analytics Dashboard: Instructors access an analytics dashboard that displays peer-review activity across the course. Metrics include comment length, rubric coverage, participation patterns, and potential indicators of disengagement. These features position the framework within the domain of learning analytics, where visualized data helps instructors monitor student progress and identify support needs. == Ongoing development == Current work focuses on enhancing rubric structure, integrating principles from human–computer interaction, DataVis and expanding learning-analytics capabilities. Ongoing studies investigate how interface design, reviewer behavior, and classroom context influence the quality of feedback and overall engagement. Continuing development positions Visual Peer Review at the intersection of data visualization education, peer assessment, and educational technology.
Read more →
Mathematical knowledge management

Mathematical knowledge management (MKM) is the study of how society can effectively make use of the vast and growing literature on mathematics. It studies approaches such as databases of mathematical knowledge, automated processing of formulae and the use of semantic information, and artificial intelligence. Mathematics is particularly suited to a systematic study of automated knowledge processing due to the high degree of interconnectedness between different areas of mathematics.
Read more →
PagedAttention

PagedAttention is an attention algorithm for efficient serving of large language models (LLMs). It was introduced in 2023 by Woosuk Kwon and colleagues in the paper Efficient Memory Management for Large Language Model Serving with PagedAttention, alongside the vLLM serving engine. The method stores the key–value cache used during autoregressive decoding in fixed-size blocks that can be mapped to non-contiguous physical memory, borrowing ideas from virtual memory, paging, and operating system design. == Background == In transformer inference, the key–value cache grows with sequence length and the number of concurrent requests. Kwon et al. argued that earlier serving systems typically reserved contiguous cache regions in advance, which caused reserved space, internal fragmentation, and external fragmentation. In their experiments, the paper reported that the effective memory utilization of previous systems could fall as low as 20.4%. == Description == PagedAttention partitions the cache of each sequence into fixed-size KV blocks. A request's cache is represented as a sequence of logical blocks, while a block table maps those logical blocks to physical GPU-memory blocks. As a result, neighboring logical blocks do not need to be contiguous in physical memory, and new blocks can be allocated on demand as generation proceeds. The design also makes it easier to share cache state across related decoding paths. In vLLM, physical blocks can be reference-counted and shared among requests or branches, with block-granularity copy-on-write used when a shared block must be modified. The original paper applied this design to parallel sampling, beam search, and prompts with shared prefixes. == Mathematical formulation == For a query token i {\displaystyle i} in causal self-attention, the standard attention output can be written as a i j = exp ⁡ ( q i ⊤ k j / d ) ∑ t = 1 i exp ⁡ ( q i ⊤ k t / d ) , o i = ∑ j = 1 i a i j v j {\displaystyle a_{ij}={\frac {\exp(\mathbf {q} _{i}^{\top }\mathbf {k} _{j}/{\sqrt {d}})}{\sum _{t=1}^{i}\exp(\mathbf {q} _{i}^{\top }\mathbf {k} _{t}/{\sqrt {d}})}},\;\mathbf {o} _{i}=\sum _{j=1}^{i}a_{ij}\mathbf {v} _{j}} where q i {\displaystyle \mathbf {q} _{i}} , k j {\displaystyle \mathbf {k} _{j}} , and v j {\displaystyle \mathbf {v} _{j}} are the query, key, and value vectors, and d {\displaystyle d} is the attention dimension. If the cache is partitioned into blocks of size B {\displaystyle B} , the key and value blocks may be written as K j = ( k ( j − 1 ) B + 1 , … , k j B ) , V j = ( v ( j − 1 ) B + 1 , … , v j B ) {\displaystyle \mathbf {K} _{j}=(\mathbf {k} _{(j-1)B+1},\ldots ,\mathbf {k} _{jB}),\;\mathbf {V} _{j}=(\mathbf {v} _{(j-1)B+1},\ldots ,\mathbf {v} _{jB})} PagedAttention then performs the computation blockwise: A i j = exp ⁡ ( q i ⊤ K j / d ) ∑ t = 1 ⌈ i / B ⌉ exp ⁡ ( q i ⊤ K t / d ) , o i = ∑ j = 1 ⌈ i / B ⌉ V j A i j ⊤ {\displaystyle \mathbf {A} _{ij}={\frac {\exp(\mathbf {q} _{i}^{\top }\mathbf {K} _{j}/{\sqrt {d}})}{\sum _{t=1}^{\lceil i/B\rceil }\exp(\mathbf {q} _{i}^{\top }\mathbf {K} _{t}/{\sqrt {d}})}},\;\mathbf {o} _{i}=\sum _{j=1}^{\lceil i/B\rceil }\mathbf {V} _{j}\mathbf {A} _{ij}^{\top }} where A i j {\displaystyle \mathbf {A} _{ij}} is the vector of attention scores for the j {\displaystyle j} -th KV block. In the formulation given by Kwon et al., this preserves the causal attention calculation while allowing the key and value blocks to reside in non-contiguous physical memory. == Performance and use == The vLLM paper reported that, on its evaluated workloads, the use of PagedAttention and the associated memory-management design improved serving throughput by 2–4× over the compared baselines, including FasterTransformer and Orca, while preserving model outputs. In experiments on OPT-13B with the Alpaca trace, the paper also reported memory savings of 6.1–9.8% for parallel sampling and 37.6–55.2% for beam search through KV-block sharing. A 2024 survey of LLM serving systems described PagedAttention as having become an industry norm in LLM serving frameworks, citing support in TGI, vLLM, and TensorRT-LLM. == Limitations and alternatives == Subsequent work has described trade-offs in the approach. The 2025 vAttention paper argued that PagedAttention requires attention kernels to be rewritten to support paging and increases software complexity, portability issues, redundancy, and execution overhead, proposing instead a memory manager that keeps the cache contiguous in virtual memory while relying on demand paging for physical allocation. === vAttention === Unlike PagedAttention, vAttention does not introduce a different attention rule; it retains the standard attention computation Attention ⁡ ( q i , K , V ) = softmax ⁡ ( q i K ⊤ s c a l e ) V . {\displaystyle \operatorname {Attention} (q_{i},K,V)=\operatorname {softmax} \left({\frac {q_{i}K^{\top }}{\mathrm {scale} }}\right)V.} In the notation of Prabhu et al., the key and value tensors for a request seen so far are K , V ∈ R L ′ × ( H × D ) {\displaystyle K,V\in \mathbb {R} ^{L'\times (H\times D)}} , where L ′ {\displaystyle L'} is the context length seen so far, H {\displaystyle H} is the number of KV heads on a worker, and D {\displaystyle D} is the dimension of each KV head. In systems prior to PagedAttention, the K cache (or V cache) at each layer of a worker is typically allocated as a 4D tensor of shape [ B , L , H , D ] , {\displaystyle [B,L,H,D],} where B {\displaystyle B} is batch size and L {\displaystyle L} is the maximum context length supported by the model. vAttention preserves this contiguous virtual-memory view while deferring physical-memory allocation to runtime. A serving framework maintains separate K and V tensors for each layer, so vAttention reserves 2 N {\displaystyle 2N} virtual-memory buffers on a worker, where N {\displaystyle N} is the number of layers managed by that worker. The maximum size of one virtual-memory buffer is B S = B × S , {\displaystyle BS=B\times S,} where S {\displaystyle S} is the maximum size of a single request's per-layer K cache (or V cache) on a worker. The paper defines S = L × H × D × P , {\displaystyle S=L\times H\times D\times P,} where P {\displaystyle P} is the number of bytes needed to store one element. In this formulation, vAttention keeps the KV cache contiguous in virtual memory and relies on demand paging for physical allocation, rather than modifying the attention kernel to operate over non-contiguous KV-cache blocks.
Read more →
Relational data stream management system

A relational data stream management system (RDSMS) is a distributed, in-memory data stream management system (DSMS) that is designed to use standards-compliant SQL queries to process unstructured and structured data streams in real-time. Unlike SQL queries executed in a traditional RDBMS, which return a result and exit, SQL queries executed in a RDSMS do not exit, generating results continuously as new data become available. Continuous SQL queries in a RDSMS use the SQL Window function to analyze, join and aggregate data streams over fixed or sliding windows. Windows can be specified as time-based or row-based. == RDSMS SQL Query Examples == Continuous SQL queries in a RDSMS conform to the ANSI SQL standards. The most common RDSMS SQL query is performed with the declarative SELECT statement. A continuous SQL SELECT operates on data across one or more data streams, with optional keywords and clauses that include FROM with an optional JOIN subclause to specify the rules for joining multiple data streams, the WHERE clause and comparison predicate to restrict the records returned by the query, GROUP BY to project streams with common values into a smaller set, HAVING to filter records resulting from a GROUP BY, and ORDER BY to sort the results. The following is an example of a continuous data stream aggregation using a SELECT query that aggregates a sensor stream from a weather monitoring station. The SELECTquery aggregates the minimum, maximum and average temperature values over a one-second time period, returning a continuous stream of aggregated results at one second intervals. RDSMS SQL queries also operate on data streams over time or row-based windows. The following example shows a second continuous SQL query using the WINDOW clause with a one-second duration. The WINDOW clause changes the behavior of the query, to output a result for each new record as it arrives. Hence the output is a stream of incrementally updated results with zero result latency.
Read more →
Enumeration algorithm

In computer science, an enumeration algorithm is an algorithm that enumerates the answers to a computational problem. Formally, such an algorithm applies to problems that take an input and produce a list of solutions, similarly to function problems. For each input, the enumeration algorithm must produce the list of all solutions, without duplicates, and then halt. The performance of an enumeration algorithm is measured in terms of the time required to produce the solutions, either in terms of the total time required to produce all solutions, or in terms of the maximal delay between two consecutive solutions and in terms of a preprocessing time, counted as the time before outputting the first solution. This complexity can be expressed in terms of the size of the input, the size of each individual output, or the total size of the set of all outputs, similarly to what is done with output-sensitive algorithms. == Formal definitions == An enumeration problem P {\displaystyle P} is defined as a relation R {\displaystyle R} over strings of an arbitrary alphabet Σ {\displaystyle \Sigma } : R ⊆ Σ ∗ × Σ ∗ {\displaystyle R\subseteq \Sigma ^{}\times \Sigma ^{}} An algorithm solves P {\displaystyle P} if for every input x {\displaystyle x} the algorithm produces the (possibly infinite) sequence y {\displaystyle y} such that y {\displaystyle y} has no duplicate and z ∈ y {\displaystyle z\in y} if and only if ( x , z ) ∈ R {\displaystyle (x,z)\in R} . The algorithm should halt if the sequence y {\displaystyle y} is finite. == Common complexity classes == Enumeration problems have been studied in the context of computational complexity theory, and several complexity classes have been introduced for such problems. A very general such class is EnumP, the class of problems for which the correctness of a possible output can be checked in polynomial time in the input and output. Formally, for such a problem, there must exist an algorithm A which takes as input the problem input x, the candidate output y, and solves the decision problem of whether y is a correct output for the input x, in polynomial time in x and y. For instance, this class contains all problems that amount to enumerating the witnesses of a problem in the class NP. Other classes that have been defined include the following. In the case of problems that are also in EnumP, these problems are ordered from least to most specific: Output polynomial, the class of problems whose complete output can be computed in polynomial time. Incremental polynomial time, the class of problems where, for all i, the i-th output can be produced in polynomial time in the input size and in the number i. Polynomial delay, the class of problems where the delay between two consecutive outputs is polynomial in the input (and independent from the output). Strongly polynomial delay, the class of problems where the delay before each output is polynomial in the size of this specific output (and independent from the input or from the other outputs). The preprocessing is generally assumed to be polynomial. Constant delay, the class of problems where the delay before each output is constant, i.e., independent from the input and output. The preprocessing phase is generally assumed to be polynomial in the input. == Common techniques == Backtracking: The simplest way to enumerate all solutions is by systematically exploring the space of possible results (partitioning it at each successive step). However, performing this may not give good guarantees on the delay, i.e., a backtracking algorithm may spend a long time exploring parts of the space of possible results that do not give rise to a full solution. Flashlight search: This technique improves on backtracking by exploring the space of all possible solutions but solving at each step the problem of whether the current partial solution can be extended to a partial solution. If the answer is no, then the algorithm can immediately backtrack and avoid wasting time, which makes it easier to show guarantees on the delay between any two complete solutions. In particular, this technique applies well to self-reducible problems. Closure under set operations: If we wish to enumerate the disjoint union of two sets, then we can solve the problem by enumerating the first set and then the second set. If the union is non disjoint but the sets can be enumerated in sorted order, then the enumeration can be performed in parallel on both sets while eliminating duplicates on the fly. If the union is not disjoint and both sets are not sorted then duplicates can be eliminated at the expense of a higher memory usage, e.g., using a hash table. Likewise, the cartesian product of two sets can be enumerated efficiently by enumerating one set and joining each result with all results obtained when enumerating the second step. == Examples of enumeration problems == The vertex enumeration problem, where we are given a polytope described as a system of linear inequalities and we must enumerate the vertices of the polytope. Enumerating the minimal transversals of a hypergraph. This problem is related to monotone dualization and is connected to many applications in database theory and graph theory. Enumerating the answers to a database query, for instance a conjunctive query or a query expressed in monadic second-order. There have been characterizations in database theory of which conjunctive queries could be enumerated with linear preprocessing and constant delay. The problem of enumerating maximal cliques in an input graph, e.g., with the Bron–Kerbosch algorithm Listing all elements of structures such as matroids and greedoids Several problems on graphs, e.g., enumerating independent sets, paths, cuts, etc. Enumerating the satisfying assignments of representations of Boolean functions, e.g., a Boolean formula written in conjunctive normal form or disjunctive normal form, a binary decision diagram such as an OBDD, or a Boolean circuit in restricted classes studied in knowledge compilation, e.g., NNF. == Connection to computability theory == The notion of enumeration algorithms is also used in the field of computability theory to define some high complexity classes such as RE, the class of all recursively enumerable problems. This is the class of sets for which there exist an enumeration algorithm that will produce all elements of the set: the algorithm may run forever if the set is infinite, but each solution must be produced by the algorithm after a finite time.
Read more →