Knowledge graph embedding

In representation learning, knowledge graph embedding (KGE), also called knowledge representation learning (KRL), or multi-relation learning, is a machine learning task of learning a low-dimensional representation of a knowledge graph's entities and relations while preserving their semantic meaning. Leveraging their embedded representation, knowledge graphs can be used for various applications such as link prediction, triple classification, entity recognition, clustering, and relation extraction. == Definition == A knowledge graph G = { E , R , F } {\displaystyle {\mathcal {G}}=\{E,R,F\}} is a collection of entities E {\displaystyle E} , relations R {\displaystyle R} , and facts F {\displaystyle F} . A fact is a triple ( h , r , t ) ∈ F {\displaystyle (h,r,t)\in F} that denotes a link r ∈ R {\displaystyle r\in R} between the head h ∈ E {\displaystyle h\in E} and the tail t ∈ E {\displaystyle t\in E} of the triple. Another notation that is often used in the literature to represent a triple (or fact) is ⟨ head , relation , tail ⟩ {\displaystyle \langle {\text{head}},{\text{relation}},{\text{tail}}\rangle } . This notation is called the Resource Description Framework (RDF). A knowledge graph represents the knowledge related to a specific domain; leveraging this structured representation, it is possible to infer a piece of new knowledge from it after some refinement steps. However, nowadays, people have to deal with the sparsity of data and the computational inefficiency to use them in a real-world application. The embedding of a knowledge graph is a function that translates each entity and each relation into a vector of a given dimension d {\displaystyle d} , called embedding dimension. It is even possible to embed the entities and relations with different dimensions. The embedding vectors can then be used for other tasks. A knowledge graph embedding is characterized by four aspects: Representation space: The low-dimensional space in which the entities and relations are represented. Scoring function: A measure of the goodness of a triple-embedded representation. Encoding models: The modality in which the embedded representation of the entities and relations interact with each other. Additional information: Any additional information coming from the knowledge graph that can enrich the embedded representation. Usually, an ad hoc scoring function is integrated into the general scoring function for each additional piece of information. == Embedding procedure == All algorithms for creating a knowledge graph embedding follow the same approach. First, the embedding vectors are initialized to random values. Then, they are iteratively optimized using a training set of triples. In each iteration, a batch of size b {\displaystyle b} triples is sampled from the training set, and a triple from it is sampled and corrupted—i.e., a triple that does not represent a true fact in the knowledge graph. The corruption of a triple involves substituting the head or the tail (or both) of the triple with another entity that makes the fact false. The original triple and the corrupted triple are added in the training batch, and then the embeddings are updated, optimizing a scoring function. Iteration stops when a stop condition is reached. Usually, the stop condition depends on the overfitting of the training set. At the end, the learned embeddings should have extracted semantic meaning from the training triples and should correctly predict unseen true facts in the knowledge graph. === Pseudocode === The following is the pseudocode for the general embedding procedure. algorithm Compute entity and relation embeddings input: The training set S = { ( h , r , t ) } {\displaystyle S=\{(h,r,t)\}} , entity set E {\displaystyle E} , relation set R {\displaystyle R} , embedding dimension k {\displaystyle k} output: Entity and relation embeddings initialization: the entities e {\displaystyle e} and relations r {\displaystyle r} embeddings (vectors) are randomly initialized while stop condition do S b a t c h ← s a m p l e ( S , b ) {\displaystyle S_{batch}\leftarrow sample(S,b)} // Sample a batch from the training set for each ( h , r , t ) {\displaystyle (h,r,t)} in S b a t c h {\displaystyle S_{batch}} do ( h ′ , r , t ′ ) ← s a m p l e ( S ′ ) {\displaystyle (h',r,t')\leftarrow sample(S')} // Sample a corrupted fact T b a t c h ← T b a t c h ∪ { ( ( h , r , t ) , ( h ′ , r , t ′ ) ) } {\displaystyle T_{batch}\leftarrow T_{batch}\cup \{((h,r,t),(h',r,t'))\}} end for Update embeddings by minimizing the loss function end while == Performance indicators == These indexes are often used to measure the embedding quality of a model. The simplicity of the indexes makes them very suitable for evaluating the performance of an embedding algorithm even on a large scale. Given Q {\displaystyle {\ce {Q}}} as the set of all ranked predictions of a model, it is possible to define three different performance indexes: Hits@K, MR, and MRR. === Hits@K === Hits@K or in short, H@K, is a performance index that measures the probability to find the correct prediction in the first top K model predictions. Usually, it is used k = 10 {\displaystyle k=10} . Hits@K reflects the accuracy of an embedding model to predict the relation between two given triples correctly. Hits@K = | { q ∈ Q : q < k } | | Q | ∈ [ 0 , 1 ] {\displaystyle ={\frac {|\{q\in Q:q

Azure Maps

Azure Maps is a suite of cloud-based, location-based services provided by Microsoft as part of the company's Azure platform. The platform provides geospatial and location-based services via REST APIs and software development kits (SDKs). The service is typically used to integrate maps or geospatial data into applications. Azure Maps differs from Microsoft's other enterprise mapping service, Bing Maps, in its pricing model, focus on privacy, and its level of integration into the broader Azure cloud ecosystem. == History == Azure Maps was first introduced in public preview mode under the name "Azure Location Based Services" in 2017, primarily as an enterprise solution. The services was intended to add mapping and location-based functionality onto the existing Azure cloud services suite, seen as a critical part of Microsoft's broader Internet-of-Things (IoT) strategy. The preview version included APIs which could be used to develop location aware apps for use cases such as logistics and mobility. In 2018, the software was renamed "Azure Maps," and became generally available to the public, and a number of new functions were added, including route calculation, travel time calculation, and incorporation of real-time traffic data and incident information. Azure Maps was integrated with Azure IoT Central in 2018, which added tracking, monitoring, and geofencing capabilities. A set of mobility APIs on were added in 2019, with applications such as use in public transport apps and shared bicycle fleet management. “Azure Maps Creator,” which converts private facility floor plans into indoor map data, was also introduced in 2019. Some commentators linked these services to Microsoft's broader development of augmented reality products. In 2020, Azure Maps Visual for Power BI was released, integrating location-based features and mapping capabilities into Microsoft's business intelligence software. An elevation API (which was later retired), geolocation services, and an iOS and Android software development kit were introduced in 2021. In 2022, support for historical weather, air quality, and tropical storm data was made generally available and custom styling for indoor maps was also introduced. In 2023, Azure Maps was certified as HIPAA compliant in a move to target healthcare and health insurance companies. == Functionality == === Geocoding === Geocoding is one of the core functionalities of Azure Maps, converting addresses or place names into geographic coordinates. Batch geocoding is used to process large amounts of address data, a function used for route optimization and spatial analysis. === Reverse geocoding === Reverse geocoding derives human-readable information from geographic coordinates like longitude and latitude, used in navigation and by geographic information systems. === Routing === Azure Maps uses map data and routing algorithms to calculate the shortest or fastest routes between locations based on factors like vehicle size and type, traffic conditions, and distance. Routing also supports multi-modal routing, which include multiple modes of transport in a single trip, including cycling, walking, and ferries. This functionality is used for location-based searches and route optimization in applications like fleet management, proximity marketing, and emergency services as well as logistics and delivery, urban planning, ride sharing apps, and outdoor activities. === Map visualization === The platform supports map visualizations that can be modified to reflect real-time data (including from IoT sensors) as well as historical data patterns. Visualizations include heat maps, street maps, satellite imagery and other custom data layers. Maps are rendered using raster or vector tiles which reduce the load of displaying large data sets or complex maps. This can be used in various applications in areas like transportation, smart cities, retail and marketing, public health, and environmental monitoring. For example, it can be used for tracking the spread of diseases or measuring the impact of changing climatic patterns. === Geofencing and spatial analytics === Azure Maps supports polygonal geofencing, which enables the definition of custom geographic boundaries. Geofenced areas can be monitored in real-time for events of interest. For example, an application could send an alert when equipment or persons enter or leave a defined area. Tools for analyzing historical geofencing data are also available via the APIs for optimization purposes. == Industry usage == Azure Maps' geofencing function has seen usage in the construction industry, designating hazardous areas for safety purposes and sending alerts if anyone enters the area. Private facility maps are used by construction companies for monitoring large construction sites to increase productivity and prevent accidents or damage. In emergency management, New Zealand based company Beca has used Azure Maps to provide analysis on the impact of earthquakes to users, including information on the severity and location of an earthquake and the impact on affected properties. Alaska's Department of Transportation uses Azure Maps as part of an information system providing weather-related warnings and analytics to road crews. Airmap, an airspace management platform for drones, uses Azure Maps. Azure Maps has also been used in conjunction with Azure Monitor for risk monitoring by an insurance company. Other companies that use or have used Azure Maps include BMW, Banco Santander, Jvion, MV Transportation, C.H. Robertson, Wise Skulls, Tata Consultancy Services, Providence Health and Services, Gas Brasiliano Distribuidora S.A., Shell plc, Persistent Systems, Phase 2 Dining and Entertainment, Symbio, HID, Globant, and Insight Enterprises. == Partnerships == Azure Maps and TomTom have been partners since 2016, and TomTom provides location data to Azure Maps and can process data from Azure Maps for mapping purposes. In 2021, Azure Maps partnered with AccuWeather to make climatic data available via its APIs, making weather data along all parts of calculated routes available for mobility and logistics purposes. Microsoft has partnered with Esri, the developer of ArcGIS, and there is cross-compatibility between Azure and ArcGIS so that data from Azure Maps can be integrated into ArcGIS and vice versa. Azure Maps partnered with Moovit in 2019, a startup providing software that interfaces with public transport data. Moovit's database on global public transit networks, including information on which stations and facilities are wheelchair accessible, was linked to Azure Maps. This service was noted for its use increasing accessibility to public transport for the visually impaired by means of voice activated route planning assistance. NORAD has used some Azure Maps functions for their NORAD Tracks Santa website during Christmas holidays. == Components == === REST APIs === Various APIs cover the major functionalities across Azure Maps: Data registry API Geolocation API Render API Route API Search API Spatial API Time zone API Traffic API Weather API === SDKs === Azure Maps SDKs uses MapLibre-style specifications and open source MapLibre GL-based libraries as a rendering engine. The Web SDK is used for developing web apps with maps and location-based data and functionality. It includes a map control module as well as modules with drawing tools. It also supports Azure Maps Creator and various spatial data formats. The platform also includes a set of REST SDKs for developers integrating Azure Maps REST APIs into Python, C#, Java or JavaScript applications. Azure Maps also includes Android and iOS SDKs used for developing applications for Android and Apple devices. === Azure Maps Creator === Azure Maps Creator is a tool for generating custom maps for locations like large office complexes, construction sites, or university campuses. These maps can then be integrated into applications and used with other Azure Maps functions for purposes such as wayfinding and maintenance and security in building automation contexts. === Azure Maps Visual for Power BI === Azure Maps is integrated with Microsoft Power BI, a graphical tool for producing data visualizations. Since July 2020, Power BI can be used in conjunction with Azure Maps for developing map-based data visualizations. This functionality entered general availability in May 2023.

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.

Document capture software

Document capture software refers to applications that provide the ability and feature set to automate the process of scanning paper documents or importing electronic documents, often for the purposes of feeding advanced document classification and data collection processes. Most scanning hardware, both scanners and copiers, provides the basic ability to scan to any number of image file formats, including: PDF, TIFF, JPG, BMP, etc. This basic functionality is augmented by document capture software, which can add efficiency and standardization to the process. == Typical features == Typical features of Document Capture Software include: Barcode recognition Patch Code recognition Separation Optical Character Recognition (OCR) Optical Mark Recognition (OMR) Quality Assurance Indexing Migration === Goal for implementation of a document capture solution === The goal for implementing a document capture solution is to reduce the amount of time spent scanning, separating, enhancing, organizing, classifying, normalizing, and collecting information from document collections, and to produce metadata along with an image/PDF file, and/or OCR text. This information is then migrated to a file share, FTP site, database, Document Management or Enterprise Content Management system. These systems often provide a search function, allowing search of the assets based on the produced metadata, and then viewed using document imaging software. == General document capture system solutions == === Integration with document management system === ECM (Enterprise Content management) and their DMS component (Document Management System) are being adopted by many organizations as a corporate document management system for all types of electronic files, e.g. MS word, PDF ... However, much of the information held by organisations is on paper and this needs to be integrated within the same document repository. By converting paper documents into digital format through scanning, organizations convert paper into image formats such as TIF, JPG, and PDF, and also extract valuable index information or business data from the document using OCR technology. Digital documents and associated metadata can easily be stored in the ECM in a variety of formats. The most popular of these formats is PDF which not only provides an accurate representation of the document but also allows all the OCR text in the document to be stored behind the PDF image. This format is known as PDF with hidden text or text-searchable PDF. This allows users to search for documents by using keywords in the metadata fields or by searching the content of PDF files across the repository. ==== Advantages of scanning documents into a ECM/DMS ==== Information held on paper is usually just as valuable to organisations as the electronic documents that are generated internally. Often this information represents a large proportion of the day to day correspondence with suppliers and customers. Having the ability to manage and share this information internally through a document management system such as SharePoint or a CMIS-compatible repository improves collaboration between departments or employees and also eliminates the risk of losing this information through disasters such as floods or fire. Organisations adopting an ECM/DMS often implement electronic workflow which allows the information held on paper to be included as part of an electronic business process and incorporated into a customer record file along with other associated office documents and emails. For business critical documents, such as purchase orders and supplier invoices, digitising documents helps speed up business transactions as well as reduce manual effort involved in keying data into business systems, such as CRM, ERP and Accounting. Scanned invoices can also be routed to managers for payment approval via email or an electronic workflow. == Electronic document capture == In the earlier implementations of Document Capture Software, the technology focused solely on the digitization and capture of information from paper documents. Document images were acquired from document scanners via TWAIN/ISIS drivers. Only image-based file formats like TIF, JPG, and BMP were typically compatible with these solutions. But in recent years, as the volume of electronically-created documents and the number of proprietary file formats continues to increase at exponential rates, the need for handling documents existing in electronic formats has grown. The relevant document capture products have adapted to function with non-image file formats with the end-goal of creating a unified processing workflow capable of handling all incoming documents The ability to import files from a variety of sources is one example of such adaptation. Importing documents from ECM/DMS software solutions, email servers, FTP, and EDI is now as much of a requirement of document capture software as is paper capture. The normalization of output files to text-based PDF format is now another critical factor in long-term archival of proprietary electronic file formats. Normalization expands access and usage of files to users throughout the enterprise, rather than only those that created the original electronic file.

Data management plan

A data management plan or DMP is a formal document that outlines how data are to be handled both during a research project, and after the project is completed. The goal of a data management plan is to consider the many aspects of data management, metadata generation, data preservation, and analysis before the project begins; this may lead to data being well-managed in the present, and prepared for preservation in the future. DMPs were originally used in 1966 to manage aeronautical and engineering projects' data collection and analysis, and expanded across engineering and scientific disciplines in the 1970s and 1980s. Up until the early 2000s, DMPs were used "for projects of great technical complexity, and for limited mid-study data collection and processing purposes". In the 2000s and later, E-research and economic policies drove the development and uptake of DMPs. == Importance == Preparing a data management plan before data are collected is claimed to ensure that data are in the correct format, organized well, and better annotated. This could arguably save time in the long term because there is no need to re-organize, re-format, or try to remember details about data. It is also claimed to increase research efficiency since both the data collector and other researchers might be able to understand and use well-annotated data in the future. One component of a data management plan is data archiving and preservation. By deciding on an archive ahead of time, the data collector can format data during collection to make its future submission to a database easier. If data are preserved, they are more relevant since they can be re-used by other researchers. It also allows the data collector to direct requests for data to the database, rather than address requests individually. A frequent argument in favor of preservation is that data that are preserved have the potential to lead to new, unanticipated discoveries, and they prevent duplication of scientific studies that have already been conducted. Data archiving also provides insurance against loss by the data collector. In the 2010s, funding agencies increasingly required data management plans as part of the proposal and evaluation process, despite little or no evidence of their efficacy. == Major components == "There is no general and definitive list of topics that should be covered in a DMP for a research project", and researchers are often left to their own devices as to how to fill out a DMP. === Information about data and data format === A description of data to be produced by the project. This might include (but is not limited to) data that are: Experimental Observational Raw or derived Physical collections Models Simulations Curriculum materials Software Images How will the data be acquired? When and where will they be acquired? After collection, how will the data be processed? Include information about Software used Algorithms Scientific workflows File formats that will be used, justify those formats, and describe the naming conventions used. Quality assurance & quality control measures that will be taken during sample collection, analysis, and processing. If existing data are used, what are their origins? How will the data collected be combined with existing data? What is the relationship between the data collected and existing data? How will the data be managed in the short-term? Consider the following: Version control for files Backing up data and data products Security & protection of data and data products Who will be responsible for management === Metadata content and format === Metadata are the contextual details, including any information important for using data. This may include descriptions of temporal and spatial details, instruments, parameters, units, files, etc. Metadata is commonly referred to as "data about data". Issues to be considered include: How detailed has the metadata to be in order to make the data meaningful? How will the metadata be created and/or captured? Examples include lab notebooks, GPS hand-held units, Auto-saved files on instruments, etc. What format will be used for the metadata? What are the metadata standards commonly used in the respective scientific discipline? There should be justification for the format chosen. === Policies for access, sharing, and re-use === Describe any obligations that exist for sharing data collected. These may include obligations from funding agencies, institutions, other professional organizations, and legal requirements. Include information about how data will be shared, including when the data will be accessible, how long the data will be available, how access can be gained, and any rights that the data collector reserves for using data. Address any ethical or privacy issues with data sharing Address intellectual property & copyright issues. Who owns the copyright? What are the institutional, publisher, and/or funding agency policies associated with intellectual property? Are there embargoes for political, commercial, or patent reasons? Describe the intended future uses/users for the data Indicate how the data should be cited by others. How will the issue of persistent citation be addressed? For example, if the data will be deposited in a public archive, will the dataset have a persistent identifier (e.g., ARK, DOI, Handle, PURL, URN) assigned to it? === Long-term storage and data management === Researchers should identify an appropriate archive for the long-term preservation of their data. By identifying the archive early in the project, the data can be formatted, transformed, and documented appropriately to meet the requirements of the archive. Researchers should consult colleagues and professional societies in their discipline to determine the most appropriate database, and include a backup archive in their data management plan in case their first choice goes out of existence. Early in the project, the primary researcher should identify what data will be preserved in an archive. Usually, preserving the data in its most raw form is desirable, although data derivatives and products can also be preserved. An individual should be identified as the primary contact person for archived data, and ensure contact information is always kept up-to-date in case there are requests for data or information about data. === Budget === Data management and preservation costs may be considerable, depending on the nature of the project. By anticipating costs ahead of time, researchers ensure that the data will be properly managed and archived. Potential expenses that should be considered are Human resources and staff as they handle data preparation, management, documentation, and preservation Hardware and/or software needed for data management, backing up, security, documentation, and preservation Costs associated with submitting the data to an archive The data management plan should include how these costs will be paid. == NSF Data Management Plan == All grant proposals submitted to National Science Foundation (NSF) must include a Data Management Plan that is no more than two pages. This is a supplement (not part of the 15-page proposal) and should describe how the proposal will conform to the Award and Administration Guide policy (see below). It may include the following: The types of data The standards to be used for data and metadata format and content Policies for access and sharing Policies and provisions for re-use Plans for archiving data Policy summarized from the NSF Award and Administration Guide, Section 4 (Dissemination and Sharing of Research Results): Promptly publish with appropriate authorship Share data, samples, physical collections, and supporting materials with others, within a reasonable time frame Share software and inventions Investigators can keep their legal rights over their intellectual property, but they still have to make their results, data, and collections available to others Policies will be implemented via Proposal review Award negotiations and conditions Support/incentives == ESRC Data Management Plan == Since 1995, the UK's Economic and Social Research Council (ESRC) have had a research data policy in place. The current ESRC Research Data Policy states that research data created as a result of ESRC-funded research should be openly available to the scientific community to the maximum extent possible, through long-term preservation and high-quality data management. ESRC requires a data management plan for all research award applications where new data are being created. Such plans are designed to promote a structured approach to data management throughout the data lifecycle, resulting in better quality data that is ready to archive for sharing and re-use. The UK Data Service, the ESRC's flagship data service, provides practical guidance on research data management planning suitable for social science researchers in the UK and around the world. ESRC has a longstanding arrangement with the UK Data A

Neighborhood operation

In computer vision and image processing a neighborhood operation is a commonly used class of computations on image data which implies that it is processed according to the following pseudo code: Visit each point p in the image data and do { N = a neighborhood or region of the image data around the point p result(p) = f(N) } This general procedure can be applied to image data of arbitrary dimensionality. Also, the image data on which the operation is applied does not have to be defined in terms of intensity or color, it can be any type of information which is organized as a function of spatial (and possibly temporal) variables in p. The result of applying a neighborhood operation on an image is again something which can be interpreted as an image, it has the same dimension as the original data. The value at each image point, however, does not have to be directly related to intensity or color. Instead it is an element in the range of the function f, which can be of arbitrary type. Normally the neighborhood N is of fixed size and is a square (or a cube, depending on the dimensionality of the image data) centered on the point p. Also the function f is fixed, but may in some cases have parameters which can vary with p, see below. In the simplest case, the neighborhood N may be only a single point. This type of operation is often referred to as a point-wise operation. == Examples == The most common examples of a neighborhood operation use a fixed function f which in addition is linear, that is, the computation consists of a linear shift invariant operation. In this case, the neighborhood operation corresponds to the convolution operation. A typical example is convolution with a low-pass filter, where the result can be interpreted in terms of local averages of the image data around each image point. Other examples are computation of local derivatives of the image data. It is also rather common to use a fixed but non-linear function f. This includes median filtering, and computation of local variances. The Nagao-Matsuyama filter is an example of a complex local neighbourhood operation that uses variance as an indicator of the uniformity within a pixel group. The result is similar to a convolution with a low-pass filter with the added effect of preserving sharp edges. There is also a class of neighborhood operations in which the function f has additional parameters which can vary with p: Visit each point p in the image data and do { N = a neighborhood or region of the image data around the point p result(p) = f(N, parameters(p)) } This implies that the result is not shift invariant. Examples are adaptive Wiener filters. == Implementation aspects == The pseudo code given above suggests that a neighborhood operation is implemented in terms of an outer loop over all image points. However, since the results are independent, the image points can be visited in arbitrary order, or can even be processed in parallel. Furthermore, in the case of linear shift-invariant operations, the computation of f at each point implies a summation of products between the image data and the filter coefficients. The implementation of this neighborhood operation can then be made by having the summation loop outside the loop over all image points. An important issue related to neighborhood operation is how to deal with the fact that the neighborhood N becomes more or less undefined for points p close to the edge or border of the image data. Several strategies have been proposed: Compute result only for points p for which the corresponding neighborhood is well-defined. This implies that the output image will be somewhat smaller than the input image. Zero padding: Extend the input image sufficiently by adding extra points outside the original image which are set to zero. The loops over the image points described above visit only the original image points. Border extension: Extend the input image sufficiently by adding extra points outside the original image which are set to the image value at the closest image point. The loops over the image points described above visit only the original image points. Mirror extension: Extend the image sufficiently much by mirroring the image at the image boundaries. This method is less sensitive to local variations at the image boundary than border extension. Wrapping: The image is tiled, so that going off one edge wraps around to the opposite side of the image. This method assumes that the image is largely homogeneous, for example a stochastic image texture without large textons.

TurboQuant

TurboQuant is an online vector quantization algorithm for compressing high-dimensional Euclidean vectors while preserving their geometric structure. It was proposed in 2025 by Amir Zandieh, Majid Daliri, Majid Hadian, and Vahab Mirrokni in the paper TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate. The paper lists Zandieh and Mirrokni as affiliated with Google Research, Daliri with New York University, and Hadian with Google DeepMind. The method was developed for applications including large language model (LLM) inference, key–value (KV) cache compression, vector databases, and nearest neighbor search. TurboQuant consists of two related algorithms: TurboQuantmse, which is optimized for mean squared error (MSE), and TurboQuantprod, which is optimized for unbiased inner product estimation. The algorithm uses a random rotation of input vectors, applies scalar quantizers to the rotated coordinates, and, for inner-product estimation, applies a one-bit Quantized Johnson–Lindenstrauss (QJL) transform to the residual error. == Background == Vector quantization is a compression method that maps high-dimensional vectors to a finite set of codewords. The problem has roots in Shannon's source coding theory and rate–distortion theory. In machine learning and information retrieval, vector quantization is used to reduce the memory required to store embeddings, activation vectors, and other numerical representations. In Transformer-based large language models, the KV cache stores key and value vectors from previous tokens during autoregressive decoding. The size of this cache grows with context length, the number of attention heads, and the number of concurrent requests, making it a major memory bottleneck in LLM serving. Similar compression problems appear in vector search, where large collections of embedding vectors must be stored and searched efficiently. Earlier approaches to vector quantization include product quantization, scalar quantization, and data-dependent k-means codebook construction. The TurboQuant paper argues that many existing methods either require offline preprocessing and calibration or suffer from suboptimal distortion guarantees in online settings. == Algorithm == === TurboQuantmse === TurboQuantmse is the version of the algorithm optimized for mean-squared error. For a unit vector x ∈ S d − 1 {\displaystyle x\in S^{d-1}} , the algorithm first applies a random rotation matrix Π ∈ R d × d {\displaystyle \Pi \in \mathbb {R} ^{d\times d}} and sets z = Π x {\displaystyle z=\Pi x} . Each coordinate of the rotated vector follows a shifted and scaled beta distribution, which converges to a normal distribution in high dimensions. In high dimensions, distinct coordinates also become nearly independent, allowing the algorithm to apply scalar quantizers independently to each coordinate. The scalar quantizer is constructed by solving a one-dimensional continuous k-means or Lloyd–Max quantization problem. If the centroids are c 1 , c 2 , … , c 2 b {\displaystyle c_{1},c_{2},\ldots ,c_{2^{b}}} , the quantization step stores, for each coordinate, i d x j = ⁡ a r g m i n k ∈ [ 2 b ] | z j − c k | . {\displaystyle \mathrm {idx} _{j}=\operatorname {} {arg\,min}_{k\in [2^{b}]}|z_{j}-c_{k}|.} During dequantization, the stored index for each coordinate is replaced by the corresponding centroid, giving a reconstructed rotated vector z ~ {\displaystyle {\tilde {z}}} . The algorithm then rotates back: x ~ = Π ⊤ z ~ . {\displaystyle {\tilde {x}}=\Pi ^{\top }{\tilde {z}}.} The paper gives the following bound for TurboQuantmse: D m s e ≤ 3 π 2 ⋅ 1 4 b . {\displaystyle D_{\mathrm {mse} }\leq {\frac {\sqrt {3\pi }}{2}}\cdot {\frac {1}{4^{b}}}.} It also reports finer-grained MSE values of approximately 0.36, 0.117, 0.03, and 0.009 for bit-widths b = 1 , 2 , 3 , 4 {\displaystyle b=1,2,3,4} , respectively. === TurboQuantprod === TurboQuantprod is optimized for unbiased inner-product estimation. The authors note that an MSE-optimized quantizer may introduce bias when used to estimate inner products. To address this, TurboQuantprod first applies TurboQuantmse with bit-width b − 1 {\displaystyle b-1} , then applies a one-bit Quantized Johnson–Lindenstrauss transform to the remaining residual vector. Let r = x − Q m s e − 1 ( Q m s e ( x ) ) {\displaystyle r=x-Q_{\mathrm {mse} }^{-1}(Q_{\mathrm {mse} }(x))} be the residual after MSE quantization, and let γ = ‖ r ‖ 2 {\displaystyle \gamma =\|r\|_{2}} . The QJL step stores a sign vector for the residual. For γ ≠ 0 {\displaystyle \gamma \neq 0} , this can be written using the normalized residual u = r / γ {\displaystyle u=r/\gamma } : q j l = sign ⁡ ( S u ) , {\displaystyle qjl=\operatorname {sign} (Su),} where S ∈ R d × d {\displaystyle S\in \mathbb {R} ^{d\times d}} is a random projection matrix. Since the sign function is invariant under positive rescaling, this is equivalent to sign ⁡ ( S r ) {\displaystyle \operatorname {sign} (Sr)} when r ≠ 0 {\displaystyle r\neq 0} . If γ = 0 {\displaystyle \gamma =0} , the residual correction is zero. TurboQuantprod stores the MSE quantization, the QJL sign vector, and the residual norm: Q p r o d ( x ) = [ Q m s e ( x ) , q j l , γ ] . {\displaystyle Q_{\mathrm {prod} }(x)=\left[Q_{\mathrm {mse} }(x),qjl,\gamma \right].} The dequantized vector is reconstructed as x ~ = x ~ m s e + π / 2 d γ S ⊤ q j l . {\displaystyle {\tilde {x}}={\tilde {x}}_{\mathrm {mse} }+{\frac {\sqrt {\pi /2}}{d}}\,\gamma S^{\top }qjl.} The paper proves that TurboQuantprod is unbiased for inner-product estimation: E x ~ [ ⟨ y , x ~ ⟩ ] = ⟨ y , x ⟩ . {\displaystyle \mathbb {E} _{\tilde {x}}\left[\langle y,{\tilde {x}}\rangle \right]=\langle y,x\rangle .} It also gives the distortion bound D p r o d ≤ 3 π 2 ⋅ ‖ y ‖ 2 2 d ⋅ 1 4 b . {\displaystyle D_{\mathrm {prod} }\leq {\frac {\sqrt {3\pi }}{2}}\cdot {\frac {\|y\|_{2}^{2}}{d}}\cdot {\frac {1}{4^{b}}}.} == Performance and applications == The TurboQuant paper reports that the algorithm achieves near-optimal distortion rates within a small constant factor of information-theoretic lower bounds. The authors report that, for KV cache quantization, TurboQuant achieved quality neutrality at 3.5 bits per channel and marginal degradation at 2.5 bits per channel. In long-context LLM experiments using Llama 3.1 8B Instruct, the paper evaluated the method on a "needle-in-a-haystack" retrieval task with document lengths from 4,000 to 104,000 tokens. It reported that TurboQuant matched the uncompressed full-precision baseline while using more than 4× compression, and compared the method against PolarQuant, SnapKV, PyramidKV, and KIVI. Google Research stated that TurboQuant was evaluated on long-context benchmarks including LongBench, Needle in a Haystack, ZeroSCROLLS, RULER, and L-Eval using open-source models including Gemma and Mistral. According to a report in Tom's Hardware, Google described the method as reducing KV-cache memory by at least six times and achieving up to an eightfold improvement in attention-logit computation on Nvidia H100 GPUs compared with unquantized 32-bit keys. TurboQuant has also been applied to nearest-neighbor vector search. The original paper reports experiments on DBpedia entity embeddings and GloVe embeddings, comparing TurboQuant with product quantization and other vector-search quantization baselines. == Relationship to other methods == TurboQuant is related to several methods for efficient large language model inference and high-dimensional search: Product quantization – a vector quantization technique widely used for approximate nearest-neighbor search Quantization (machine learning) – reducing the numerical precision of weights, activations, or cached tensors in machine learning models PagedAttention – a memory-management algorithm for LLM serving that reduces fragmentation in the KV cache Johnson–Lindenstrauss lemma – a result in high-dimensional geometry used in random projection methods Lloyd's algorithm – an algorithm for scalar and vector quantization, including k-means-style codebook construction Unlike PagedAttention, which focuses on memory allocation and cache layout, TurboQuant reduces the numerical storage cost of the vectors themselves. Unlike many product-quantization methods, TurboQuant is designed to be data-oblivious and online, avoiding dataset-specific codebook training. == Limitations == The strongest performance claims for TurboQuant come from the original paper and Google Research's own publication. Coverage in technology media has noted that the broader impact of the method will depend on real-world implementation details, workloads, and hardware architectures.