Galaksija BASIC was the BASIC interpreter of the Galaksija build-it-yourself home computer from Yugoslavia. While being partially based on code taken from TRS-80 Level 1 BASIC, which the creator believed to have been a Microsoft BASIC, the extensive modifications of Galaksija BASIC—such as to include rudimentary array support, video generation code (as the CPU itself did it in absence of dedicated video circuitry) and generally improvements to the programming language—is said to have left not much more than flow-control and floating point code remaining from the original. The core implementation of the interpreter was fully contained in the 4 KiB ROM "A" or "1". The computer's original mainboard had a reserved slot for an extension ROM "B" or "2" that added more commands and features such as a built-in Zilog Z80 assembler. == ROM "A"/"1" symbols and keywords == The core implementation, in ROM "A" or "1", contained 3 special symbols and 32 keywords: ! begins a comment (equivalent of standard BASIC REM command) # Equivalent of standard BASIC DATA statement & prefix for hex numbers ARR$(n) Allocates an array of strings, like DIM, but can allocate only array with name A$ BYTE serves as PEEK when used as a function (e.g. PRINT BYTE(11123)) and POKE when used as a command (e.g. BYTE 11123,123). CALL n Calls BASIC subroutine as GOSUB in most other BASICs (e.g. CALL 100+4X) CHR$(n) converts an ASCII numeric code into a corresponding character (string) DOT x, y draws (command) or inspects (function) a pixel at given coordinates (0<=x<=63, 0<=y<=47). DOT displays the clock or time controlled by content of Y$ variable. Not in standard ROM EDIT n causes specified program line to be edited ELSE standard part of IF-ELSE construct (Galaksija did not use THEN) EQ compare alphanumeric values X$ and Y$ FOR standard FOR loop GOTO standard GOTO command HOME equivalent of standard BASIC CLS command - clears the screen HOME n protects n characters from the top of the screen from being scrolled away IF standard part of IF-ELSE construct (Galaksija did not use THEN) INPUT user entry of variable INT(n) a function that returns the greatest integer value equal to or lesser than n KEY(n) test whether a particular keyboard key is pressed LIST lists the program. Optional numeric argument specifies the first line number to begin listing with. MEM returns memory consumption data (need details here) NEW clears the current BASIC program NEW n clears BASIC program and moves beginning of BASIC area NEXT standard terminator of FOR loop OLD loads a program from tape OLD n loads program to different address PTR Returns address of the variable PRINT Printing numeric or string expression. RETURN Return from BASIC subroutine RND function (takes no arguments) that returns a random number between 0 and 1. RUN runs (executes) BASIC program. Optional numeric argument specifies the line number to begin execution with. SAVE saves a program to tape. Optional two arguments specify memory range to be saved (need details here). STEP standard part of FOR loop STOP stops execution of BASIC program TAKE replacement for READ and RESTORE. If the parameter is variable name, acts as READ, if it is number, acts as RESTORE UNDOT x, y "undraws" (resets) at specified coordinates (see DOT) UNDOT Stops the clock, not part of ROM USR Calls machine code subroutine WORD Double byte PEEK and POKE == ROM "B"/"2" additional symbols and keywords == The extended BASIC features, in ROM "B" or "2", contained one extra reserved symbol and 22 extra keywords: % /LABEL ABS(x) ARCTG(x) COS(x) COSD(x) DEL DUMP EXP(x) INP(x) LDUMP LLIST LN(x) LPRINT OUT PI POW(x,y) REN SIN(x), SIND(x) SQR(x) TG(x) TGD(x)
Database
In computing, a database is an organized collection of data or a type of data store based on the use of a database management system (DBMS), the software that interacts with end users, applications, and the database itself to capture and analyze the data. The DBMS additionally encompasses the core facilities provided to administer the database. The sum total of the database, the DBMS and the associated applications can be referred to as a database system. Often the term "database" is also used loosely to refer to any of the DBMS, the database system or an application associated with the database. Before digital storage and retrieval of data became widespread, index cards were used for data storage in a wide range of applications and environments: in the home to record and store recipes, shopping lists, contact information and other organizational data; in business to record presentation notes, project research and notes, and contact information; in schools as flash cards or other visual aids; and in academic research to hold data such as bibliographical citations or notes in a card file. Professional book indexers used index cards in the creation of book indexes until they were replaced by indexing software in the 1980s and 1990s. Small databases can be stored on a file system, while large databases are hosted on computer clusters or cloud storage. The design of databases spans formal techniques and practical considerations, including data modeling, efficient data representation and storage, query languages, security and privacy of sensitive data, and distributed computing issues, including supporting concurrent access and fault tolerance. Computer scientists may classify database management systems according to the database models that they support. Relational databases became dominant in the 1980s. These model data as rows and columns in a series of tables, and the vast majority use SQL for writing and querying data. In the 2000s, non-relational databases became popular, collectively referred to as NoSQL, because they use different query languages. == Terminology and overview == Formally, a "database" refers to a set of related data accessed through the use of a "database management system" (DBMS), which is an integrated set of computer software that allows users to interact with one or more databases and provides access to all of the data contained in the database (although restrictions may exist that limit access to particular data). The DBMS provides various functions that allow entry, storage and retrieval of large quantities of information and provides ways to manage how that information is organized. Because of the close relationship between them, the term "database" is often used casually to refer to both a database and the DBMS used to manipulate it. Outside the world of professional information technology, the term database is often used to refer to any collection of related data (such as a spreadsheet or a card index) as size and usage requirements typically necessitate use of a database management system. Existing DBMSs provide various functions that allow management of a database and its data which can be classified into four main functional groups: Data definition – Creation, modification and removal of definitions that detail how the data is to be organized. Update – Insertion, modification, and deletion of the data itself. Retrieval – Selecting data according to specified criteria (e.g., a query, a position in a hierarchy, or a position in relation to other data) and providing that data either directly to the user, or making it available for further processing by the database itself or by other applications. The retrieved data may be made available in a more or less direct form without modification, as it is stored in the database, or in a new form obtained by altering it or combining it with existing data from the database. Administration – Registering and monitoring users, enforcing data security, monitoring performance, maintaining data integrity, dealing with concurrency control, and recovering information that has been corrupted by some event such as an unexpected system failure. Both a database and its DBMS conform to the principles of a particular database model. "Database system" refers collectively to the database model, database management system, and database. Physically, database servers are dedicated computers that hold the actual databases and run only the DBMS and related software. Database servers are usually multiprocessor computers, with generous memory and RAID disk arrays used for stable storage. Hardware database accelerators, connected to one or more servers via a high-speed channel, are also used in large-volume transaction processing environments. DBMSs are found at the heart of most database applications. DBMSs may be built around a custom multitasking kernel with built-in networking support, but modern DBMSs typically rely on a standard operating system to provide these functions. Since DBMSs comprise a significant market, computer and storage vendors often take into account DBMS requirements in their own development plans. Databases and DBMSs can be categorized according to the database model(s) that they support (such as relational or XML), the type(s) of computer they run on (from a server cluster to a mobile phone), the query language(s) used to access the database (such as SQL or XQuery), and their internal engineering, which affects performance, scalability, resilience, and security. == History == The sizes, capabilities, and performance of databases and their respective DBMSs have grown in orders of magnitude. These performance increases were enabled by the technology progress in the areas of processors, computer memory, computer storage, and computer networks. The concept of a database was made possible by the emergence of direct access storage media such as magnetic disks, which became widely available in the mid-1960s; earlier systems relied on sequential storage of data on magnetic tape. The subsequent development of database technology can be divided into three eras based on data model or structure: navigational, SQL/relational, and post-relational. The two main early navigational data models were the hierarchical model and the CODASYL model (network model). These were characterized by the use of pointers (often physical disk addresses) to follow relationships from one record to another. The relational model, first proposed in 1970 by Edgar F. Codd, departed from this tradition by insisting that applications should search for data by content, rather than by following links. The relational model employs sets of ledger-style tables, each used for a different type of entity. Only in the mid-1980s did computing hardware become powerful enough to allow the wide deployment of relational systems (DBMSs plus applications). By the early 1990s, however, relational systems dominated in all large-scale data processing applications, and as of 2018 they remain dominant: IBM Db2, Oracle, MySQL, and Microsoft SQL Server are the most searched DBMS. The dominant database language, standardized SQL for the relational model, has influenced database languages for other data models. Object databases were developed in the 1980s to overcome the inconvenience of object–relational impedance mismatch, which led to the coining of the term "post-relational" and also the development of hybrid object–relational databases. The next generation of post-relational databases in the late 2000s became known as NoSQL databases, introducing fast key–value stores and document-oriented databases. A competing "next generation" known as NewSQL databases attempted new implementations that retained the relational/SQL model while aiming to match the high performance of NoSQL compared to commercially available relational DBMSs. === 1960s, navigational DBMS === The introduction of the term database coincided with the availability of direct-access storage (disks and drums) from the mid-1960s onwards. The term represented a contrast with the tape-based systems of the past, allowing shared interactive use rather than daily batch processing. The Oxford English Dictionary cites a 1962 report by the System Development Corporation of California as the first to use the term "data-base" in a specific technical sense. As computers grew in speed and capability, a number of general-purpose database systems emerged; by the mid-1960s a number of such systems had come into commercial use. Interest in a standard began to grow, and Charles Bachman, author of one such product, the Integrated Data Store (IDS), founded the Database Task Group within CODASYL, the group responsible for the creation and standardization of COBOL. In 1971, the Database Task Group delivered their standard, which generally became known as the CODASYL approach, and soon a number of commercial products based on this approach entered the market. The CODASYL approach of
Generative adversarial network
A generative adversarial network (GAN) is a class of machine learning frameworks and a prominent framework for approaching generative artificial intelligence. The concept was initially developed by Ian Goodfellow and his colleagues in June 2014. In a GAN, two neural networks compete with each other in the form of a zero-sum game, where one agent's gain is another agent's loss. Given a training set, this technique learns to generate new data with the same statistics as the training set. For example, a GAN trained on photographs can generate new photographs that look at least superficially authentic to human observers, having many realistic characteristics. Though originally proposed as a form of generative model for unsupervised learning, GANs have also proved useful for semi-supervised learning, fully supervised learning, and reinforcement learning. The core idea of a GAN is based on the "indirect" training through the discriminator, another neural network that can tell how "realistic" the input seems, which itself is also being updated dynamically. This means that the generator is not trained to minimize the distance to a specific image, but rather to fool the discriminator. This enables the model to learn in an unsupervised manner. GANs are similar to mimicry in evolutionary biology, with an evolutionary arms race between both networks. == Definition == === Mathematical === The original GAN is defined as the following game: Each probability space ( Ω , μ ref ) {\displaystyle (\Omega ,\mu _{\text{ref}})} defines a GAN game. There are 2 players: generator and discriminator. The generator's strategy set is P ( Ω ) {\displaystyle {\mathcal {P}}(\Omega )} , the set of all probability measures μ G {\displaystyle \mu _{G}} on Ω {\displaystyle \Omega } . The discriminator's strategy set is the set of Markov kernels μ D : Ω → P [ 0 , 1 ] {\displaystyle \mu _{D}:\Omega \to {\mathcal {P}}[0,1]} , where P [ 0 , 1 ] {\displaystyle {\mathcal {P}}[0,1]} is the set of probability measures on [ 0 , 1 ] {\displaystyle [0,1]} . The GAN game is a zero-sum game, with objective function L ( μ G , μ D ) := E x ∼ μ ref , y ∼ μ D ( x ) [ ln y ] + E x ∼ μ G , y ∼ μ D ( x ) [ ln ( 1 − y ) ] . {\displaystyle L(\mu _{G},\mu _{D}):=\operatorname {E} _{x\sim \mu _{\text{ref}},y\sim \mu _{D}(x)}[\ln y]+\operatorname {E} _{x\sim \mu _{G},y\sim \mu _{D}(x)}[\ln(1-y)].} The generator aims to minimize the objective, and the discriminator aims to maximize the objective. The generator's task is to approach μ G ≈ μ ref {\displaystyle \mu _{G}\approx \mu _{\text{ref}}} , that is, to match its own output distribution as closely as possible to the reference distribution. The discriminator's task is to output a value close to 1 when the input appears to be from the reference distribution, and to output a value close to 0 when the input looks like it came from the generator distribution. === In practice === The generative network generates candidates while the discriminative network evaluates them. This creates a contest based on data distributions, where the generator learns to map from a latent space to the true data distribution, aiming to produce candidates that the discriminator cannot distinguish from real data. The discriminator's goal is to correctly identify these candidates, but as the generator improves, its task becomes more challenging, increasing the discriminator's error rate. A known dataset serves as the initial training data for the discriminator. Training involves presenting it with samples from the training dataset until it achieves acceptable accuracy. The generator is trained based on whether it succeeds in fooling the discriminator. Typically, the generator is seeded with randomized input that is sampled from a predefined latent space (e.g. a multivariate normal distribution). Thereafter, candidates synthesized by the generator are evaluated by the discriminator. Independent backpropagation procedures are applied to both networks so that the generator produces better samples, while the discriminator becomes more skilled at flagging synthetic samples. When used for image generation, the generator is typically a deconvolutional neural network, and the discriminator is a convolutional neural network. === Relation to other statistical machine learning methods === GANs are implicit generative models, which means that they do not explicitly model the likelihood function nor provide a means for finding the latent variable corresponding to a given sample, unlike alternatives such as flow-based generative model. Compared to fully visible belief networks such as WaveNet and PixelRNN and autoregressive models in general, GANs can generate one complete sample in one pass, rather than multiple passes through the network. Compared to Boltzmann machines and linear ICA, there is no restriction on the type of function used by the network. Since neural networks are universal approximators, GANs are asymptotically consistent. Variational autoencoders might be universal approximators, but it is not proven as of 2017. == Mathematical properties == === Measure-theoretic considerations === This section provides some of the mathematical theory behind these methods. In modern probability theory based on measure theory, a probability space also needs to be equipped with a σ-algebra. As a result, a more rigorous definition of the GAN game would make the following changes:Each probability space ( Ω , B , μ ref ) {\displaystyle (\Omega ,{\mathcal {B}},\mu _{\text{ref}})} defines a GAN game. The generator's strategy set is P ( Ω , B ) {\displaystyle {\mathcal {P}}(\Omega ,{\mathcal {B}})} , the set of all probability measures μ G {\displaystyle \mu _{G}} on the measure-space ( Ω , B ) {\displaystyle (\Omega ,{\mathcal {B}})} . The discriminator's strategy set is the set of Markov kernels μ D : ( Ω , B ) → P ( [ 0 , 1 ] , B ( [ 0 , 1 ] ) ) {\displaystyle \mu _{D}:(\Omega ,{\mathcal {B}})\to {\mathcal {P}}([0,1],{\mathcal {B}}([0,1]))} , where B ( [ 0 , 1 ] ) {\displaystyle {\mathcal {B}}([0,1])} is the Borel σ-algebra on [ 0 , 1 ] {\displaystyle [0,1]} .Since issues of measurability never arise in practice, these will not concern us further. === Choice of the strategy set === In the most generic version of the GAN game described above, the strategy set for the discriminator contains all Markov kernels μ D : Ω → P [ 0 , 1 ] {\displaystyle \mu _{D}:\Omega \to {\mathcal {P}}[0,1]} , and the strategy set for the generator contains arbitrary probability distributions μ G {\displaystyle \mu _{G}} on Ω {\displaystyle \Omega } . However, as shown below, the optimal discriminator strategy against any μ G {\displaystyle \mu _{G}} is deterministic, so there is no loss of generality in restricting the discriminator's strategies to deterministic functions D : Ω → [ 0 , 1 ] {\displaystyle D:\Omega \to [0,1]} . In most applications, D {\displaystyle D} is a deep neural network function. As for the generator, while μ G {\displaystyle \mu _{G}} could theoretically be any computable probability distribution, in practice, it is usually implemented as a pushforward: μ G = μ Z ∘ G − 1 {\displaystyle \mu _{G}=\mu _{Z}\circ G^{-1}} . That is, start with a random variable z ∼ μ Z {\displaystyle z\sim \mu _{Z}} , where μ Z {\displaystyle \mu _{Z}} is a probability distribution that is easy to compute (such as the uniform distribution, or the Gaussian distribution), then define a function G : Ω Z → Ω {\displaystyle G:\Omega _{Z}\to \Omega } . Then the distribution μ G {\displaystyle \mu _{G}} is the distribution of G ( z ) {\displaystyle G(z)} . Consequently, the generator's strategy is usually defined as just G {\displaystyle G} , leaving z ∼ μ Z {\displaystyle z\sim \mu _{Z}} implicit. In this formalism, the GAN game objective is L ( G , D ) := E x ∼ μ ref [ ln D ( x ) ] + E z ∼ μ Z [ ln ( 1 − D ( G ( z ) ) ) ] . {\displaystyle L(G,D):=\operatorname {E} _{x\sim \mu _{\text{ref}}}[\ln D(x)]+\operatorname {E} _{z\sim \mu _{Z}}[\ln(1-D(G(z)))].} === Generative reparametrization === The GAN architecture has two main components. One is casting optimization into a game, of form min G max D L ( G , D ) {\displaystyle \min _{G}\max _{D}L(G,D)} , which is different from the usual kind of optimization, of form min θ L ( θ ) {\displaystyle \min _{\theta }L(\theta )} . The other is the decomposition of μ G {\displaystyle \mu _{G}} into μ Z ∘ G − 1 {\displaystyle \mu _{Z}\circ G^{-1}} , which can be understood as a reparametrization trick. To see its significance, one must compare GAN with previous methods for learning generative models, which were plagued with "intractable probabilistic computations that arise in maximum likelihood estimation and related strategies". At the same time, Kingma and Welling and Rezende et al. developed the same idea of reparametrization into a general stochastic backpropagation method. Among its first applications was the variational autoencoder. === Move order and st
2025 Abu Dhabi Autonomous Racing League
The 2025 season of the Abu Dhabi Autonomous Racing League began on 11 April 2025 in Abu Dhabi. This year marks the first multi-format season of the A2RL, racing both drones and self-driving cars. The venue of choice for the Car Race, set for 15 November 2025, is the Yas Marina Circuit, same as the previous year, while the Drone Race was held at the ADNEC Marina Hall. == Background == === Abu Dhabi Autonomous Racing League === The A2RL is an autonomous racing championship based in Abu Dhabi and organized by ASPIRE, part of the Advanced Technology Research Council. It is one of two active autonomous car racing championships, the second being the US-based Indy Autonomous Challenge. However, it was a shame fans were unable to follow the live stream on YouTube as promised. Unlike the IAC, which primarily focuses on time trials and simulated races, the A2RL's car races are closer to a standard grand prix formula race format. Both use Dallara-supplied racecars; the IAC uses the AV-24 chassis derived from Indy NXT's IL-15, while the A2RL chassis is designated EAV-24 and is derived from the SF-23 chassis used in Japanese Super Formula races. === Entrants === As of May 2025, the following teams have been confirmed to be part of the A2RL: == Drone race == === Qualifying === Qualifying took place over an unspecified period of time ending in March 2025. 14 teams qualified. === Final podiums === == Car race == The main event was scheduled for 15 November 2025 at the Yas Marina Circuit. === Pre-season testing === Pre-season testing took place in early 2025. According to the organizers, over 300 terabytes of data were gathered and 1640 laps were logged between all teams. === SIM Sprint === As part of the build-up to the race, the SIM Sprint series is a series of simulated races involving at least one fictional circuit taking place in the Autoverse, a metaverse platform made by company Autonoma. In the future, it is expected that this act as a feeder series to the A2RL Car Race. ==== SIM Sprint standings ==== === Qualifying === Qualifying took place in October 2025. The top 6 in the 3-kilometer short-course time trials qualified for the main race. ==== Qualifying report ==== Once the qualifying cars were determined, there were a pair of sprint races to set the grid for the main event. One race was disputed by the top three qualifying teams and determined the pole-sitting car and the other two cars' starting positions, the other was disputed among the teams that scored P4 though P6 in the time trials and determined the remaining grid positions. ==== Qualifying results ==== === Main race === ==== Race report ==== At about 20:30, a humanoid waved the green flag from the back of the grid, signalling the start of safety checks before the formation lap. It was a rolling start. On Lap 1, just a few corners after crossing the line, Hailey (for team Technical University of Munich, or TUM) and Gianna (for team Unimore) quickly pushed out front, with what the commentators described as “aggressive” from Gianna. On Lap 2 at Turn 6, Gianna dives up the inside of Hailey to take the lead. Hailey takes evasive action and slows down slightly. At the end of Lap 6/start of Lap 7, both Gianna and Hailey lap slow-moving Constructor AI (for Constructor University), now 35 seconds behind Eva (team PoliMove). Gianna was slowed down by Constructor AI, causing Hailey to close the gap to Gianna. On Lap 12, while trying to lap Constructor AI again and simultaneously defend from Hailey, Gianna rear-ended Constructor AI, causing Gianna to run into the barriers at Turn 1 and both cars to retire. This brought out a red flag, followed by a Full Course Yellow. During the Full Course Yellow, on Lap 13, Turn 5, Sparkz (for team Kinetiz) span, presumably from cold tyre temperatures (a big concern after 2024's race), and dropping from second place down to fourth and last of the remaining cars. On Lap 15, the green flag was shown, and the race was resumed. On Lap 20, Hailey took the chequered flag and won the race for team TUM, as they did in 2024. Musa for TII Racing came second, over 47 seconds behind Hailey. Eva for PoliMove finished third. ==== Final race classification ==== Source:
Legal Knowledge Interchange Format
The Legal Knowledge Interchange Format (LKIF) was developed in the European ESTRELLA project and was designed with the goal of becoming a standard for representing and interchanging policy, legislation and cases, including their justificatory arguments, in the legal domain. LKIF builds on and uses the Web Ontology Language (OWL) for representing concepts and includes a reusable basic ontology of legal concepts. The core of LKIF consists of a combination of OWL-DL and SWRL. LKIF was designed with two main roles in mind: the translation of legal knowledge bases written in different representation formats and formalisms and to be a knowledge representation formalism which could be part of larger architectures for developing legal knowledge systems.
Phase stretch transform
Phase stretch transform (PST) is a computational approach to signal and image processing. One of its utilities is for feature detection and classification. PST is related to time stretch dispersive Fourier transform. It transforms the image by emulating propagation through a diffractive medium with engineered 3D dispersive property (refractive index). The operation relies on symmetry of the dispersion profile and can be understood in terms of dispersive eigenfunctions or stretch modes. PST performs similar functionality as phase-contrast microscopy, but on digital images. PST can be applied to digital images and temporal (time series) data. It is a physics-based feature engineering algorithm. == Operation principle == Here the principle is described in the context of feature enhancement in digital images. The image is first filtered with a spatial kernel followed by application of a nonlinear frequency-dependent phase. The output of the transform is the phase in the spatial domain. The main step is the 2-D phase function which is typically applied in the frequency domain. The amount of phase applied to the image is frequency dependent, with higher amount of phase applied to higher frequency features of the image. Since sharp transitions, such as edges and corners, contain higher frequencies, PST emphasizes the edge information. Features can be further enhanced by applying thresholding and morphological operations. PST is a pure phase operation whereas conventional edge detection algorithms operate on amplitude. == Physical and mathematical foundations of phase stretch transform == Photonic time stretch technique can be understood by considering the propagation of an optical pulse through a dispersive fiber. By disregarding the loss and non-linearity in fiber, the non-linear Schrödinger equation governing the optical pulse propagation in fiber upon integration reduces to: E o ( z , t ) = 1 2 π ∫ − ∞ ∞ E ~ i ( 0 , ω ) ⋅ e − i β 2 z ω 2 2 ⋅ e i ω t d ω {\displaystyle E_{o}(z,t)={\frac {1}{2\pi }}\int _{-\infty }^{\infty }{\tilde {E}}_{i}(0,\omega )\cdot e^{\frac {-i\beta _{2}z\omega ^{2}}{2}}\cdot e^{i\omega {t}}\,d\omega } (1) where β 2 {\displaystyle \beta _{2}} = GVD parameter, z is propagation distance, E o ( z , t ) {\displaystyle E_{o}(z,t)} is the reshaped output pulse at distance z and time t. The response of this dispersive element in the time-stretch system can be approximated as a phase propagator as presented in H ( ω ) = e i φ ( ω ) = e i ∑ m = 0 ∞ φ m ( ω ) = ∏ m = 0 ∞ H m ( ω ) {\displaystyle H(\omega )=e^{i\varphi (\omega )}=e^{i\sum _{m=0}^{\infty }\varphi _{m}(\omega )}=\prod _{m=0}^{\infty }H_{m}(\omega )} (2) Therefore, Eq. 1 can be written as following for a pulse that propagates through the time-stretch system and is reshaped into a temporal signal with a complex envelope given by E o ( t ) = 1 2 π ∫ − ∞ ∞ E ~ i ( ω ) ⋅ H ( ω ) ⋅ e i ω t d ω {\displaystyle E_{o}(t)={\frac {1}{2\pi }}\int _{-\infty }^{\infty }{\tilde {E}}_{i}(\omega )\cdot H(\omega )\cdot e^{i\omega t}\,d\omega } (3) The time stretch operation is formulated as generalized phase and amplitude operations, S { E i ( t ) } = ∫ − ∞ + ∞ F { E i ( t ) } ⋅ e i φ ( ω ) ⋅ L ~ ( ω ) ⋅ e i ω t d ω {\displaystyle \mathbb {S} \{E_{i}(t)\}=\int _{-\infty }^{+\infty }{\mathcal {F}}\{E_{i}(t)\}\cdot e^{i\varphi (\omega )}\cdot {\tilde {L}}(\omega )\cdot e^{i\omega {t}}d\omega } (4) where e i φ ( ω ) {\displaystyle e^{i\varphi (\omega )}} is the phase filter and L ~ ( ω ) {\displaystyle {\tilde {L}}(\omega )} is the amplitude filter. Next the operator is converted to discrete domain, S { E i [ n ] } = 1 N ∑ u = 0 N − 1 F F T { E i ( n ) } ⋅ K ~ ( u ) ⋅ L ~ ( u ) ⋅ e i 2 π N u n {\displaystyle \mathbb {S} \{E_{i}[n]\}={\frac {1}{N}}\sum _{u=0}^{N-1}FFT\{E_{i}(n)\}\cdot {\tilde {K}}(u)\cdot {\tilde {L}}(u)\cdot e^{i{\frac {2\pi }{N}}un}} (5) where u {\displaystyle u} is the discrete frequency, K ~ ( u ) {\displaystyle {\tilde {K}}(u)} is the phase filter, L ~ ( u ) {\displaystyle {\tilde {L}}(u)} is the amplitude filter and FFT is fast Fourier transform. The stretch operator S { } {\displaystyle \mathbb {S} \{\}} for a digital image is then S { E i [ n , m ] } = 1 M N ∑ v = 0 N − 1 ∑ u = 0 M − 1 F F T 2 { E i ( n , m ) } ⋅ K ~ ( u , v ) ⋅ L ~ ( u , v ) ⋅ e i 2 π M u m ⋅ e i 2 π N v n {\displaystyle \mathbb {S} \{E_{i}[n,m]\}={\frac {1}{MN}}\sum _{v=0}^{N-1}\sum _{u=0}^{M-1}FFT^{2}\{E_{i}(n,m)\}\cdot {\tilde {K}}(u,v)\cdot {\tilde {L}}(u,v)\cdot e^{i{\frac {2\pi }{M}}um}\cdot e^{i{\frac {2\pi }{N}}vn}} (6) In the above equations, E i [ n , m ] {\displaystyle E_{i}[n,m]} is the input image, n {\displaystyle n} and m {\displaystyle m} are the spatial variables, F F T 2 {\displaystyle FFT^{2}} is the two-dimensional fast Fourier transform, and u {\displaystyle u} and v {\displaystyle v} are spatial frequency variables. The function K ~ ( u , v ) {\displaystyle {\tilde {K}}(u,v)} is the warped phase kernel and the function L ~ ( u , v ) {\displaystyle {\tilde {L}}(u,v)} is a localization kernel implemented in frequency domain. PST operator is defined as the phase of the Warped Stretch Transform output as follows P S T { E i [ n , m ] } ≜ ∡ { S { E i [ x , y ] } } {\displaystyle PST\{E_{i}[n,m]\}\triangleq \measuredangle \{\mathbb {S} \{E_{i}[x,y]\}\}} (7) where ∡ { } {\displaystyle \measuredangle \{\}} is the angle operator. == PST kernel implementation == The warped phase kernel K ~ ( u , v ) {\displaystyle {\tilde {K}}(u,v)} can be described by a nonlinear frequency dependent phase K ~ ( u , v ) = e i φ ( u , v ) {\displaystyle {\tilde {K}}(u,v)=e^{i\varphi (u,v)}} While arbitrary phase kernels can be considered for PST operation, here we study the phase kernels for which the kernel phase derivative is a linear or sublinear function with respect to frequency variables. A simple example for such phase derivative profiles is the inverse tangent function. Consider the phase profile in the polar coordinate system φ ( u , v ) = φ polar ( r , θ ) = φ polar ( r ) {\displaystyle \varphi (u,v)=\varphi _{\text{polar}}(r,\theta )=\varphi _{\text{polar}}(r)} From d φ ( r ) d r = tan − 1 ( r ) {\displaystyle {\frac {d\varphi (r)}{dr}}=\tan ^{-1}(r)} we have φ ( r ) = r tan − 1 ( r ) − 1 2 log ( r 2 + 1 ) {\displaystyle \varphi (r)=r\tan ^{-1}(r)-{\frac {1}{2}}\log(r^{2}+1)} Therefore, the PST kernel is implemented as φ ( r ) = S ⋅ ( W r ) ⋅ tan − 1 ( W r ) − 1 2 log ( 1 + ( W r ) 2 ) ( W r max ) ⋅ tan − 1 ( W r max ) − 1 2 log ( 1 + ( W r max ) 2 ) {\displaystyle \varphi (r)=S\cdot {\frac {(Wr)\cdot \tan ^{-1}(Wr)-{\frac {1}{2}}\log(1+(Wr)^{2})}{(Wr_{\max })\cdot \tan ^{-1}(Wr_{\max })-{\frac {1}{2}}\log(1+(Wr_{\max })^{2})}}} where S {\displaystyle S} and W {\displaystyle W} are real-valued numbers related to the strength and warp of the phase profile == Applications == PST has been used for edge detection in biological and biomedical images as well as synthetic-aperture radar (SAR) image processing, as well as detail and feature enhancement for digital images. PST has also been applied to improve the point spread function for single molecule imaging in order to achieve super-resolution. The transform exhibits intrinsic superior properties compared to conventional edge detectors for feature detection in low contrast visually impaired images. The PST function can also be performed on 1-D temporal waveforms in the analog domain to reveal transitions and anomalies in real time. == Open source code release == On February 9, 2016, a UCLA Engineering research group has made public the computer code for PST algorithm that helps computers process images at high speeds and "see" them in ways that human eyes cannot. The researchers say the code could eventually be used in face, fingerprint, and iris recognition systems for high-tech security, as well as in self-driving cars' navigation systems or for inspecting industrial products. The Matlab implementation for PST can also be downloaded from Matlab Files Exchange. However, it is provided for research purposes only, and a license must be obtained for any commercial applications. The software is protected under a US patent. The code was then significantly refactored and improved to support GPU acceleration. In May 2022, it became one algorithm in PhyCV: the first physics-inspired computer vision library.
Whisper (speech recognition system)
Whisper is a machine learning model for speech recognition and transcription, created by OpenAI and first released as open-source software in September 2022. It is capable of transcribing speech in English and multiple other languages, and can translate several non-English languages into English. Whisper is a weakly-supervised deep learning acoustic model, made using an encoder-decoder transformer architecture. OpenAI claims that the combination of different training data and post-training filtering used in its development has led to improved recognition of accents, background noise, and jargon compared to previous approaches. While the model does not outperform larger, more specialized models and still experiences AI hallucination, it has been showed to be useful for general sound recognition and has many applications across different industries. == Background == Speech recognition has had a long history in research; the first approaches made use of statistical methods, such as dynamic time warping, and later hidden Markov models. At around the 2010s, deep neural network approaches became more common for speech recognition models, which were enabled by the availability of large datasets ("big data") and increased computational performance. Early approaches to deep learning in speech recognition included convolutional neural networks, which were limited due to their inability to capture sequential data, which later led to developments of Seq2seq approaches, which include recurrent neural networks, which made use of long short-term memory. Transformers, introduced in 2017 by Google, displaced many prior state-of-the-art approaches across a wide range in machine learning, and started becoming the core neural architecture in fields such as language modeling and computer vision. Weakly-supervised approaches to training acoustic models were recognized in the early 2020s as promising for speech recognition approaches using deep neural networks. According to a NYT report, in 2021 OpenAI believed they exhausted sources of higher-quality data to train their large language models and decided to complement scraped web text with transcriptions of YouTube videos and podcasts, and developed Whisper to solve this task. Whisper Large V2 was released on December 8, 2022, followed by Whisper Large V3 being released in November 2023, during the OpenAI Dev Day. In March 2025, OpenAI released new transcription models based on GPT-4o and GPT-4o mini, both of which have lower error rates than Whisper. == Architecture == The Whisper architecture is based on an encoder-decoder transformer. Input audio is resampled to 16,000 Hertz (Hz) and converted to an 80-channel Log-magnitude Mel spectrogram using 25 ms windows with a 10 ms stride. The spectrogram is then normalized to a [-1, 1] range with near-zero mean. The encoder takes this Mel spectrogram as input and processes it. It first passes through two convolutional layers. Sinusoidal positional embeddings are added. It is then processed by a series of Transformer encoder blocks (with pre-activation residual connections). The encoder's output is layer normalized. The decoder is a standard transformer decoder. It has the same width and Transformer blocks as the encoder. It uses learned positional embeddings and tied input-output token representations (using the same weight matrix for both the input and output embeddings). It uses a byte-pair encoding tokenizer, of the same kind as used in GPT-2. English-only models use the GPT-2 vocabulary, while multilingual models employ a re-trained multilingual vocabulary with the same number of words. Special tokens are used to allow the decoder to perform multiple tasks: Tokens that denote language (one unique token per language). Tokens that specify task (<|transcribe|> or <|translate|>). Tokens that specify if no timestamps are present (<|notimestamps|>). If the token is not present, then the decoder predicts timestamps relative to the segment, and quantized to 20 ms intervals. <|nospeech|> for voice activity detection. <|startoftranscript|>, and <|endoftranscript|> . Any text that appears before <|startoftranscript|> is not generated by the decoder, but given to the decoder as context. Loss is only computed over non-contextual parts of the sequence, i.e. tokens between these two special tokens. == Training data == The training dataset consists of 680,000 hours of labeled audio-transcript pairs sourced from the internet using semi-supervised learning. This includes 117,000 hours in 96 non-English languages and 125,000 hours of X→English translation data, where X stands for any non-English language. Preprocessing involved standardization of transcripts, filtering to remove machine-generated transcripts using heuristics (e.g., punctuation, capitalization), language identification and matching with transcripts, fuzzy deduplication, and deduplication with evaluation datasets to avoid data contamination. Speechless segments were also included to allow voice activity detection training. For the files still remaining after the filtering process, audio files were then broken into 30-second segments paired with the subset of the transcript that occurs within that time. If this predicted spoken language differed from the language of the text transcript associated with the audio, that audio-transcript pair was not used for training the speech recognition models, but instead for training translation. The model was trained using the AdamW optimizer with gradient norm clipping and a linear learning rate decay with warmup, with batch size 256 segments. Training proceeded for 1 million updates (approximately 2-3 epochs). No data augmentation or regularization, except for the Large V2 model, which used SpecAugment, Stochastic Depth, and BPE Dropout. The training used data parallelism with float16, dynamic loss scaling, and activation checkpointing. === Post-training filtering === After training the first model, researchers ran it on different subsets of the training data, each representing a distinct source. Data sources were ranked by a combination of their error rate and size. Manual inspection of the top-ranked sources (high error, large size) helped determine if the source was low quality (e.g., partial transcriptions, inaccurate alignment). After training, it was fine-tuned to suppress the prediction of speaker names and low-quality sources were then removed. == Capacity == While Whisper does not outperform models which specialize in the LibriSpeech dataset, when tested across many datasets, it is more robust and makes 55.2% fewer errors than other models. Whisper has a differing error rate with respect to transcribing different languages, with a higher word error rate in languages not well-represented in the training data. The authors found that multi-task learning improved overall performance compared to models specialized to one task. They conjectured that the best Whisper model trained is still underfitting the dataset, and larger models and longer training can result in better models. Third-party evaluations have found varying levels of AI hallucination. A study of transcripts of public meetings found hallucinations in eight out of every 10 transcripts, while an engineer discovered hallucinations in "about half" of 100 hours of transcriptions and a developer identified them in "nearly every one" of 26,000 transcripts. A study of 13,140 short audio segments (averaging 10 seconds) found 187 hallucinations (1.4%), 38% of which generated text that could be harmful because it inserted false references to things like race, non-existent medications, or violent events that were not in the audio. == Applications == The model has been used as the base for many applications, such as a unified model for speech recognition and more general sound recognition. Whisper has also been integrated into the workflow of biomedical research. In 2025, a study on Alzheimer's disease detection used the model to transcribe spontaneous speech recordings. The transcripts that were generated by the model were combined with LLM vector embeddings and traditional classifiers to help classify the patients' health. Another application is when OVALYTICS incorporated Whisper to transcribe YouTube videos and automate content moderation systems, which improved its detection of offensive content. The model has also been used in academic libraries and cultral heritage institutions to generate transcripts and captions for their digitized audiovisual collections. In a 2025 case study, Emory University Libraries found that Whisper reduced the labor used in transcription by around 30-35%, shifting work from text creation to text correction. However, human review is still necessary to make sure accuracy, formatting, and accessibility are all standard.