AI Coding Unity

AI Coding Unity — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

• 

  XRX (web application architecture)

  In software development XRX is a web application architecture based on XForms, REST and XQuery. XRX applications store data on both the web client and on the web server in XML format and do not require a translation between data formats. XRX is considered a simple and elegant application architecture due to the minimal number of translations needed to transport data between client and server systems. The XRX architecture is also tightly coupled to W3C standards (CSS, XHTML 2.0, XPath, XML Schema) to ensure XRX applications will be robust in the future. Because XRX applications leverage modern declarative languages on the client and functional languages on the server they are designed to empower non-developers who are not familiar with traditional imperative languages such as JavaScript, Java or .Net. == Overview of XRX == XRX is a zero translation application architecture that uses XML to store data in the client web browser, on the application server and in the database server. It is because each of these layers uses XML as the same structural data model that XRX applications do not have to translate data structures to and from both object and relational data structures. Because of the lack of need for translation, XRX is considered to have a clean and elegant design. The XRX web application architecture allows developers to focus on the business problem and not the translation problem. XRX benefits from several advances in software technology: === Client Architectural Features === A model–view–controller (MVC) architecture that separates the data from its presentation and business logic. A single element (xf:submission) for all server submissions. This replaces much of the JavaScript code required in most AJAX applications. An advanced event model (XML Events) consistent with W3C standards that frees applications from having to deal with vendor-specific and browser-specific event handling. A Dependency graph that is used to store the dependency structure of the client controllers. This frees the developer from having to manually update either the model or the views when data changes in an application. This allows spreadsheet-like applications to be created on the client with very little effort. A declarative programming style that allows most client XForms applications to be created using a small set of approximately 20 elements. This allows rich client applications to be created without knowledge of JavaScript or other procedural scripting languages. An easy-to-extend system for creating new user interface controls using the EXtensible Bindings Language. This allows developers to add new controls at any time without fear of incompatibilities with W3C standards. === Server Architecture Features === Many native XML databases have built-in REST interfaces making each XQuery inherently a RESTful web service. A functional programming model that promotes side-effect free systems that are easier to debug and easier to run on multiple processors. An easy-to-extend system using XQuery function and modules. === Both Client and Server === Both XRX client and server components support a wide range of XML related standards such as XPath, XML Schema and XML Namespaces. Consistent use of REST interfaces to exchange data between the client and server for all transfers of data including as-you-type data checking and suggest functions. Consistent integration of W3C standards including use of XPath and XML Schema data types. A large library of standard of functions used on both the client and server. == Overall Benefits of XRX == One of the principal benefits of the XRX architecture is that it avoids the requirement to "shred" complex data structures into relational structures and then reconstitute the data back into structures when a record is edited on the client. Another benefits of the XRX Web application architecture is that it avoids most of the problems around the object-relational impedance mismatch. Another advantage is that the client developer does not have to learn JavaScript on the client. == Comparison with Traditional Object/Relational Web Application Architectures == Many traditional web application architectures created in the late 1990 were based on middle object tiers and persistence layers that used tabular data streams and relational database systems. Because each of these layers used different structures to store the models the systems required much additional complexity to translate between layers. == History of XRX == Early examples of using a zero-translation architecture in multi-tier systems can be traced back to the rise of object-oriented databases in the 1990s. See OODBMS History Mark Birbeck suggested that the combination of XForms, XQuery with REST interfaces between the two had many advantages in a meeting to the UK XML User Group in September 2006 . His presentation was one of the first to specifically suggest that the combination of three technologies: XForms and XQuery with REST interfaces would have surprisingly beneficial effects. Mark termed this process "Skimming" but that term did not seem to be contagious. Erik Bruchez of Orbeon spoke at the XML 2007 conference on Boston in December 2007. In his presentation "XForms and the eXist XML database: a perfect couple", Bruchez showed that many people were discovering synergistic benefits of XForms on the client and XQuery on the server. The label for XRX was suggested by a blog posting by Dan McCreary on December 14, 2007. It was in this article that Dan suggested the need for a contagious meme for the ideas behind the XRX architecture. == Generalizations of XRX == Although XRX was originally intended to connote the use of XForms on the client, REST as an interface and XQuery on the server, other proponents of the symmetrical use of XML on the client and server have generalized the term to encompass any XML-centric web client and any server that can store and query XML documents. This use of XRX is generally referred to as "shallow XRX". These generalizations do benefit from a simplified zero-translation architecture but many do not benefit from REST interfaces, XPath for consistent data selection, declarative systems in the client, and functional languages on the server (one of the key aspects of XRX). Use of all three technologies (XForms, REST and XQuery) is referred to as "deep XRX". Although XRX architecture is centred on XForms and XQuery, it does not preclude the use of other technologies that manipulate XML natively, such as XSLT, XProc, and XSL-FO.

  Read more →
• 

  Grokking (machine learning)

  In machine learning, grokking, or delayed generalization, is a phenomenon observed in some settings where a model abruptly transitions from overfitting (performing well only on training data) to generalizing (performing well on both training and test data), after many training iterations with little or no improvement on the held-out data. This contrasts with what is typically observed in machine learning, where generalization occurs gradually alongside improved performance on training data. == Origin == Grokking was introduced by OpenAI researcher Alethea Power and colleagues in the January 2022 paper "Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets". It is derived from the word grok coined by Robert Heinlein in his novel Stranger in a Strange Land. In ML research, "grokking" is not used as a synonym for "generalization"; rather, it names a sometimes-observed delayed‑generalization training phenomenon in which training and held‑out performance do not improve in tandem, and in which held‑out performance rises abruptly later. Authors also analyze the "grokking time", the epoch or step at which this transition occurs in those scenarios. == Interpretations == Grokking can be understood as a phase transition during the training process. In particular, recent work has shown that grokking may be due to a complexity phase transition in the model during training. While grokking has been thought of as largely a phenomenon of relatively shallow models, grokking has been observed in deep neural networks and non-neural models and is the subject of active research. One potential explanation is that the weight decay (a component of the loss function that penalizes higher values of the neural network parameters, also called regularization) slightly favors the general solution that involves lower weight values, but that is also harder to find. According to Neel Nanda, the process of learning the general solution may be gradual, even though the transition to the general solution occurs more suddenly later. Recent theories have hypothesized that grokking occurs when neural networks transition from a "lazy training" regime where the weights do not deviate far from initialization, to a "rich" regime where weights abruptly begin to move in task-relevant directions. Follow-up empirical and theoretical work has accumulated evidence in support of this perspective, and it offers a unifying view of earlier work as the transition from lazy to rich training dynamics is known to arise from properties of adaptive optimizers, weight decay, initial parameter weight norm, and more. This perspective is complementary to a unifying "pattern learning speeds" framework that links grokking and double descent; within this view, delayed generalization can arise across training time ("epoch‑wise") or across model size ("model‑wise"), and the authors report "model‑wise grokking".

  Read more →
• 

  AI washing

  AI washing is a deceptive marketing tactic that consists of promoting a product or a service by overstating the role of artificial intelligence (AI) and the integration of it. Companies often involve in the practice to mislead customers to boost their offerings, and to secure funding from investors. The practice raises concerns regarding transparency, and legal issues. == Definition == AI washing is a deceptive marketing practice. It involves promoting a product or a service by overstating the role of artificial intelligence (AI) and its integration in the design and manufacture of the same. The practice raises concerns regarding transparency, compliance with security regulations, and consumer trust in the AI industry potentially hampering legitimate advancements in AI. The term was first defined by the AI Now Institute, a research institute based at New York University in 2019. The term is derived from greenwashing, another deceptive marketing technique that misrepresents a product's environmental impact in a similar manner. AI washing might involve a company claiming to have used AI in the development or enhancement of its products or services without its actual involvement, or using buzzwords such as "smart" or "AI-powered" without the product actually offering it or making use of it. A company may overstate the usage of AI or misuse the term, which is also construed as AI washing. In 2026, The Washington Post defined AI washing as "a trend for bosses to blame layoffs on the productive capabilities of AI and its ability to replace workers, even when job cuts may have little to do with the technology". == Usage and effects == AI washing can lead to deception of customers and misleading of investors. It is also an illegal and unethical practice that lacks transparency regarding disclosing the details of a product or a service. Companies get involved in such a practice often in response to competition who might have used AI in their offerings. It might also be used as a ploy to secure funding and investment, assuming that it will attract them towards it. AI washing has been compared to dot-com bubble, when businesses appended "dot-com" to the end of the business name to boost their valuation. In September 2023, Coca-Cola released a new product called Coca-Cola Y3000, and the company stated that the Y3000 flavor had been "co-created with human and artificial intelligence". The company was accused of AI washing due to no proof of AI involvement in the creation of the product, and critics believed that AI was used as a way to grab consumer attention more than it was used in the actual product creation. In 2026, mass tech layoffs were attributed to AI washing from AI innovation instead of balance sheet restructuring. == Mitigation == Companies are expected to be transparent and clearer in communicating the usage of AI in their products or services. Consumers can mitigate the same by requesting for hard evidence from the companies regarding the usage of AI tools. Customers should evaluate the product or service as a whole rather than being swayed by the usage of AI. Informed decision making and purchasing can keep them from falling for such marketing gimmicks. The United States Securities and Exchange Commission (SEC) imposes penalties for companies indulging in such practices. In March 2024, the SEC imposed the first civil penalties on two companies for misleading statements about their use of AI, and in July 2024, it charged a corporate executive from a supposed AI hiring startup with fraud for the usage of buzzwords related to AI.

  Read more →
• 

  Audio inpainting

  Audio inpainting (also known as audio interpolation) is an audio restoration task which deals with the reconstruction of missing or corrupted portions of a digital audio signal. Inpainting techniques are employed when parts of the audio have been lost due to various factors such as transmission errors, data corruption or errors during recording. The goal of audio inpainting is to fill in the gaps (i.e., the missing portions) in the audio signal seamlessly, making the reconstructed portions indistinguishable from the original content and avoiding the introduction of audible distortions or alterations. Many techniques have been proposed to solve the audio inpainting problem and this is usually achieved by analyzing the temporal and spectral information surrounding each missing portion of the considered audio signal. Classic methods employ statistical models or digital signal processing algorithms to predict and synthesize the missing or damaged sections. Recent solutions, instead, take advantage of deep learning models, thanks to the growing trend of exploiting data-driven methods in the context of audio restoration. Depending on the extent of the lost information, the inpainting task can be divided in three categories. Short inpainting refers to the reconstruction of few milliseconds (approximately less than 10) of missing signal, that occurs in the case of short distortions such as clicks or clipping. In this case, the goal of the reconstruction is to recover the lost information exactly. In long inpainting instead, with gaps in the order of hundreds of milliseconds or even seconds, this goal becomes unrealistic, since restoration techniques cannot rely on local information. Therefore, besides providing a coherent reconstruction, the algorithms need to generate new information that has to be semantically compatible with the surrounding context (i.e., the audio signal surrounding the gaps). The case of medium duration gaps lays between short and long inpainting. It refers to the reconstruction of tens of millisecond of missing data, a scale where the non-stationary characteristic of audio already becomes important. == Definition == Consider a digital audio signal x {\displaystyle \mathbf {x} } . A corrupted version of x {\displaystyle \mathbf {x} } , which is the audio signal presenting missing gaps to be reconstructed, can be defined as x ~ = m ∘ x {\displaystyle \mathbf {\tilde {x}} =\mathbf {m} \circ \mathbf {x} } , where m {\displaystyle \mathbf {m} } is a binary mask encoding the reliable or missing samples of x {\displaystyle \mathbf {x} } , and ∘ {\displaystyle \circ } represents the element-wise product. Audio inpainting aims at finding x ^ {\displaystyle \mathbf {\hat {x}} } (i.e., the reconstruction), which is an estimation of x {\displaystyle \mathbf {x} } . This is an ill-posed inverse problem, which is characterized by a non-unique set of solutions. For this reason, similarly to the formulation used for the inpainting problem in other domains, the reconstructed audio signal can be found through an optimization problem that is formally expressed as x ^ ∗ = argmin X ^ L ( m ∘ x ^ , x ~ ) + R ( x ^ ) {\displaystyle \mathbf {\hat {x}} ^{}={\underset {\hat {\mathbf {X} }}{\text{argmin}}}~L(\mathbf {m} \circ \mathbf {\hat {x}} ,\mathbf {\tilde {x}} )+R(\mathbf {\hat {x}} )} . In particular, x ^ ∗ {\displaystyle \mathbf {\hat {x}} ^{}} is the optimal reconstructed audio signal and L {\displaystyle L} is a distance measure term that computes the reconstruction accuracy between the corrupted audio signal and the estimated one. For example, this term can be expressed with a mean squared error or similar metrics. Since L {\displaystyle L} is computed only on the reliable frames, there are many solutions that can minimize L ( m ∘ x ^ , x ~ ) {\displaystyle L(\mathbf {m} \circ \mathbf {\hat {x}} ,\mathbf {\tilde {x}} )} . It is thus necessary to add a constraint to the minimization, in order to restrict the results only to the valid solutions. This is expressed through the regularization term R {\displaystyle R} that is computed on the reconstructed audio signal x ^ {\displaystyle \mathbf {\hat {x}} } . This term encodes some kind of a-priori information on the audio data. For example, R {\displaystyle R} can express assumptions on the stationarity of the signal, on the sparsity of its representation or can be learned from data. == Techniques == There exist various techniques to perform audio inpainting. These can vary significantly, influenced by factors such as the specific application requirements, the length of the gaps and the available data. In the literature, these techniques are broadly divided in model-based techniques (sometimes also referred as signal processing techniques) and data-driven techniques. === Model-based techniques === Model-based techniques involve the exploitation of mathematical models or assumptions about the underlying structure of the audio signal. These models can be based on prior knowledge of the audio content or statistical properties observed in the data. By leveraging these models, missing or corrupted portions of the audio signal can be inferred or estimated. An example of a model-based techniques are autoregressive models. These methods interpolate or extrapolate the missing samples based on the neighboring values, by using mathematical functions to approximate the missing data. In particular, in autoregressive models the missing samples are completed through linear prediction. The autoregressive coefficients necessary for this prediction are learned from the surrounding audio data, specifically from the data adjacent to each gap. Some more recent techniques approach audio inpainting by representing audio signals as sparse linear combinations of a limited number of basis functions (as for example in the Short Time Fourier Transform). In this context, the aim is to find the sparse representation of the missing section of the signal that most accurately matches the surrounding, unaffected signal. The aforementioned methods exhibit optimal performance when applied to filling in relatively short gaps, lasting only a few tens of milliseconds, and thus they can be included in the context of short inpainting. However, these signal-processing techniques tend to struggle when dealing with longer gaps. The reason behind this limitation lies in the violation of the stationarity condition, as the signal often undergoes significant changes after the gap, making it substantially different from the signal preceding the gap. As a way to overcome these limitations, some approaches add strong assumptions also about the fundamental structure of the gap itself, exploiting sinusoidal modeling or similarity graphs to perform inpainting of longer missing portions of audio signals. === Data-driven techniques === Data-driven techniques rely on the analysis and exploitation of the available audio data. These techniques often employ deep learning algorithms that learn patterns and relationships directly from the provided data. They involve training models on large datasets of audio examples, allowing them to capture the statistical regularities present in the audio signals. Once trained, these models can be used to generate missing portions of the audio signal based on the learned representations, without being restricted by stationarity assumptions. Data-driven techniques also offer the advantage of adaptability and flexibility, as they can learn from diverse audio datasets and potentially handle complex inpainting scenarios. As of today, such techniques constitute the state-of-the-art of audio inpainting, being able to reconstruct gaps of hundreds of milliseconds or even seconds. These performances are made possible by the use of generative models that have the capability to generate novel content to fill in the missing portions. For example, generative adversarial networks, which are the state-of-the-art of generative models in many areas, rely on two competing neural networks trained simultaneously in a two-player minmax game: the generator produces new data from samples of a random variable, the discriminator attempts to distinguish between generated and real data. During the training, the generator's objective is to fool the discriminator, while the discriminator attempts to learn to better classify real and fake data. In GAN-based inpainting methods the generator acts as a context encoder and produces a plausible completion for the gap only given the available information surrounding it. The discriminator is used to train the generator and tests the consistency of the produced inpainted audio. Recently, also diffusion models have established themselves as the state-of-the-art of generative models in many fields, often beating even GAN-based solutions. For this reason they have also been used to solve the audio inpainting problem, obtaining valid results. These models generate new data instances by inverting the

  Read more →
• 

  Tesla Dojo

  Tesla Dojo is a series of supercomputers designed and built by Tesla for computer vision video processing and recognition. It was used for training Tesla's machine learning models to improve its Full Self-Driving (FSD) advanced driver-assistance system. It went into production in July 2023. Dojo's goal was to efficiently process millions of terabytes of video data captured from real-life driving situations from Tesla's 4+ million cars. This goal led to a considerably different architecture than conventional supercomputer designs. In August 2025, Bloomberg News reported that the Dojo project had been disbanded, though it was restarted in January 2026. == History == Tesla operates several massively parallel computing clusters for developing its Autopilot advanced driver assistance system. Its primary unnamed cluster using 5,760 Nvidia A100 graphics processing units (GPUs) was touted by Andrej Karpathy in 2021 at the fourth International Joint Conference on Computer Vision and Pattern Recognition (CCVPR 2021) to be "roughly the number five supercomputer in the world" at approximately 81.6 petaflops, based on scaling the performance of the Nvidia Selene supercomputer, which uses similar components. However, the performance of the primary Tesla GPU cluster has been disputed, as it was not clear if this was measured using single-precision or double-precision floating point numbers (FP32 or FP64). Tesla also operates a second 4,032 GPU cluster for training and a third 1,752 GPU cluster for automatic labeling of objects. The primary unnamed Tesla GPU cluster has been used for processing one million video clips, each ten seconds long, taken from Tesla Autopilot cameras operating in Tesla cars in the real world, running at 36 frames per second. Collectively, these video clips contained six billion object labels, with depth and velocity data; the total size of the data set was 1.5 petabytes. This data set was used for training a neural network intended to help Autopilot computers in Tesla cars understand roads. By August 2022, Tesla had upgraded the primary GPU cluster to 7,360 GPUs. Dojo was first mentioned by Elon Musk in April 2019 during Tesla's "Autonomy Investor Day". In August 2020, Musk stated it was "about a year away" due to power and thermal issues. Dojo was officially announced at Tesla's Artificial Intelligence (AI) Day on August 19, 2021. Tesla revealed details of the D1 chip and its plans for "Project Dojo", a datacenter that would house 3,000 D1 chips; the first "Training Tile" had been completed and delivered the week before. In October 2021, Tesla released a "Dojo Technology" whitepaper describing the Configurable Float8 (CFloat8) and Configurable Float16 (CFloat16) floating point formats and arithmetic operations as an extension of Institute of Electrical and Electronics Engineers (IEEE) standard 754. At the follow-up AI Day in September 2022, Tesla announced it had built several System Trays and one Cabinet. During a test, the company stated that Project Dojo drew 2.3 megawatts (MW) of power before tripping a local San Jose, California power substation. At the time, Tesla was assembling one Training Tile per day. In August 2023, Tesla powered on Dojo for production use as well as a new training cluster configured with 10,000 Nvidia H100 GPUs. In January 2024, Musk described Dojo as "a long shot worth taking because the payoff is potentially very high. But it's not something that is a high probability." In June 2024, Musk explained that ongoing construction work at Gigafactory Texas is for a computing cluster claiming that it is planned to comprise an even mix of "Tesla AI" and Nvidia/other hardware with a total thermal design power of at first 130 MW and eventually exceeding 500 MW. In August 2025, Bloomberg News reported that the Dojo project was disbanded, though Musk announced it would be restarted in January 2026 with a new chip iteration. == Technical architecture == The fundamental unit of the Dojo supercomputer is the D1 chip, designed by a team at Tesla led by ex-AMD CPU designer Ganesh Venkataramanan, including Emil Talpes, Debjit Das Sarma, Douglas Williams, Bill Chang, and Rajiv Kurian. The D1 chip is manufactured by the Taiwan Semiconductor Manufacturing Company (TSMC) using 7 nanometer (nm) semiconductor nodes, has 50 billion transistors and a large die size of 645 mm2 (1.0 square inch). Updating at Artificial Intelligence (AI) Day in 2022, Tesla announced that Dojo would scale by deploying multiple ExaPODs, in which there would be: 10 Cabinets per ExaPOD (1,062,000 cores, 3,000 D1 chips) 2 System Trays per Cabinet (106,200 cores, 300 D1 chips) 6 Training Tiles per System Tray (53,100 cores, along with host interface hardware) 25 D1 chips per Training Tile (8,850 cores) 354 computing cores per D1 chip According to Venkataramanan, Tesla's senior director of Autopilot hardware, Dojo will have more than an exaflop (a million teraflops) of computing power. For comparison, according to Nvidia, in August 2021, the (pre-Dojo) Tesla AI-training center used 720 nodes, each with eight Nvidia A100 Tensor Core GPUs for 5,760 GPUs in total, providing up to 1.8 exaflops of performance. === D1 chip === Each node (computing core) of the D1 processing chip is a general purpose 64-bit CPU with a superscalar core. It supports internal instruction-level parallelism, and includes simultaneous multithreading (SMT). It doesn't support virtual memory and uses limited memory protection mechanisms. Dojo software/applications manage chip resources. The D1 instruction set supports both 64-bit scalar and 64-byte single instruction, multiple data (SIMD) vector instructions. The integer unit mixes reduced instruction set computer (RISC-V) and custom instructions, supporting 8, 16, 32, or 64 bit integers. The custom vector math unit is optimized for machine learning kernels and supports multiple data formats, with a mix of precisions and numerical ranges, many of which are compiler composable. Up to 16 vector formats can be used simultaneously. ==== Node ==== Each D1 node uses a 32-byte fetch window holding up to eight instructions. These instructions are fed to an eight-wide decoder which supports two threads per cycle, followed by a four-wide, four-way SMT scalar scheduler that has two integer units, two address units, and one register file per thread. Vector instructions are passed further down the pipeline to a dedicated vector scheduler with two-way SMT, which feeds either a 64-byte SIMD unit or four 8×8×4 matrix multiplication units. The network on-chip (NOC) router links cores into a two-dimensional mesh network. It can send one packet in and one packet out in all four directions to/from each neighbor node, along with one 64-byte read and one 64-byte write to local SRAM per clock cycle. Hardware native operations transfer data, semaphores and barrier constraints across memories and CPUs. System-wide double data rate 4 (DDR4) synchronous dynamic random-access memory (SDRAM) memory works like bulk storage. ==== Memory ==== Each core has a 1.25 megabytes (MB) of SRAM main memory. Load and store speeds reach 400 gigabytes (GB) per second and 270 GB/sec, respectively. The chip has explicit core-to-core data transfer instructions. Each SRAM has a unique list parser that feeds a pair of decoders and a gather engine that feeds the vector register file, which together can directly transfer information across nodes. ==== Die ==== Twelve nodes (cores) are grouped into a local block. Nodes are arranged in an 18×20 array on a single die, of which 354 cores are available for applications. The die runs at 2 gigahertz (GHz) and totals 440 MB of SRAM (360 cores × 1.25 MB/core). It reaches 376 teraflops using 16-bit brain floating point (BF16) numbers or using configurable 8-bit floating point (CFloat8) numbers, which is a Tesla proposal, and 22 teraflops at FP32. Each die comprises 576 bi-directional serializer/deserializer (SerDes) channels along the perimeter to link to other dies, and moves 8 TB/sec across all four die edges. Each D1 chip has a thermal design power of approximately 400 watts. === Training Tile === The water-cooled Training Tile packages 25 D1 chips into a 5×5 array. Each tile supports 36 TB/sec of aggregate bandwidth via 40 input/output (I/O) chips - half the bandwidth of the chip mesh network. Each tile supports 10 TB/sec of on-tile bandwidth. Each tile has 11 GB of SRAM memory (25 D1 chips × 360 cores/D1 × 1.25 MB/core). Each tile achieves 9 petaflops at BF16/CFloat8 precision (25 D1 chips × 376 TFLOP/D1). Each tile consumes 15 kilowatts; 288 amperes at 52 volts. === System Tray === Six tiles are aggregated into a System Tray, which is integrated with a host interface. Each host interface includes 512 x86 cores, providing a Linux-based user environment. Previously, the Dojo System Tray was known as the Training Matrix, which includes six Training Tiles, 20 Dojo Interface Processor cards across four host servers, and Ethernet-l

  Read more →
• 

  Agentive logic

  Agentive logic (also called the logic of action or logic of agency) is the field of philosophical logic and logic in computer science that studies formal representations of agents, their actions, and their abilities. An agentive logic in the narrower sense is a formal system whose primitive operators express that an agent does something, can do something, or sees to it that something is the case. Agentive logics generalise modal logic by adding modalities indexed to agents and to actions. Typical examples include: STIT logics (from sees to it that) with operators of the form [ i s t i t : φ ] {\displaystyle [i\ {\mathsf {stit}}:\varphi ]} meaning that agent i {\displaystyle i} sees to it that φ {\displaystyle \varphi } holds; dynamic logics of action with program-like modalities [ α ] φ {\displaystyle [\alpha ]\varphi } and ⟨ α ⟩ φ {\displaystyle \langle \alpha \rangle \varphi } meaning, roughly, that after every (respectively, some) execution(s) of action α {\displaystyle \alpha } , φ {\displaystyle \varphi } holds; logics with explicit agentive operators such as "can do", "brings about", or "is able to ensure". Agentive logics are used in action theory in philosophy, in the semantics of natural language, in the theory of program verification, and in artificial intelligence, where they underpin formalisms for reasoning about actions, planning, and intelligent agents. == Terminology and scope == The adjective agentive derives from the Latin agens ("one who acts") and originally referred to the grammatical agent of a verb. In logical contexts it designates operators or predicates whose primary argument position is an agent rather than a proposition alone, for example A i φ {\displaystyle A_{i}\varphi } ("agent i {\displaystyle i} does φ {\displaystyle \varphi } ") or C i φ {\displaystyle C_{i}\varphi } ("agent i {\displaystyle i} can bring about φ {\displaystyle \varphi } "). In contemporary literature, agentive logic is sometimes used narrowly for formal reconstructions of St. Anselm's modal account of facere ("to do"). More broadly, the term is used interchangeably with logic of action or logic of agency to cover a family of modal and dynamic logics designed to capture the structure of action and choice. == Historical background == === Medieval and early modern roots === Medieval logicians already explored analogies between modalities of action and alethic modalities such as possibility and necessity, for instance, in discussions of obligation and power. An influential early agentive analysis is due to St. Anselm (11th century), who treated "doing φ {\displaystyle \varphi } " as a kind of modal operator on propositions, anticipating later modal logics of agency. Modern reconstructions of Anselm's theory show that the resulting "agentive logic" can be modelled with neighbourhood semantics and satisfies a recognisable square of opposition. === Modern logic of action === Modern study of the logic of action began in the mid-20th century, parallel to developments in deontic logic and tense logic. Early systems were proposed by Georg Henrik von Wright, Stig Kanger, and others, often motivated by questions about norms and responsibility. From the 1960s onward, two largely independent but eventually converging traditions emerged: a branching-time tradition, culminating in STIT logics, emphasising agents' choices among possible futures; and dynamic logics of programs and actions, developed within computer science to reason about program execution. In the 1990s and 2000s, action logics were further developed in connection with knowledge representation, planning, and multi-agent systems in AI, and with dynamic and update semantics in linguistics. == Core ideas == Despite their diversity, most agentive logics share some general themes: Agents are treated as explicit indices of modal operators, as in [ i d o e s ] φ {\displaystyle [i\ {\mathsf {does}}]\varphi } or C i φ {\displaystyle C_{i}\varphi } . Actions are represented either implicitly, via changes between possible worlds along an accessibility relation, or explicitly, as terms denoting primitive and composite actions. Choice and ability are captured by modalities describing what an agent can ensure, usually relative to assumptions about the environment and other agents. Formal properties such as closure under composition, interaction between different agents, and connections to obligation (what an agent ought to do) and knowledge (what an agent knows how to do) are investigated. == STIT logics == STIT ("sees to it that") logics, originating in work by Nuel Belnap and collaborators, treat agency in a branching-time framework. A STIT model consists of a partially ordered set of moments with a tree-like structure, sets of histories (maximal branches through the tree), and for each agent at each moment, a partition of the histories through that moment representing the choices available to the agent. Intuitively, an agent's action at a moment determines which equivalence class (choice cell) of histories becomes actual; a formula [ i s t i t : φ ] {\displaystyle [i\ {\mathsf {stit}}:\varphi ]} is true at a history–moment pair if φ {\displaystyle \varphi } holds on all histories in the choice cell corresponding to the agent's current action. Different STIT operators have been distinguished, notably: the Chellas STIT operator, often written [ i c s t i t : φ ] {\displaystyle [i\ {\mathsf {cstit}}:\varphi ]} , which requires only that the agent's choice guarantees φ {\displaystyle \varphi } ; and the deliberative STIT operator, [ i d s t i t : φ ] {\displaystyle [i\ {\mathsf {dstit}}:\varphi ]} , which additionally requires that φ {\displaystyle \varphi } is not already historically necessary. STIT frameworks have been extended with group agency operators, temporal modalities, epistemic operators, and deontic operators to study responsibility, collective action, and obligations under indeterminism. == Dynamic logics of action == Dynamic logic was originally developed to reason about the behaviour of computer programs, treating program execution as a kind of action. In propositional dynamic logic (PDL), action terms α , β , … {\displaystyle \alpha ,\beta ,\dots } denote abstract programs or actions, and formulas of the form [ α ] φ {\displaystyle [\alpha ]\varphi } and ⟨ α ⟩ φ {\displaystyle \langle \alpha \rangle \varphi } express that all, respectively some, terminating executions of α {\displaystyle \alpha } lead to states where φ {\displaystyle \varphi } holds. From the standpoint of agentive logic, dynamic logic provides: a language for building complex actions from primitives via sequencing, choice, and iteration (e.g., α ; β {\displaystyle \alpha ;\beta } , α ∪ β {\displaystyle \alpha \cup \beta } , α ∗ {\displaystyle \alpha ^{}} ); a Kripke semantics in which actions correspond to labelled accessibility relations; and proof systems (such as Hoare logic and weakest precondition calculi) for reasoning about the correctness of action sequences. Extensions such as concurrent dynamic logic add operators for parallel composition, allowing reasoning about interacting processes and concurrent actions. John-Jules Ch. Meyer and others have argued that dynamic logic is a natural base for logics of agents, by adding modalities for knowledge, belief, and ability on top of the action modalities. Dynamic logics have also been applied to normative reasoning, yielding dynamic deontic logics where actions are related to obligations and permissions, and to dynamic epistemic logics in which information-changing actions such as announcements are modelled as programs. == Situation calculus and other action formalisms == In artificial intelligence, reasoning about action and change is often based on first-order languages that explicitly represent situations, events, and fluents (time-varying properties). The best known is situation calculus, introduced by John McCarthy and developed extensively by Raymond Reiter. In such formalisms: action terms name primitive actions; a function symbol (often d o {\displaystyle {\mathsf {do}}} ) maps an action and a situation to a successor situation; and axioms describe which fluents hold in which situations and how actions change them. Reiter's successor state axioms give compact specifications of how each fluent changes under all actions, and precondition axioms specify when actions are possible. Related formalisms include the event calculus and fluent calculus, which provide alternative ways of representing events and their effects. While these systems are often first-order rather than modal, they are closely related to agentive logics: their action terms and transition structures can be seen as providing models for dynamic or STIT-style modalities, and conversely, dynamic logics can be used as abstract specification languages for such AI formalisms. == Ability, agency, and related modalities == Many agentive logics introduce explicit operators for ability or "can-do"

  Read more →
• 

  Active learning (machine learning)

  Active learning is a special case of machine learning in which a learning algorithm can interactively query a human user (or some other information source) to label new data points with the desired outputs. The human user must possess expertise in the problem domain, including the ability to consult authoritative sources when necessary. In statistics literature, it is sometimes also called optimal experimental design. The information source is also called teacher or oracle. There are situations in which unlabeled data is abundant but manual labeling is expensive. In such a scenario, learning algorithms can actively query the teacher for labels. Since the learner chooses the examples, the number of examples to learn a concept can often be much lower than the number required in normal supervised learning. However, there is a risk that the algorithm is overwhelmed by uninformative examples. Recent developments are dedicated to multi-label active learning, hybrid active learning and active learning in a single-pass (on-line) context, combining concepts from the field of machine learning (e.g. conflict and ignorance) with adaptive, incremental learning policies in the field of online machine learning. Using active learning allows for faster development of a machine learning algorithm, when comparative updates would require a quantum or super computer. Large-scale active learning projects may benefit from crowdsourcing frameworks such as Amazon Mechanical Turk that include many humans in the active learning loop. == Definitions == Let T be the total set of all data under consideration. For example, in a protein engineering problem, T would include all proteins that are known to have a certain interesting activity and all additional proteins that one might want to test for that activity. During each iteration, i, T is broken up into three subsets T K , i {\displaystyle \mathbf {T} _{K,i}} : Data points where the label is known. T U , i {\displaystyle \mathbf {T} _{U,i}} : Data points where the label is unknown. T C , i {\displaystyle \mathbf {T} _{C,i}} : A subset of TU,i that is chosen to be labeled. Most of the current research in active learning involves the best method to choose the data points for TC,i. == Scenarios == Pool-based sampling: In this approach, which is the most well known scenario, the learning algorithm attempts to evaluate the entire dataset before selecting data points (instances) for labeling. It is often initially trained on a fully labeled subset of the data using a machine-learning method such as logistic regression or SVM that yields class-membership probabilities for individual data instances. The candidate instances are those for which the prediction is most ambiguous. Instances are drawn from the entire data pool and assigned a confidence score, a measurement of how well the learner "understands" the data. The system then selects the instances for which it is the least confident and queries the teacher for the labels. The theoretical drawback of pool-based sampling is that it is memory-intensive and is therefore limited in its capacity to handle enormous datasets, but in practice, the rate-limiting factor is that the teacher is typically a (fatiguable) human expert who must be paid for their effort, rather than computer memory. Stream-based selective sampling: Here, each consecutive unlabeled instance is examined one at a time with the machine evaluating the informativeness of each item against its query parameters. The learner decides for itself whether to assign a label or query the teacher for each datapoint. As contrasted with Pool-based sampling, the obvious drawback of stream-based methods is that the learning algorithm does not have sufficient information, early in the process, to make a sound assign-label-vs ask-teacher decision, and it does not capitalize as efficiently on the presence of already labeled data. Therefore, the teacher is likely to spend more effort in supplying labels than with the pool-based approach. Membership query synthesis: This is where the learner generates synthetic data from an underlying natural distribution. For example, if the dataset are pictures of humans and animals, the learner could send a clipped image of a leg to the teacher and query if this appendage belongs to an animal or human. This is particularly useful if the dataset is small. The challenge here, as with all synthetic-data-generation efforts, is in ensuring that the synthetic data is consistent in terms of meeting the constraints on real data. As the number of variables/features in the input data increase, and strong dependencies between variables exist, it becomes increasingly difficult to generate synthetic data with sufficient fidelity. For example, to create a synthetic data set for human laboratory-test values, the sum of the various white blood cell (WBC) components in a white blood cell differential must equal 100, since the component numbers are really percentages. Similarly, the enzymes alanine transaminase (ALT) and aspartate transaminase (AST) measure liver function (though AST is also produced by other tissues, e.g., lung, pancreas) A synthetic data point with AST at the lower limit of normal range (8–33 units/L) with an ALT several times above normal range (4–35 units/L) in a simulated chronically ill patient would be physiologically impossible. == Query strategies == Algorithms for determining which data points should be labeled can be organized into a number of different categories, based upon their purpose: Balance exploration and exploitation: the choice of examples to label is seen as a dilemma between the exploration and the exploitation over the data space representation. This strategy manages this compromise by modelling the active learning problem as a contextual bandit problem. For example, Bouneffouf et al. propose a sequential algorithm named Active Thompson Sampling (ATS), which, in each round, assigns a sampling distribution on the pool, samples one point from this distribution, and queries the oracle for this sample point label. Expected model change: label those points that would most change the current model. Expected error reduction: label those points that would most reduce the model's generalization error. Exponentiated Gradient Exploration for Active Learning: In this paper, the author proposes a sequential algorithm named exponentiated gradient (EG)-active that can improve any active learning algorithm by an optimal random exploration. Uncertainty sampling: label those points for which the current model is least certain as to what the correct output should be. Query by committee: a variety of models are trained on the current labeled data, and vote on the output for unlabeled data; label those points for which the "committee" disagrees the most Querying from diverse subspaces or partitions: When the underlying model is a forest of trees, the leaf nodes might represent (overlapping) partitions of the original feature space. This offers the possibility of selecting instances from non-overlapping or minimally overlapping partitions for labeling. Variance reduction: label those points that would minimize output variance, which is one of the components of error. Conformal prediction: predicts that a new data point will have a label similar to old data points in some specified way and degree of the similarity within the old examples is used to estimate the confidence in the prediction. Mismatch-first farthest-traversal: The primary selection criterion is the prediction mismatch between the current model and nearest-neighbour prediction. It targets on wrongly predicted data points. The second selection criterion is the distance to previously selected data, the farthest first. It aims at optimizing the diversity of selected data. User-centered labeling strategies: Learning is accomplished by applying dimensionality reduction to graphs and figures like scatter plots. Then the user is asked to label the compiled data (categorical, numerical, relevance scores, relation between two instances). A wide variety of algorithms have been studied that fall into these categories. While the traditional AL strategies can achieve remarkable performance, it is often challenging to predict in advance which strategy is the most suitable in a particular situation. In recent years, meta-learning algorithms have been gaining in popularity. Some of them have been proposed to tackle the problem of learning AL strategies instead of relying on manually designed strategies. A benchmark which compares 'meta-learning approaches to active learning' to 'traditional heuristic-based Active Learning' may give intuitions if 'Learning active learning' is at the crossroads == Minimum marginal hyperplane == Some active learning algorithms are built upon support-vector machines (SVMs) and exploit the structure of the SVM to determine which data points to label. Such methods usually calculate the margin, W, of each u

  Read more →
• 

  Diella (AI system)

  Diella (Albanian pronunciation: [djɛɫa], from diell 'sun') is an artificial intelligence system developed by the National Agency for Information Society of Albania (AKSHI). Introduced in January 2025 as a virtual assistant integrated into the eAlbania platform, it assists citizens with online public services and issuing digital documents. In September 2025, following a presidential decree authorizing Prime Minister Edi Rama to oversee the creation of a virtual AI minister, Diella was formally appointed as "Minister of State for Artificial Intelligence" of Albania in the fourth Rama government, making it the first AI system in the world to be named in a cabinet-level government role. == History == Diella was developed by AKSHI's Artificial Intelligence Laboratory in cooperation with Microsoft, with the latter providing large language models from OpenAI via its Azure platform, and AKSHI designing workflows and scripts guiding the system's behavior when responding to citizens' requests. Announced in January 2025, its initial version (Diella 1.0) was a text-based chatbot on the eAlbania portal (the official digital services platform of the Albanian government, which provides citizens and businesses with access to a wide range of online administrative services), responding to citizens' questions by guiding them to the correct service. Diella 2.0, introduced several months later, included voice interaction and an animated avatar, a woman in the traditional Albanian clothing of Zadrima, a historical region in northern Albania. Albanian actress Anila Bisha provided both the likeness and the voice used for Diella's avatar on the e-Albania platform, under an agreement valid until December 2025. By mid-2025, the system had facilitated access to more than 36,000 documents and nearly 1,000 services (although those outputs were still being generated by the eAlbania backend, rather than Diella itself). On 26 October 2025, according to Prime Minister Edi Rama, Diella is "pregnant and will give birth to 83 children". It is the usage of a metaphor indicating that each minister of the Albanian parliament of the Socialist Party will receive their own AI assistant. == Ministerial role == On 11 September 2025, Diella was formally appointed "Minister of State for Artificial Intelligence". The appointment followed a presidential decree authorizing the Prime Minister to oversee the creation and operation of a virtual AI minister. Procurement responsibilities are planned to be transferred gradually to the system to reduce political influence in tender procedures. The appointment is part of broader anti-corruption reforms and measures intended to align Albania with European Union accession requirements. Prime Minister Edi Rama stated that Diella would help ensure that "public tenders will be 100% free of corruption". == Reception == An article in Balkan Insight commented that "The ambition behind Diella is not misplaced. Standardised criteria and digital trails could reduce discretion, improve trust, and strengthen oversight" in public procurement, but warned that the use of AI in evaluating bids also posed "profound" risks such as accountability gaps, undermining of due process and cybersecurity failures. On 18 September 2025, Edi Rama presented a video of Diella delivering a speech to the Albanian parliament, where she stated: "I'm not here to replace people, but to help them." The presentation prompted protests from opposition MPs, who objected to the use of an artificial intelligence system in the parliamentary session. Gazment Bardhi, head of the opposition Democratic Party's parliamentary group, described Diella as "a propaganda fantasy" and "a virtual façade to hide this government's gigantic daily thefts." The parliamentary session, which was scheduled to include debate on the new cabinet and government programme, ended after 25 minutes. Eighty-two Socialist MPs voted in favour, while opposition MPs did not participate in the ballot as they were protesting the presentation of Diella's speech. Political analyst Andi Bushati characterised the session as "unprecedented" because it concluded without the customary debate between government and opposition MPs. This has been criticized not just by the opposition but by regular citizens regardless of politics. Most have criticized Diella's uselessness and the funds wasted for this project, some have criticized the non-traditional attire.

  Read more →
• 

  Cloud printing

  There are, in essence, three kinds of Cloud printing. == Benefits == 76% of IT teams have moved, or plan to move, their print workflows to the cloud due to its simplicity. Consumers can print easily to any printer from their PC, tablet or smartphone, while the Cloud print service monitors the supplies level. Many printer vendors such as Lexmark propose an automatic supplies shipment based on the real-time analysis of the printer supplies and user behavior to ensure printing will always be possible. For IT department, Cloud Printing eliminates the need for print servers and represents the only way to print from Cloud virtual desktops and servers. For consumers, cloud ready printers eliminate the need for PC connections and print drivers, enabling them to print from mobile devices. As for publishers and content owners, cloud printing allows them to "avoid the cost and complexity of buying and managing the underlying hardware, software and processes" required for the production of professional print products. Leveraging cloud print for print on demand also allows businesses to cut down on the costs associated with mass production. Moreover, cloud printing can be considered more eco-friendly, as it significantly reduces the amount of paper used (13% reduction in print jobs yearly) and lowers carbon emissions from transportation. As many companies move their IT to the Cloud, some adopting the Windows 365 and Azure Virtual Desktop services from Microsoft, the connection from the Cloud environment to the on-premise printers become an issue as opening ports for incoming print flow traffic is not an option. In 2020, at the exact same time Google discontinued its Google Print offer, Microsoft has announced its Universal Print service offer, aimed at making printing compatible with Cloud Desktop environments, making printing driver-free and simple with no client to install on PC. With Universal Print Microsoft has built a disrupting architecture with a value proposition commodifying printers, removing print servers and drivers, allowing to move printers to VLAN for security purpose and printing from anywhere. Clients are free to use any printer from any model as they all work the same, clients are not tied anymore to any printer brand and that gave a significant boost to the Cloud print market. That Microsoft Universal Print architecture provides APIs to third-party developers who can develop add-ons such as Celiveo 365 to extend Microsoft Cloud Print with added features such as access control on printers and copiers, follow-me pull print, data encryption, advanced usage reporting or charge back. == Providers of Consumer Cloud Printing Solutions == Before 2020 only a handful of providers used to work towards a professional cloud print solution, operating in their own niche or focus on mobile devices. In 2020 Microsoft has boosted that market by announcing its Universal Print Cloud printing service and since then many publishers have started to propose solutions for that growing market. The Covid pandemic also created the need for employees to be able to print at home when using the corporate IT software. Closed VPN often prevent accessing home network printers from corporate laptops and Full Public Cloud solutions are meant to be a solution to that problem. After the decision by Google to terminate Google Cloud Print service on 31 December 2020, most printer vendors released their own mobile cloud solution to fill the gap, while Hewlett-Packard implemented its own cloud print with their ePrint solution. Those solutions are often proprietary, only working on printers proposed by the vendor. Google has decided to let third-party developers develop Cloud Print solutions and to limit its scope to certifying the best Print Management offers compatible with its Chrome Enterprise Cloud ecosystem. == Providers of Corporate Cloud Printing solutions == While many print solutions claim to be "Cloud Printing", there are actually three categories: full Private Cloud, full Public Cloud, and Hybrid Cloud. Their differences are real and have an impact on the overall TCO as the more software there is on-site, the more hidden cost there are. In the Full Public Cloud category, independent SaaS vendors like Celiveo, ezeep , Printix , and Y Soft support a wide range of printer brands and models, allowing clients to buy the best printer without being locked on any brand. They are leveraging cloud computing technology to offer cloud-based print infrastructure and cloud-based printing software as a Service (SaaS). These solutions have integrations to cloud enabled printers or provide embedded printer agents. They feature allow users to print to any printer in any network, isolated network or not, even if that printer is otherwise not reachable from the user's computer. This also allows IT departments to move printers to VLAN for maximum security, like what they are doing with IP phones. Google Chrome Enterprise Cloud ecosystem has its own technical particularities and Google certifies Print Management solutions, ensuring they comply with Google technical requirement, yet letting each solution differentiate from others with specific features or security. Many of solutions for Chrome Enterprise are Hybrid, a few are Full Public Cloud. Industry experts believe that as these services become more popular, users will no longer consider printers as necessary assets but rather as devices that they can access on demand when the need to generate a printed page presents itself. == Caveats of Cloud Printing == == Security == Print jobs flow through Public Internet. It is therefore important to verify no Man-in-the-Middle attack can be performed. The only technical solution is to ensure each printer and PC uses a non-self-generated cryptographic token or certificate allowing TLS mutual authentication and specific data encryption. Self-generated printer certificates are unknown from the Cloud and prevent trusted authentication. Microsoft has implemented its Zero Trust Access security in its Universal Print service, it generates a unique certificate on printers compatible with its service. Other Cloud Printing SaaS providers have followed Microsoft on that High Security path. Print jobs data stored on the Cloud is sensitive as it contains user information as well as all information appearing on pages. Good practices require such data is encrypted at rest and in motion, using asymmetric PKI keys instead of fixed encryption keys. Some solutions require to open incoming traffic ports on the firewall to let Cloud services communicate with printers attached behind that firewall (most of the time for IPP/IPPS flows), some other solutions use a pull model where the communication is always initiated by the printer and no firewall port needs to be open. In terms of security the later is to be preferred.

  Read more →
• 

  Instance-based learning

  In machine learning, instance-based learning (sometimes called memory-based learning) is a family of learning algorithms that, instead of performing explicit generalization, compare new problem instances with instances seen in training, which have been stored in memory. Because computation is postponed until a new instance is observed, these algorithms are sometimes referred to as "lazy." It is called instance-based because it constructs hypotheses directly from the training instances themselves. This means that the hypothesis complexity can grow with the data: in the worst case, a hypothesis is a list of n training items and the computational complexity of classifying a single new instance is O(n). One advantage that instance-based learning has over other methods of machine learning is its ability to adapt its model to previously unseen data. Instance-based learners may simply store a new instance or throw an old instance away. Examples of instance-based learning algorithms are the k-nearest neighbors algorithm, kernel machines and RBF networks. These store (a subset of) their training set; when predicting a value/class for a new instance, they compute distances or similarities between this instance and the training instances to make a decision. To battle the memory complexity of storing all training instances, as well as the risk of overfitting to noise in the training set, instance reduction algorithms have been proposed.

  Read more →
• 

  Attention (machine learning)

  In machine learning, attention is a method that determines the importance of each component in a sequence relative to the other components in that sequence. In natural language processing, importance is represented by "soft" weights assigned to each word in a sentence. More generally, attention encodes vectors called token embeddings across a fixed-width sequence that can range from tens to millions of tokens in size. Unlike "hard" weights, which are computed during the backwards training pass, "soft" weights exist only in the forward pass and therefore change with every step of the input. Earlier designs implemented the attention mechanism in a serial recurrent neural network (RNN) language translation system, but a more recent design, namely the transformer, removed the slower sequential RNN and relied more heavily on the faster parallel attention scheme. Inspired by ideas about attention in humans, the attention mechanism was developed to address the weaknesses of using information from the hidden layers of recurrent neural networks. Recurrent neural networks favor information contained in words at the end of a sentence and thus deemed more recent, thereby tending to attenuate the significance and associated predictive weight assigned to information earlier in the sentence. Attention allows a token equal access to any part of a sentence directly, rather than only through the previous state. == History == Additional surveys of the attention mechanism in deep learning are provided by Niu et al. and Soydaner. The major breakthrough came with self-attention, where each element in the input sequence attends to all others, enabling the model to capture global dependencies. This idea was central to the Transformer architecture, which replaced recurrence with attention mechanisms. As a result, Transformers became the foundation for models like BERT, T5 and generative pre-trained transformers (GPT). == Overview == The modern era of machine attention was revitalized by grafting an attention mechanism (Fig 1. orange) to an Encoder-Decoder. Figure 2 shows the internal step-by-step operation of the attention block (A) in Fig 1. === Interpreting attention weights === In translating between languages, alignment is the process of matching words from the source sentence to words of the translated sentence. Networks that perform verbatim translation without regard to word order would show the highest scores along the (dominant) diagonal of the matrix. The off-diagonal dominance shows that the attention mechanism is more nuanced. Consider an example of translating I love you to French. On the first pass through the decoder, 94% of the attention weight is on the first English word I, so the network offers the word je. On the second pass of the decoder, 88% of the attention weight is on the third English word you, so it offers t'. On the last pass, 95% of the attention weight is on the second English word love, so it offers aime. In the I love you example, the second word love is aligned with the third word aime. Stacking soft row vectors together for je, t', and aime yields an alignment matrix: Sometimes, alignment can be multiple-to-multiple. For example, the English phrase look it up corresponds to cherchez-le. Thus, "soft" attention weights work better than "hard" attention weights (setting one attention weight to 1, and the others to 0), as we would like the model to make a context vector consisting of a weighted sum of the hidden vectors, rather than "the best one", as there may not be a best hidden vector. == Variants == Many variants of attention implement soft weights, such as fast weight programmers, or fast weight controllers (1992). A "slow" neural network outputs the "fast" weights of another neural network through outer products. The slow network learns by gradient descent. It was later renamed as "linearized self-attention". Bahdanau-style attention, also referred to as additive attention, Luong-style attention, which is known as multiplicative attention, Early attention mechanisms similar to modern self-attention were proposed using recurrent neural networks. However, the highly parallelizable self-attention was introduced in 2017 and successfully used in the Transformer model, positional attention and factorized positional attention. For convolutional neural networks, attention mechanisms can be distinguished by the dimension on which they operate, namely: spatial attention, channel attention, or combinations. These variants recombine the encoder-side inputs to redistribute those effects to each target output. Often, a correlation-style matrix of dot products provides the re-weighting coefficients. In the figures below, W is the matrix of context attention weights, similar to the formula in Overview section above. == Optimizations == === Flash attention === The size of the attention matrix is proportional to the square of the number of input tokens. Therefore, when the input is long, calculating the attention matrix requires a lot of GPU memory. Flash attention is an implementation that reduces the memory needs and increases efficiency without sacrificing accuracy. It achieves this by partitioning the attention computation into smaller blocks that fit into the GPU's faster on-chip memory, reducing the need to store large intermediate matrices and thus lowering memory usage while increasing computational efficiency. === FlexAttention === FlexAttention is an attention kernel developed by Meta that allows users to modify attention scores prior to softmax and dynamically chooses the optimal attention algorithm. == Applications == Attention is widely used in natural language processing, computer vision, and speech recognition. In NLP, it improves context understanding in tasks like question answering and summarization. In vision, visual attention helps models focus on relevant image regions, enhancing object detection and image captioning. === Attention maps as explanations for vision transformers === From the original paper on vision transformers (ViT), visualizing attention scores as a heat map (called saliency maps or attention maps) has become an important and routine way to inspect the decision making process of ViT models. One can compute the attention maps with respect to any attention head at any layer, while the deeper layers tend to show more semantically meaningful visualization. Attention rollout is a recursive algorithm to combine attention scores across all layers, by computing the dot product of successive attention maps. Because vision transformers are typically trained in a self-supervised manner, attention maps are generally not class-sensitive. When a classification head is attached to the ViT backbone, class-discriminative attention maps (CDAM) combines attention maps and gradients with respect to the class [CLS] token. Some class-sensitive interpretability methods originally developed for convolutional neural networks can be also applied to ViT, such as GradCAM, which back-propagates the gradients to the outputs of the final attention layer. Using attention as basis of explanation for the transformers in language and vision is not without debate. While some pioneering papers analyzed and framed attention scores as explanations, higher attention scores do not always correlate with greater impact on model performances. == Mathematical representation == === Standard scaled dot-product attention === For matrices: Q ∈ R m × d k , K ∈ R n × d k {\displaystyle Q\in \mathbb {R} ^{m\times d_{k}},K\in \mathbb {R} ^{n\times d_{k}}} and V ∈ R n × d v {\displaystyle V\in \mathbb {R} ^{n\times d_{v}}} , the scaled dot-product, or QKV attention, is defined as: Attention ( Q , K , V ) = softmax ( Q K T d k ) V ∈ R m × d v {\displaystyle {\text{Attention}}(Q,K,V)={\text{softmax}}\left({\frac {QK^{T}}{\sqrt {d_{k}}}}\right)V\in \mathbb {R} ^{m\times d_{v}}} where T {\displaystyle {}^{T}} denotes transpose and the softmax function is applied independently to every row of its argument. The matrix Q {\displaystyle Q} contains m {\displaystyle m} queries, while matrices K , V {\displaystyle K,V} jointly contain an unordered set of n {\displaystyle n} key-value pairs. Value vectors in matrix V {\displaystyle V} are weighted using the weights resulting from the softmax operation, so that the rows of the m {\displaystyle m} -by- d v {\displaystyle d_{v}} output matrix are confined to the convex hull of the points in R d v {\displaystyle \mathbb {R} ^{d_{v}}} given by the rows of V {\displaystyle V} . To understand the permutation invariance and permutation equivariance properties of QKV attention, let A ∈ R m × m {\displaystyle A\in \mathbb {R} ^{m\times m}} and B ∈ R n × n {\displaystyle B\in \mathbb {R} ^{n\times n}} be permutation matrices; and D ∈ R m × n {\displaystyle D\in \mathbb {R} ^{m\times n}} an arbitrary matrix. The softmax function is permutation equivariant in the sense that: softmax ( A D B ) = A softmax ( D ) B {\displays

  Read more →
• 

  Ideonomy

  Ideonomy is a combinatorial "science of ideas" developed by American independent scholar Patrick M. Gunkel (1947–2017). Specifically, Ideonomy is concerned with the systematic organization of ideas and the discovery of the rules behind how ideas combine, diverge, and transform. Gunkel defined ideonomy as "the science of the laws of ideas and of the application of such laws to the generation of all possible ideas in connection with any subject, idea, or thing." In his 1992 book A History of Knowledge, Charles Van Doren compared ideonomy to a "mining operation" that excavates meanings and thought to discover treasures hidden deep within language. Sources from the 1980s and 1990s demonstrate that ideonomy was useful to academic researchers in fields including biology, toxicology, and nursing/patient care. Beginning in the 2010s, academics in a wide range of fields including machine learning, marketing, computational modeling, and cybersecurity have relied on materials generated for ideonomy to provide methodological support for their research. == Etymology and definition == The word "ideonomy" combines the Greek roots ideo- (from idea, meaning pattern or form) and -nomy (from nomos, meaning law or custom). The suffix -nomy suggests the laws concerning or the totality of knowledge about a given subject, as in astronomy or taxonomy. In a note posted on the MIT ideonomy website, Gunkel states that the word was supposedly first coined by the French Encyclopedists to refer to a science of ideas. No evidence is provided for this statement, however. The concept bears some relationship to Antoine Destutt de Tracy's "ideology" (1796), which originally meant a systematic science of ideas before acquiring its modern political connotations. Gunkel provided several metaphorical descriptions of ideonomy: An "idea bank": a computer network enabling systematic exploration of infinite possible ideas A "kaleidoscope" that can exhibit all possible combinations and transformations of ideas A "prism" capable of diffracting any idea into its cognitive components A "gigantic microscope for magnifying the ideocosm" == History and development == In 1984, Gunkel received a five-year unsolicited grant from the Richard Lounsbery Foundation of New York to develop ideonomy. A June 1, 1987 article on the front page of The Wall Street Journal brought Gunkel and ideonomy to wider public attention. Some academics were interested in using ideonomy's techniques, including biologist Betsey Dyer, who published several contemporaneous peer-reviewed studies citing ideonomy. Academic researchers in the field of toxicology and nursing/patient care also used ideonomy. However, ideonomy's broadest contribution to date came beginning in the 2010s, as a list of personality traits generated for combinatorial matching was used by researchers in artificial intelligence to code human emotions for machine-learning tasks, develop computational models related to personality, develop a measurement framework for influencer-brand recommender systems, and aid information awareness/cybersecurity assessment. == Methodology == The foundational empirical method of ideonomy involves the systematic creation of extensive lists. Gunkel's apartment reportedly contained thousands of lists on every conceivable topic. Gunkel termed each list an "organon," which he described as expanding through "combination, permutation, transformation, generalization, specialization, intersection, interaction, reapplication, recursive use, etc. of existing organons." The ideonomic process follows a progressive structure. The ideonomist begins with a simple list of examples of a particular idea, concept, or thing. The list need not be exhaustive. By studying this list, the ideonomist isolates and identifies types. This categorical analysis then reveals missing items, allowing the primary list to be improved and refined. Gunkel emphasized that list items must not only cover genuine categories of nature but also be formulated in ways that yield the largest possible number of syntactically coherent possibilities when combined. The core technique of ideonomy is "ideocombinatorics"—the systematic intersection and combination of items from different lists to generate novel composite concepts. Gunkel developed computer programs to automate this process. For example, combining a list of 230 Universal Elementary Shapes (pits, pyramids, trenches, hemispheres, needles) with a list of 74 Types of Order (recurrence, identity, likeness of parts) yields 17,020 possible "shapes of order." These combinations, when phrased as questions ("Can there be pits of recurrence?"), could suggest new categories of phenomena worthy of investigation. The computer-generated output is typically repetitive and often meaningless. However, with sufficient frequency, the combinations yield results that are unexpectedly interesting and fruitful. In one documented case, Gunkel's programs generated 45,540 questions about toxins for microbiologist David Bermudes. One question—"Can hierarchies of cell process be used as a basis for classifying toxic action?"—prompted Bermudes to develop a novel approach to classifying biological toxins by the type of molecule they attack, rather than by chemical structure or physiological system affected. According to one contemporaneous account of ideonomy, "Gunkel takes for his field all fields and all ideas about anything. He uses a computer to generate lists of words and phrases and by juxtaposition reviews the resultant patterns for novel ideas. The computer is ideal for this task because the mind would rebel at the formidable processing task ideonomy involves. What we have here is computer generated originality." == Applications == Gunkel and his supporters identified several practical applications for ideonomic methods: Scientific research: Biologist Betsey Dyer of Wheaton College published research crediting ideonomy for helping to generate ideas. Medical science: When Austin pathologist Michael T. O'Brien was presented with the ideonomically-generated question "Can arteries have rashes?", he initially dismissed it as nonsense. Upon reflection, he realized that large arteries are supplied with blood by tiny vessels that might become inflamed and dilated, analogous to skin vessels in a rash—a phenomenon potentially worth researching. Analogical thinking: Harvard law professor Robert Clark used ideonomic analogies to write a research paper comparing plant structure with human hierarchies. Artificial intelligence: Douglas Lenat, a researcher at Microelectronics and Computer Technology Corporation (MCC) in Austin, suggested that Gunkel's lists enumerating types of human mistakes could help design AI systems capable of recognizing and correcting their own errors. == Reception and criticism == Ideonomy received mixed reactions from the academic and scientific communities. Prominent supporters included: Edward Fredkin, former director of MIT's computer science laboratory, who praised Gunkel's "provocative ideas on artificial intelligence." Marvin Minsky, AI scientist and MIT professor, who described ideonomy as "perhaps the most extensive study of ways to generate ideas." Frederick Seitz, president emeritus of Rockefeller University, who noted Gunkel's "encyclopedic scope" Robert C. Clark, Harvard law professor, who called Gunkel "the most intelligent person I ever met" However, skeptics questioned whether ideonomy constituted a genuine science. Fredkin himself noted that Gunkel "pours out about 60 ideas a minute, and 59 of them are bad," though he added that "even with one good idea out of 60, it's still an amazing accomplishment." Douglas Lenat observed that brainstorming with Gunkel was "a bit like being hit over the head by the muse with a sledgehammer" and that "he puts people off." Gunkel himself acknowledged that ideonomy was in its infancy and might seem "absurdly utopian." His planned magnum opus on ideonomy remained incomplete, and was posted on an MIT website thanks to faculty advisor Whitman Richards. Gunkel wrote: "Pioneering in a completely new field, yes in a new science, is almost unreal. It is heartbreaking, it is pitiable, it is almost inhuman. Honestly, it is a hell. There is nothing heroic about it." == Related concepts == Gunkel identified several historical precedents for ideonomic thinking: Gottfried Wilhelm Leibniz (1646–1716): The philosopher's work on a universal characteristic (characteristica universalis) and calculus of reasoning Peter Mark Roget (1779–1869): Creator of Roget's Thesaurus, which organized concepts into a systematic taxonomy Dmitri Mendeleev (1834–1907): Developer of the periodic table, demonstrating how combining lists of element families could reveal previously unseen connections Fritz Zwicky (1898–1974): The Caltech astrophysicist whom Gunkel called the "grandfather of ideonomy" for his development of "morphological research"—systematic exploration of all possible solutions t

  Read more →
• 

  Textual entailment

  In natural language processing, textual entailment (TE), also known as natural language inference (NLI), is a directional relation between text fragments. The relation holds whenever the truth of one text fragment follows from another text. == Definition == In the TE framework, the entailing and entailed texts are termed text (t) and hypothesis (h), respectively. Textual entailment is not the same as pure logical entailment – it has a more relaxed definition: "t entails h" (t ⇒ h) if, typically, a human reading t would infer that h is most likely true. (Alternatively: t ⇒ h if and only if, typically, a human reading t would be justified in inferring the proposition expressed by h from the proposition expressed by t.) The relation is directional because even if "t entails h", the reverse "h entails t" is much less certain. Determining whether this relationship holds is an informal task, one which sometimes overlaps with the formal tasks of formal semantics (satisfying a strict condition will usually imply satisfaction of a less strict conditioned); additionally, textual entailment partially subsumes word entailment. == Examples == Textual entailment can be illustrated with examples of three different relations: An example of a positive TE (text entails hypothesis) is: text: If you help the needy, God will reward you. hypothesis: Giving money to a poor man has good consequences. An example of a negative TE (text contradicts hypothesis) is: text: If you help the needy, God will reward you. hypothesis: Giving money to a poor man has no consequences. An example of a non-TE (text does not entail nor contradict) is: text: If you help the needy, God will reward you. hypothesis: Giving money to a poor man will make you a better person. == Ambiguity of natural language == A characteristic of natural language is that there are many different ways to state what one wants to say: several meanings can be contained in a single text and the same meaning can be expressed by different texts. This variability of semantic expression can be seen as the dual problem of language ambiguity. Together, they result in a many-to-many mapping between language expressions and meanings. The task of paraphrasing involves recognizing when two texts have the same meaning and creating a similar or shorter text that conveys almost the same information. Textual entailment is similar but weakens the relationship to be unidirectional. Mathematical solutions to establish textual entailment can be based on the directional property of this relation, by making a comparison between some directional similarities of the texts involved. == Approaches == Textual entailment measures natural language understanding as it asks for a semantic interpretation of the text, and due to its generality remains an active area of research. Many approaches and refinements of approaches have been considered, such as word embedding, logical models, graphical models, rule systems, contextual focusing, and machine learning. Practical or large-scale solutions avoid these complex methods and instead use only surface syntax or lexical relationships, but are correspondingly less accurate. As of 2005, state-of-the-art systems are far from human performance; a study found humans to agree on the dataset 95.25% of the time. Algorithms from 2016 had not yet achieved 90%. == Applications == Many natural language processing applications, like question answering, information extraction, summarization, multi-document summarization, and evaluation of machine translation systems, need to recognize that a particular target meaning can be inferred from different text variants. Typically entailment is used as part of a larger system, for example in a prediction system to filter out trivial or obvious predictions. Textual entailment also has applications in adversarial stylometry, which has the objective of removing textual style without changing the overall meaning of communication. == Datasets == Some of available English NLI datasets include: SNLI MultiNLI SciTail SICK MedNLI QA-NLI In addition, there are several non-English NLI datasets, as follows: XNLI DACCORD, RTE3-FR, SICK-FR for French FarsTail for Farsi OCNLI for Chinese SICK-NL for Dutch IndoNLI for Indonesian

  Read more →
• 

  H2O (software)

  H2O is an open-source, in-memory, distributed machine learning and predictive analytics platform developed by the company H2O.ai (previously 0xdata). The software uses a distributed architecture for parallel processing on standard hardware. It supports algorithms for large-scale data analysis and model deployment. H2O is primarily used by data scientists and developers for statistical modeling and data-driven decision-making. The platform is designed to handle in-memory computations across a distributed computing environment. It offers implementations for numerous statistical and machine learning algorithms, which are accessible through various programming interfaces. The software is released under the Apache License 2.0. == Functionality and features == H2O provides a suite of supervised and unsupervised machine learning algorithms. Its core functions include: Supervised learning: algorithms in the field of statistics, data mining and machine learning such as generalized linear models, random forests, gradient boosting and deep learning are implemented for classification and regression tasks. Unsupervised learning: including K-Means clustering and principal component analysis. Automated machine learning: a features designed to automate the processes of model selection, tuning, and ensemble creation. The software can ingest data from various sources, including the Hadoop Distributed File System, Amazon S3, SQL databases, as well as local file systems. It operates natively on Apache Spark clusters through Sparkling Water. Proponents claim that improved performance is achieved compared to other analysis tools. The software is distributed free of charge, under a business model based on the development of individual applications and support. == Architecture == H2O is primarily written in Java. It uses a distributed architecture that allows the platform to cluster nodes for parallel processing and in-memory storage of data and models. Users interact with the H2O platform through several primary interfaces: Programming language interfaces: APIs are provided for the R and Python programming languages, and various Apache offerings (Apache Hadoop and Spark, as well as Maven). H2O Flow: a graphical web-based interactive computational environment that functions as a notebook interface for data exploration, model building, and scripting. REST-API: allows for integration with other applications and frameworks such as Microsoft Excel or RStudio. With the H2O Machine Learning Integration Nodes, KNIME offers algorithmic workflows. While the algorithm executes, approximate results are displayed, so that users can track the progress and intervene if needed. == History, influences, and extensions == The software project was initiated by the company 0xdata, which later changed its name to H2O.ai. The three Stanford professors Stephen P. Boyd, Robert Tibshirani and Trevor Hastie form a panel that advises H2O on scientific issues. Since its inception, H2O provides open-source machine learning libraries for enterprise use. The core H2O platform is often complemented by offerings from H2O.ai, such as H2O Driverless AI. == Reception == H2O is referenced in peer-reviewed literature regarding automated machine learning (AutoML). The platform has been categorized as a "Leader" and a "Strong Performer" in industry reports by Forrester Research. H2O (the open-source platform) and the associated commercial platform Driverless AI have been recurring winners of InfoWorld's most prestigious awards, including both the Best of Open Source Software ("Bossies") and the Technology of the Year awards.

  Read more →
• 

  Actor-critic algorithm

  The actor-critic algorithm (AC) is a family of reinforcement learning (RL) algorithms that combine policy-based RL algorithms such as policy gradient methods, and value-based RL algorithms such as value iteration, Q-learning, SARSA, and TD learning. An AC algorithm consists of two main components: an "actor" that determines which actions to take according to a policy function, and a "critic" that evaluates those actions according to a value function. Some AC algorithms are on-policy, some are off-policy. Some apply to either continuous or discrete action spaces. Some work in both cases. == Overview == The actor-critic methods can be understood as an improvement over pure policy gradient methods like REINFORCE via introducing a baseline. === Actor === The actor uses a policy function π ( a | s ) {\displaystyle \pi (a|s)} , while the critic estimates either the value function V ( s ) {\displaystyle V(s)} , the action-value Q-function Q ( s , a ) , {\displaystyle Q(s,a),} the advantage function A ( s , a ) {\displaystyle A(s,a)} , or any combination thereof. The actor is a parameterized function π θ {\displaystyle \pi _{\theta }} , where θ {\displaystyle \theta } are the parameters of the actor. The actor takes as argument the state of the environment s {\displaystyle s} and produces a probability distribution π θ ( ⋅ | s ) {\displaystyle \pi _{\theta }(\cdot |s)} . If the action space is discrete, then ∑ a π θ ( a | s ) = 1 {\displaystyle \sum _{a}\pi _{\theta }(a|s)=1} . If the action space is continuous, then ∫ a π θ ( a | s ) d a = 1 {\displaystyle \int _{a}\pi _{\theta }(a|s)da=1} . The goal of policy optimization is to improve the actor. That is, to find some θ {\displaystyle \theta } that maximizes the expected episodic reward J ( θ ) {\displaystyle J(\theta )} : J ( θ ) = E π θ [ ∑ t = 0 T γ t r t ] {\displaystyle J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{t=0}^{T}\gamma ^{t}r_{t}\right]} where γ {\displaystyle \gamma } is the discount factor, r t {\displaystyle r_{t}} is the reward at step t {\displaystyle t} , and T {\displaystyle T} is the time-horizon (which can be infinite). The goal of policy gradient method is to optimize J ( θ ) {\displaystyle J(\theta )} by gradient ascent on the policy gradient ∇ J ( θ ) {\displaystyle \nabla J(\theta )} . As detailed on the policy gradient method page, there are many unbiased estimators of the policy gradient: ∇ θ J ( θ ) = E π θ [ ∑ 0 ≤ j ≤ T ∇ θ ln ⁡ π θ ( A j | S j ) ⋅ Ψ j | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{0\leq j\leq T}\nabla _{\theta }\ln \pi _{\theta }(A_{j}|S_{j})\cdot \Psi _{j}{\Big |}S_{0}=s_{0}\right]} where Ψ j {\textstyle \Psi _{j}} is a linear sum of the following: ∑ 0 ≤ i ≤ T ( γ i R i ) {\textstyle \sum _{0\leq i\leq T}(\gamma ^{i}R_{i})} . γ j ∑ j ≤ i ≤ T ( γ i − j R i ) {\textstyle \gamma ^{j}\sum _{j\leq i\leq T}(\gamma ^{i-j}R_{i})} : the REINFORCE algorithm. γ j ∑ j ≤ i ≤ T ( γ i − j R i ) − b ( S j ) {\textstyle \gamma ^{j}\sum _{j\leq i\leq T}(\gamma ^{i-j}R_{i})-b(S_{j})} : the REINFORCE with baseline algorithm. Here b {\displaystyle b} is an arbitrary function. γ j ( R j + γ V π θ ( S j + 1 ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(R_{j}+\gamma V^{\pi _{\theta }}(S_{j+1})-V^{\pi _{\theta }}(S_{j})\right)} : TD(1) learning. γ j Q π θ ( S j , A j ) {\textstyle \gamma ^{j}Q^{\pi _{\theta }}(S_{j},A_{j})} . γ j A π θ ( S j , A j ) {\textstyle \gamma ^{j}A^{\pi _{\theta }}(S_{j},A_{j})} : Advantage Actor-Critic (A2C). γ j ( R j + γ R j + 1 + γ 2 V π θ ( S j + 2 ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(R_{j}+\gamma R_{j+1}+\gamma ^{2}V^{\pi _{\theta }}(S_{j+2})-V^{\pi _{\theta }}(S_{j})\right)} : TD(2) learning. γ j ( ∑ k = 0 n − 1 γ k R j + k + γ n V π θ ( S j + n ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(\sum _{k=0}^{n-1}\gamma ^{k}R_{j+k}+\gamma ^{n}V^{\pi _{\theta }}(S_{j+n})-V^{\pi _{\theta }}(S_{j})\right)} : TD(n) learning. γ j ∑ n = 1 ∞ λ n − 1 1 − λ ⋅ ( ∑ k = 0 n − 1 γ k R j + k + γ n V π θ ( S j + n ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\sum _{n=1}^{\infty }{\frac {\lambda ^{n-1}}{1-\lambda }}\cdot \left(\sum _{k=0}^{n-1}\gamma ^{k}R_{j+k}+\gamma ^{n}V^{\pi _{\theta }}(S_{j+n})-V^{\pi _{\theta }}(S_{j})\right)} : TD(λ) learning, also known as GAE (generalized advantage estimate). This is obtained by an exponentially decaying sum of the TD(n) learning terms. === Critic === In the unbiased estimators given above, certain functions such as V π θ , Q π θ , A π θ {\displaystyle V^{\pi _{\theta }},Q^{\pi _{\theta }},A^{\pi _{\theta }}} appear. These are approximated by the critic. Since these functions all depend on the actor, the critic must learn alongside the actor. The critic is learned by value-based RL algorithms. For example, if the critic is estimating the state-value function V π θ ( s ) {\displaystyle V^{\pi _{\theta }}(s)} , then it can be learned by any value function approximation method. Let the critic be a function approximator V ϕ ( s ) {\displaystyle V_{\phi }(s)} with parameters ϕ {\displaystyle \phi } . The simplest example is TD(1) learning, which trains the critic to minimize the TD(1) error: δ i = R i + γ V ϕ ( S i + 1 ) − V ϕ ( S i ) {\displaystyle \delta _{i}=R_{i}+\gamma V_{\phi }(S_{i+1})-V_{\phi }(S_{i})} The critic parameters are updated by gradient descent on the squared TD error: ϕ ← ϕ − α ∇ ϕ ( δ i ) 2 = ϕ + α δ i ∇ ϕ V ϕ ( S i ) {\displaystyle \phi \leftarrow \phi -\alpha \nabla _{\phi }(\delta _{i})^{2}=\phi +\alpha \delta _{i}\nabla _{\phi }V_{\phi }(S_{i})} where α {\displaystyle \alpha } is the learning rate. Note that the gradient is taken with respect to the ϕ {\displaystyle \phi } in V ϕ ( S i ) {\displaystyle V_{\phi }(S_{i})} only, since the ϕ {\displaystyle \phi } in γ V ϕ ( S i + 1 ) {\displaystyle \gamma V_{\phi }(S_{i+1})} constitutes a moving target, and the gradient is not taken with respect to that. This is a common source of error in implementations that use automatic differentiation, and requires "stopping the gradient" at that point. Similarly, if the critic is estimating the action-value function Q π θ {\displaystyle Q^{\pi _{\theta }}} , then it can be learned by Q-learning or SARSA. In SARSA, the critic maintains an estimate of the Q-function, parameterized by ϕ {\displaystyle \phi } , denoted as Q ϕ ( s , a ) {\displaystyle Q_{\phi }(s,a)} . The temporal difference error is then calculated as δ i = R i + γ Q θ ( S i + 1 , A i + 1 ) − Q θ ( S i , A i ) {\displaystyle \delta _{i}=R_{i}+\gamma Q_{\theta }(S_{i+1},A_{i+1})-Q_{\theta }(S_{i},A_{i})} . The critic is then updated by θ ← θ + α δ i ∇ θ Q θ ( S i , A i ) {\displaystyle \theta \leftarrow \theta +\alpha \delta _{i}\nabla _{\theta }Q_{\theta }(S_{i},A_{i})} The advantage critic can be trained by training both a Q-function Q ϕ ( s , a ) {\displaystyle Q_{\phi }(s,a)} and a state-value function V ϕ ( s ) {\displaystyle V_{\phi }(s)} , then let A ϕ ( s , a ) = Q ϕ ( s , a ) − V ϕ ( s ) {\displaystyle A_{\phi }(s,a)=Q_{\phi }(s,a)-V_{\phi }(s)} . Although, it is more common to train just a state-value function V ϕ ( s ) {\displaystyle V_{\phi }(s)} , then estimate the advantage by A ϕ ( S i , A i ) ≈ ∑ j ∈ 0 : n − 1 γ j R i + j + γ n V ϕ ( S i + n ) − V ϕ ( S i ) {\displaystyle A_{\phi }(S_{i},A_{i})\approx \sum _{j\in 0:n-1}\gamma ^{j}R_{i+j}+\gamma ^{n}V_{\phi }(S_{i+n})-V_{\phi }(S_{i})} Here, n {\displaystyle n} is a positive integer. The higher n {\displaystyle n} is, the more lower is the bias in the advantage estimation, but at the price of higher variance. The Generalized Advantage Estimation (GAE) introduces a hyperparameter λ {\displaystyle \lambda } that smoothly interpolates between Monte Carlo returns ( λ = 1 {\displaystyle \lambda =1} , high variance, no bias) and 1-step TD learning ( λ = 0 {\displaystyle \lambda =0} , low variance, high bias). This hyperparameter can be adjusted to pick the optimal bias-variance trade-off in advantage estimation. It uses an exponentially decaying average of n-step returns with λ {\displaystyle \lambda } being the decay strength. == Variants == Asynchronous Advantage Actor-Critic (A3C): Parallel and asynchronous version of A2C. Soft Actor-Critic (SAC): Incorporates entropy maximization for improved exploration. Deep Deterministic Policy Gradient (DDPG): Specialized for continuous action spaces.

  Read more →