AI Video Generator Tools — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.
Stop Motion Studio is a stop motion animation software developed by Cateater LLC. It is available as both an app for iOS and Android and as a software for Windows and Mac. Two versions of the software exist, the standard Stop Motion Studio for free, and the paid Stop Motion Studio Pro, which contains extra, more advanced features. The software is commonly used in brickfilming.
Read more →
Lernmatrix (German for "learning matrix") is a special type of artificial neural network (ANN) architecture, similar to associative memory, invented around 1960 by Karl Steinbuch, a pioneer in computer science and ANNs. This model for learning systems could establish complex associations between certain sets of characteristics (e.g., letters of an alphabet) and their meanings. == Function == The Lernmatrix generally consists of n "characteristic lines" and m "meaning lines," where each characteristic line is connected to each meaning line, similar to how neurons in the brain are connected by synapses. (This can be realized in various ways – according to Steinbuch, this could be done by hardware or software). To train a Lernmatrix, values are specified on the corresponding characteristic and meaning lines (binary or real); then the connections between all pairs of characteristic and meaning lines are strengthened by the Hebb rule. A trained Lernmatrix, when given a specific input on the characteristic lines, activates the corresponding meaning lines. In modern language, it is a linear projection module. By appropriately interconnecting several Lernmatrices, a switching system can be built that, after completing certain training phases, is ultimately able to automatically determine the most probable associated meaning for an input sequence of features.
Read more →
The Fifth Generation Computer Systems (FGCS; Japanese: 第五世代コンピュータ, romanized: daigosedai konpyūta) was a 10-year initiative launched in 1982 by Japan's Ministry of International Trade and Industry (MITI) to develop computers based on massively parallel computing and logic programming. The project aimed to create an "epoch-making computer" with supercomputer-like performance and to establish a platform for future advancements in artificial intelligence. Although FGCS was noted as ahead of its time, and its ambitious goals contributed significantly to the development of concurrent logic programming, it ultimately ended in commercial failure. The term "fifth generation" was chosen to emphasize the system's advanced nature. In the history of computing hardware, there had been four prior "generations" of computers: the first generation utilized vacuum tubes; the second, transistors and diodes; the third, integrated circuits; and the fourth, microprocessors. While earlier generations focused on increasing the number of logic elements within a single CPU, it was widely believed at the time that the fifth generation would achieve enhanced performance through the use of massive numbers of CPUs. == Background == In the late 1960s until the early 1970s, there was much talk about "generations" of computer hardware, then usually organized into three generations First generation: Thermionic vacuum tubes. Mid-1940s. IBM pioneered the arrangement of vacuum tubes in pluggable modules. The IBM 650 was a first-generation computer. Second generation: Transistors. 1956. The era of miniaturization begins. Transistors are much smaller than vacuum tubes, draw less power, and generate less heat. Discrete transistors are soldered to circuit boards, with interconnections accomplished by stencil-screened conductive patterns on the reverse side. The IBM 7090 was a second-generation computer. Third generation: Integrated circuits (silicon chips containing multiple transistors). 1964. A pioneering example is the ACPX module used in the IBM 360/91, which, by stacking layers of silicon over a ceramic substrate, accommodated over 20 transistors per chip; the chips could be packed together onto a circuit board to achieve unprecedented logic densities. The IBM 360/91 was a hybrid second and third-generation computer. Omitted from this taxonomy is the "zeroth-generation" computer based on metal gears (such as the IBM 407) or mechanical relays (such as the Mark I), and the post-third-generation computers based on Very Large Scale Integrated (VLSI) circuits. There was also a parallel set of generations for software: First generation: Machine language. Second generation: Low-level programming languages such as Assembly language. Third generation: Structured high-level programming languages such as C, COBOL and FORTRAN. Fourth generation: "Non-procedural" high-level programming languages (such as object-oriented languages). Throughout these multiple generations up to the 1970s, Japan built computers following U.S. and British leads. In the mid-1970s, the Ministry of International Trade and Industry stopped following western leads and started looking into the future of computing on a small scale. They asked the Japan Information Processing Development Center (JIPDEC) to indicate a number of future directions, and in 1979 offered a three-year contract to carry out more in-depth studies along with industry and academia. It was during this period that the term "fifth-generation computer" started to be used. Prior to the 1970s, MITI guidance had successes such as an improved steel industry, the creation of the oil supertanker, the automotive industry, consumer electronics, and computer memory. MITI decided that the future was going to be information technology. However, the Japanese language, particularly in its written form, presented and still presents obstacles for computers. As a result of these hurdles, MITI held a conference to seek assistance from experts. The primary fields for investigation from this initial project were: Inference computer technologies for knowledge processing Computer technologies to process large-scale data bases and knowledge bases High-performance workstations Distributed functional computer technologies Super-computers for scientific calculation == Project launch == The aim was to build parallel computers for artificial intelligence applications using concurrent logic programming. The project imagined an "epoch-making" computer with supercomputer-like performance running on top of large databases (as opposed to a traditional filesystem) using a logic programming language to define and access the data using massively parallel computing/processing. They envisioned building a prototype machine with performance between 100M and 1G LIPS, where a LIPS is a Logical Inference Per Second. At the time typical workstation machines were capable of about 100k LIPS. They proposed to build this machine over a ten-year period, 3 years for initial R&D, 4 years for building various subsystems, and a final 3 years to complete a working prototype system. In 1982 the government decided to go ahead with the project, and established the Institute for New Generation Computer Technology (ICOT) through joint investment with various Japanese computer companies. After the project ended, MITI would consider an investment in a new "sixth generation" project. Ehud Shapiro captured the rationale and motivations driving this project: "As part of Japan's effort to become a leader in the computer industry, the Institute for New Generation Computer Technology has launched a revolutionary ten-year plan for the development of large computer systems which will be applicable to knowledge information processing systems. These Fifth Generation computers will be built around the concepts of logic programming. In order to refute the accusation that Japan exploits knowledge from abroad without contributing any of its own, this project will stimulate original research and will make its results available to the international research community." === Logic programming === The target defined by the FGCS project was to develop "Knowledge Information Processing systems" (roughly meaning, applied Artificial Intelligence). The chosen tool to implement this goal was logic programming. Logic programming approach as was characterized by Maarten Van Emden – one of its founders – as: The use of logic to express information in a computer. The use of logic to present problems to a computer. The use of logical inference to solve these problems. More technically, it can be summed up in two equations: Program = Set of axioms. Computation = Proof of a statement from axioms. The Axioms typically used are universal axioms of a restricted form, called Horn-clauses or definite-clauses. The statement proved in a computation is an existential statement. The proof is constructive, and provides values for the existentially quantified variables: these values constitute the output of the computation. Logic programming was thought of as something that unified various gradients of computer science (software engineering, databases, computer architecture and artificial intelligence). It seemed that logic programming was a key missing connection between knowledge engineering and parallel computer architectures. == Results == After having influenced the consumer electronics field during the 1970s and the automotive world during the 1980s, the Japanese had developed a strong reputation. The launch of the FGCS project spread the belief that parallel computing was the future of all performance gains, producing a wave of apprehension in the computer field. Soon parallel projects were set up in the US as the Strategic Computing Initiative and the Microelectronics and Computer Technology Corporation (MCC), in the UK as Alvey, and in Europe as the European Strategic Program on Research in Information Technology (ESPRIT), as well as the European Computer‐Industry Research Centre (ECRC) in Munich, a collaboration between ICL in Britain, Bull in France, and Siemens in Germany. The project ran from 1982 to 1994, spending a little less than ¥57 billion (about US$320 million) total. After the FGCS Project, MITI stopped funding large-scale computer research projects, and the research momentum developed by the FGCS Project dissipated. However MITI/ICOT embarked on a neural-net project which some called the Sixth Generation Project in the 1990s, with a similar level of funding. Per-year spending was less than 1% of the entire R&D expenditure of the electronics and communications equipment industry. For example, the project's highest expenditure year was 7.2 million yen in 1991, but IBM alone spent 1.5 billion dollars (370 billion yen) in 1982, while the industry spent 2150 billion yen in 1990. === Concurrent logic programming === In 1982, during a visit to the ICOT, Ehud Shapiro invented Concurrent Prolog, a novel programming language t
Read more →
OpenNN (Open Neural Networks Library) is a software library written in the C++ programming language which implements neural networks, a main area of deep learning research. The library is open-source, licensed under the GNU Lesser General Public License. == Characteristics == The software implements any number of layers of non-linear processing units for supervised learning. This deep architecture allows the design of neural networks with universal approximation properties. Additionally, it allows multiprocessing programming by means of OpenMP, in order to increase computer performance. OpenNN contains machine learning algorithms as a bundle of functions. These can be embedded in other software tools, using an application programming interface, for the integration of the predictive analytics tasks. In this regard, a graphical user interface is missing but some functions can be supported by specific visualization tools. == History == The development started in 2003 at the International Center for Numerical Methods in Engineering, within the research project funded by the European Union called RAMFLOOD (Risk Assessment and Management of FLOODs). Then it continued as part of similar projects. OpenNN is being developed by the startup company Artelnics. == Applications == OpenNN is a general purpose artificial intelligence software package. It uses machine learning techniques for solving predictive analytics tasks in different fields. For instance, the library has been applied in the engineering, energy, or chemistry sectors.
Read more →
vLLM is an open-source software framework for inference and serving of large language models and related multimodal models. Originally developed at the University of California, Berkeley's Sky Computing Lab, the project is centered on PagedAttention, a memory-management method for transformer key–value caches, and supports features such as continuous batching, distributed inference, quantization, and OpenAI-compatible APIs. According to a project maintainer, the "v" in vLLM originally referred to "virtual", inspired by virtual memory. == History == vLLM was introduced in 2023 by researchers affiliated with the Sky Computing Lab at UC Berkeley. Its core ideas were described in the 2023 paper Efficient Memory Management for Large Language Model Serving with PagedAttention, which presented the system as a high-throughput and memory-efficient serving engine for large language models. In 2025, the PyTorch Foundation announced that vLLM had become a Foundation-hosted project. PyTorch's project page states that the University of California, Berkeley contributed vLLM to the Linux Foundation in July 2024. In January 2026, TechCrunch reported that the creators of vLLM had launched the startup Inferact to commercialize the project, raising $150 million in seed funding. == Architecture == According to its 2023 paper, vLLM was designed to improve the efficiency of large language model serving by reducing memory waste in the key–value cache used during transformer inference. The paper introduced PagedAttention, an algorithm inspired by virtual memory and paging techniques in operating systems, and described vLLM as using block-level memory management and request scheduling to increase throughput while maintaining similar latency. The project documentation and repository describe support for continuous batching, chunked prefill, speculative decoding, prefix caching, quantization, and multiple forms of distributed inference and serving. PyTorch has described vLLM as a high-throughput, memory-efficient inference and serving engine that supports a range of hardware back ends, including NVIDIA and AMD GPUs, Google TPUs, AWS Trainium, and Intel processors.
Read more →
Logic Programming Associates (LPA) is a company specializing in logic programming and artificial intelligence software. LPA was founded in 1980 and is widely known for its range of Prolog compilers, the Flex expert system toolkit and most recently, VisiRule. LPA was established to exploit research at the Department of Computing and Control at Imperial College London into logic programming carried out under the supervision of Prof Robert Kowalski. == History of LPA Prolog == One of the first Prolog implementations made available by LPA was micro-PROLOG which ran on popular 8-bit home computers such as the Sinclair ZX Spectrum and Apple II. The 8-bit micro-PROLOG interpreter was soon followed by micro-PROLOG Professional one of the first Prolog implementations for the IBM PC running MS-DOS. micro-PROLOG Professional could access all of the 640K memory available under MS-DOS and therefore manage much larger programs In 1985, LPA released LPA MacProlog which ran on the MacPlus and Mac II computers which could access up to 4 Mb memory. MacProlog was later licensed to Quintus for re-distribution in the USA. In 1989, LPA started work on a new 32-bit Prolog compiler which could use DOS-extender technology to access up to 4GB memory. This became the basis for LPA Prolog for Windows, aka WIN-PROLOG, which was then released for Windows 3.0 in 1990. LPA's core Prolog product is LPA Prolog for Windows, a compiler and development system for the Microsoft Windows platform. The current LPA software range comprises an integrated AI toolset which covers various aspects of Artificial Intelligence including Logic Programming, Expert Systems, Knowledge-based Systems, Data Mining, Agents and Case-based reasoning etc. As well as continuing with Prolog compiler technology development, LPA has a track record of creating innovative associated tools and products to address specific challenges and opportunities. == Flex Expert System toolkit == In 1989, in response to the rise of interest in Expert Systems and the emergence of products such as Crystal, GoldWorks, NExpert, LPA developed the Flex expert system toolkit, which incorporated frame-based reasoning with inheritance, rule-based programming and data-driven procedures. Flex has its own English-like Knowledge Specification Language (KSL) which means that knowledge and rules are defined in an easy-to-read and understand way. LPA supported Flex on Windows, DOS and Macintosh PCs, as an add-on toolkit to its various LPA Prolog systems and eanbled LPA to enter the then quick vibrant Expert Systems rules-market. Flex was quickly established as the leading Prolog-based expert system toolkit and was licensed to other Prolog providors on other hardware platforms including Telecomputing Plc to supplement Top One on IBM and ICL mainframes. Other implementations included Quintec-Flex, Quintus Flex, Poplog Flex and BIM Flex which were all running on Unix and/or Vax/VMS platforms. POPLOG-Flex was used to build BRAND EVALUATOR - an expert system to assist brand specialists in evaluating the worth of branded products Quintec-Flex was used to build a hybrid system for the non-linear dynamic analysis/design of coupled shear walls Flex was adopted by the Open University as part of its course T396, "Artificial intelligence for technology" which was designed by Prof Adrian Hopgood. Some of the teaching material is now available on his AI tookit website. Flex was also used by David A Ferrucci and Selmer Bringsjord in their storytelling machine, BRUTUS. == PVG == In 1992, LPA helped set up the Prolog Vendors Group, a not-for-profit organization whose aim was to help promote Prolog by making people aware of its usage in industry. == Business Integrity Ltd and Contract Express == Between 1996 and 1998, based on work co-funded through a DTI Smart award, LPA developed ScaffoldIT, a tool for building dynamic documents and intelligent web sites. This technology, built using the LPA Prolog engine and associated ProWeb Server, was able to generate complex, personalised documents such as insurance policy schedules, legal contracts, and complex sales proposals, over the Web. In 1999/2000, LPA helped set up Business Integrity Ltd, as a Joint Venture with Tarlo-Lyons, to bring the above document assembly technology to market. This product eventually became Contract Express. Contract Express became very popular amongst large law firms and was sold worldwide for both internal and external use. Partners and GCs liked Contract Express because lawyers were able to quickly and accurately automate and update their legal templates in Word without requiring IT specialists to convert them into programs. As a result of the commercial success of Contract Express, BIL was acquired by Thomson Reuters in 2015. The very early days of BIL are described by Clive Spenser here. == VisiRule == In 2004, LPA launched VisiRule a graphical tool for developing knowledge-based and decision support systems. VisiRule was described in IEEE Potentials in 2007 (see Drawing on your knowledge with VisiRule): VisiRule has been used in various sectors, to build legal expert systems, machine diagnostic programs, medical and financial advice systems, etc. In 2013, VisiRule was incorporated into Ecosystem Management Decision Support (EMDS) where it has been used to provide enhanced decision support capabilities. EMDS integrates state-of-the-art geographic information system (GIS) as well as logic programming and decision modeling technologies on multiple platforms (Windows, Linux, Mac OS X) to provide decision support for a substantial portion of the adaptive management process of ecosystem management. EMDS is actively used, extended, supported and maintained by Mountain View Business Group (for an in-depth reprise of EMDS see the article in Frontiers in Environmental Science). In 2023, VisiRule was listed as one of the 5 best decision support software for large enterprises in 2024. == Customers == For many years, LPA has worked closely with Valdis Krebs, an American-Latvian researcher, author, and consultant in the field of social and organizational network analysis. Valdis is the founder and chief scientist of Orgnet, and the creator of the popular Inflow software package. LPA Prolog and Flex were used to create Allergenius, an expert system for the interpretation of allergen microarray results. Rules representing the knowledge base (KB) were derived from the literature and specialized databases. The input data included the patient's ID and disease(s), the results of either a skin prick test or specific IgE assays and ISAC results. The output was a medical report.
Read more →
Partial-order planning is an approach to automated planning that maintains a partial ordering between actions and only commits ordering between actions when forced to, that is, ordering of actions is partial. Also this planning doesn't specify which action will come out first when two actions are processed. By contrast, total-order planning maintains a total ordering between all actions at every stage of planning. Given a problem in which some sequence of actions is needed to achieve a goal, a partial-order plan specifies all actions that must be taken, but specifies an ordering between actions only where needed. Consider the following situation: a person must travel from the start to the end of an obstacle course. The course is composed of a bridge, a see-saw, and a swing-set. The bridge must be traversed before the see-saw and swing-set are reachable. Once reachable, the see-saw and swing-set can be traversed in any order, after which the end is reachable. In a partial-order plan, ordering between these obstacles is specified only when needed. The bridge must be traversed first. Second, either the see-saw or swing-set can be traversed. Third, the remaining obstacle can be traversed. Then the end can be traversed. Partial-order planning relies upon the principle of least commitment for its efficiency. == Partial-order plan == A partial-order plan or partial plan is a plan which specifies all actions that must be taken, but only specifies the order between actions when needed. It is the result of a partial-order planner. A partial-order plan consists of four components: A set of actions (also known as operators). A partial order for the actions. It specifies the conditions about the order of some actions. A set of causal links. It specifies which actions meet which preconditions of other actions. Alternatively, a set of bindings between the variables in actions. A set of open preconditions. It specifies which preconditions are not fulfilled by any action in the partial-order plan. To keep the possible orders of the actions as open as possible, the set of order conditions and causal links must be as small as possible. A plan is a solution if the set of open preconditions is empty. A linearization of a partial order plan is a total order plan derived from the particular partial order plan; in other words, both order plans consist of the same actions, with the order in the linearization being a linear extension of the partial order in the original partial order plan. === Example === For example, a plan for baking a cake might start: go to the store get eggs; get flour; get milk pay for all goods go to the kitchen This is a partial plan because the order for finding eggs, flour and milk is not specified, the agent can wander around the store reactively accumulating all the items on its shopping list until the list is complete. == Partial-order planner == A partial-order planner is an algorithm or program which will construct a partial-order plan and search for a solution. The input is the problem description, consisting of descriptions of the initial state, the goal and possible actions. The problem can be interpreted as a search problem where the set of possible partial-order plans is the search space. The initial state would be the plan with the open preconditions equal to the goal conditions. The final state would be any plan with no open preconditions, i.e. a solution. The initial state is the starting conditions, and can be thought of as the preconditions to the task at hand. For a task of setting the table, the initial state could be a clear table. The goal is simply the final action that needs to be accomplished, for example setting the table. The operators of the algorithm are the actions by which the task is accomplished. For this example there may be two operators: lay (tablecloth), and place (glasses, plates, and silverware). === Plan space === The plan space of the algorithm is constrained between its start and finish. The algorithm starts, producing the initial state and finishes when all parts of the goal have been achieved. In the setting a table example, two types of actions exist that must be addressed: the put-out and lay operators. Four unsolved operators also exist: Action 1, lay-tablecloth, Action 2, Put-out (plates), Action 3, Put-out (silverware), and Action 4, Put-out (glasses). However, a threat arises if Action 2, 3, or 4 comes before Action 1. This threat is that the precondition to the start of the algorithm will be unsatisfied as the table will no longer be clear. Thus, constraints exist that must be added to the algorithm that force Actions 2, 3, and 4 to come after Action 1. Once these steps are completed, the algorithm will finish and the goal will have been completed. === Threats === As seen in the algorithm presented above, partial-order planning can encounter certain threats, meaning orderings that threaten to break connected actions, thus potentially destroying the entire plan. There are two ways to resolve threats: Promotion Demotion Promotion orders the possible threat after the connection it threatens. Demotion orders the possible threat before the connection it threatens. Partial-order planning algorithms are known for being both sound and complete, with sound being defined as the total ordering of the algorithm, and complete being defined as the capability to find a solution, given that a solution does in fact exist. == Partial-order vs. total-order planning == Partial-order planning is the opposite of total-order planning, in which actions are sequenced all at once and for the entirety of the task at hand. The question arises when one has two competing processes, which one is better? Anthony Barret and Daniel Weld have argued in their 1993 book, that partial-order planning is superior to total-order planning, as it is faster and thus more efficient. They tested this theory using Korf’s taxonomy of subgoal collections, in which they found that partial-order planning performs better because it produces more trivial serializability than total-order planning. Trivial serializability facilitates a planner’s ability to perform quickly when dealing with goals that contain subgoals. Planners perform more slowly when dealing with laboriously serializable or nonserializable subgoals. The determining factor that makes a subgoal trivially or laboriously serializable is the search space of different plans. They found that partial-order planning is more adept at finding the quickest path, and is therefore the more efficient of these two main types of planning. == The Sussman anomaly == Partial-order plans are known to easily and optimally solve the Sussman anomaly. Using this type of incremental planning system solves this problem quickly and efficiently. This was a result of partial-order planning that solidified its place as an efficient planning system. == Disadvantages to partial-order planning == One drawback of this type of planning system is that it requires a lot more computational power for each node. This higher per-node cost occurs because the algorithm for partial-order planning is more complex than others. This has important artificial intelligence implications. When coding a robot to do a certain task, the creator needs to take into account how much energy is needed. Though a partial-order plan may be quicker it may not be worth the energy cost for the robot. The creator must be aware of and weigh these two options to build an efficient robot.
Read more →
Cristóbal Valenzuela (born 1989) is a Chilean-born technologist, software developer, and CEO of Runway. In 2018, Valenzuela co-founded the AI research company Runway in New York City with Anastasis Germanidis and Alejandro Matamala. == Education == Valenzuela graduated from Adolfo Ibáñez University (AIU), a research private university in Chile. From there, Valenzuela obtained a bachelor's degree in economics and business management, along with a master's degree in arts in design in 2012. In 2018, he graduated with a media arts degree from ITP NYU's Tisch School of the Arts. == Career and recognition == One of Valenzuela's first jobs was as a teaching and research assistant at the Adolfo Ibáñez University School of Design, and later an adjunct professor in the same department. In 2018, he became a researcher at NYU's Tisch School of the Arts ITP program, where he worked with Daniel Shiffman. He contributes to open-source software projects, including ml5.js, an open-source machine learning software. He co-founded Runway with two colleagues from ITP, Anastasis Germanidis, and Alejandro Matamala. The goal of Runway is to create new tools for human imagination using generative AI. In recent years, Valenzuela's work has been sponsored by Google and the Processing Foundation and his projects have been exhibited throughout Latin America and the US, including the Santiago Museum of Contemporary Art, Lollapalooza, NYC Media Lab, New Latin Wave, Inter-American Development Bank, Stanford University and New York University. In September 2023, Valenzuela was named as one of the TIME 100 Most Influential People in AI (TIME100 AI).
Read more →
Machine unlearning is a branch of machine learning focused on removing specific undesired element, such as private data, wrong or manipulated training data, outdated information, copyrighted material, harmful content, dangerous abilities, or misinformation, without needing to rebuild models from the ground up. Large language models, like the ones powering ChatGPT, may be asked not just to remove specific elements but also to unlearn a "concept," "fact," or "knowledge," which aren't easily linked to specific examples. New terms such as "model editing," "concept editing," and "knowledge unlearning" have emerged to describe this process. == History == Early research efforts were largely motivated by Article 17 of the GDPR, the European Union's privacy regulation commonly known as the "right to be forgotten" (RTBF), introduced in 2014. The GDPR did not anticipate that the development of large language models would make data erasure a complex task. This issue has since led to research on "machine unlearning," with a growing focus on removing copyrighted material, harmful content, dangerous capabilities, and misinformation. Just as early experiences in humans shape later ones, some concepts are more fundamental and harder to unlearn. A piece of knowledge may be so deeply embedded in the model's knowledge graph that unlearning it could cause internal contradictions, requiring adjustments to other parts of the graph to resolve them. Researchers have now also started studying unlearning in the context of removing incorrect or adversarially manipulated training data such as systematically biased labels or poisoning attacks. == Motivations == At present, machine unlearning is motivated by a growing range of concerns that extend well beyond the field's original focus on data privacy. A widely used taxonomy in the literature distinguishes two high-level categories of motivation. Access revocation covers cases where a data subject or rights holder requests the removal of data they own or control. This is most commonly associated with RTBF established by the European Union's General Data Protection Regulation (GDPR) and analogous legislation such as the California Consumer Privacy Act (CCPA). These regulations grant individuals the legal right to request erasure of their personal data from any system that has processed it, including models that were trained on it. Access revocation also encompasses the removal of copyrighted or pay-walled content that was incorporated into training corpora without the necessary licenses, a concern that has become prominent with the widespread use of largely web-scraped pre-training datasets. Model correction covers cases where the model exhibits undesirable behavior arising from the training data, regardless of any individual's request. This includes: Removal of toxic, biased, or unsafe outputs introduced by harmful content in the training set Correction of stale or factually incorrect associations, such as outdated knowledge encoded in a deployed model Removal of dangerous capabilities, such as detailed knowledge of the synthesis of chemical or biological agents Correction of the influence of data poisoning or adversarial attacks that have corrupted model behavior This second category has been formalized as corrective machine unlearning, which frames unlearning as a post-training mechanism for repairing the effects of bad or harmful training data. It is closely related to the AI safety literature, where data filtering alone has been found insufficient to prevent hazardous knowledge from being encoded in model weights, motivating unlearning as a complementary risk mitigation strategy. A further distinction has been drawn in the literature between removal {eliminating the influence of specific training data on model parameters) and suppression (preventing the model from generating specific outputs regardless of how that knowledge is encoded). These two goals are not equivalent: removing training data does not guarantee meaningful output suppression, and suppressing outputs does not constitute removal of the underlying training data's influence. == SISA Training == SISA is a training strategy consisting of four mechanisms designed to make machine unlearning more efficient by structuring how models are trained and updated. Its goal is to allow a system to remove the influence of specific data points without retraining an entire model from scratch. By reorganizing training data and workflows, SISA reduces the computational burden of unlearning requests. Sharding divides the training dataset into multiple disjoint subsets, or shards. Each shard is used to train a separate model instance. This ensures that a single data point affects only one shard, so unlearning it requires updating only the corresponding shard rather than the full model. Isolation refers to training each shard independently, with nothing shared across shards during the training process. This separation prevents cross-contamination between shards, ensuring that forgetting data in one shard does not require adjustments to any others. Slicing breaks the data within each shard into sequential slices and stores model states after each slice is trained on. When an unlearning request targets a piece of data, the system can roll back to the checkpoint before the point was seen and retrain only from that slice forward. This reduces retraining time even within a shard. Aggregation occurs at inference, when the model is queried. It combines the outputs of each shard to determine the output of the overall model. This is often through majority voting or averaging. This allows SISA-trained systems to behave like a single model despite being composed of multiple shard-level models. Together, these mechanisms enable machine learning systems to forget specific data points with far lower computational cost than full retraining. The trade-off is that sharding and slicing can lead to reduced model accuracy, worse generalization, and increased storage requirements for the intermediate checkpoints. This can be tolerable based on the needs of the individual or organization to comply with "right to be forgotten" or efficiently recover from backdoor attacks. == Algorithms == Machine unlearning algorithms are broadly categorized into exact and approximate methods, reflecting a fundamental trade-off between formal guarantees and computational tractability. === Exact Unlearning === Exact unlearning methods produce a model that is statistically indistinguishable from one retrained from scratch on the dataset with the forget data removed. The canonical framework for exact unlearning is SISA Training (Sharded, Isolated, Sliced, and Aggregated), introduced by Bourtoule et al. (2021). SISA partitions the training dataset into disjoint shards and trains a separate sub-model on each. At inference time, predictions are aggregated across sub-models. When an unlearning request is received, only the sub-model corresponding to the shard containing the target data requires retraining, reducing computational overhead proportionally to the number of shards. Exact methods provide the strongest guarantees but become prohibitively expensive for large pre-trained neural networks and are generally limited to settings where training can be structured in advance. === Approximate Unlearning === Approximate unlearning methods seek to produce a model whose behavior is sufficiently close to an exactly unlearned model without the cost of full retraining. These methods dominate practical applications. Common approaches include: Gradient Ascent: The model is fine-tuned by maximizing the loss on the forget set, directly degrading its performance on targeted data. This is the most direct approach but risks destabilizing performance on retained data. Random Labelling: The model is fine-tuned on the forget set using randomly shuffled labels, confusing its associations with the targeted data while producing a less aggressive weight shift than pure gradient ascent. Gradient Difference: Combines gradient ascent on the forget set with simultaneous gradient descent on the retain set, using the retain objective as a regularizer to preserve general model utility. KL Divergence Regularization: Minimizes the KL divergence between the outputs of the unlearned model and the original model on the retain set, anchoring behavior on data the model should remember. Weight Pruning and Fine-tuning: Parameters with the smallest L1-norm are pruned — targeting weights most weakly associated with general knowledge and potentially most associated with the forget set — followed by fine-tuning on the retain set to restore utility. Layer Reset and Fine-tuning: The first or last k layers are re-initialized to random weights and the model is subsequently fine-tuned on the retain set. This is a coarse but computationally simple approach. Selective Synaptic Dampening: Uses influence functions to estimate the effect of individual trainin
Read more →
A capsule neural network (CapsNet) is a machine learning system that is a type of artificial neural network (ANN) that can be used to better model hierarchical relationships. The approach is an attempt to more closely mimic biological neural organization. The idea is to add structures called "capsules" to a convolutional neural network (CNN), and to reuse output from several of those capsules to form more stable (with respect to various perturbations) representations for higher capsules. The output is a vector consisting of the probability of an observation, and a pose for that observation. This vector is similar to what is done for example when doing classification with localization in CNNs. Among other benefits, capsnets address the "Picasso problem" in image recognition: images that have all the right parts but that are not in the correct spatial relationship (e.g., in a "face", the positions of the mouth and one eye are switched). For image recognition, capsnets exploit the fact that while viewpoint changes have nonlinear effects at the pixel level, they have linear effects at the part/object level. This can be compared to inverting the rendering of an object of multiple parts. == History == In 2000, Geoffrey Hinton et al. described an imaging system that combined segmentation and recognition into a single inference process using parse trees. So-called credibility networks described the joint distribution over the latent variables and over the possible parse trees. That system proved useful on the MNIST handwritten digit database. A dynamic routing mechanism for capsule networks was introduced by Hinton and his team in 2017. The approach was claimed to reduce error rates on MNIST and to reduce training set sizes. Results were claimed to be considerably better than a CNN on highly overlapped digits. In Hinton's original idea one minicolumn would represent and detect one multidimensional entity. == Transformations == An invariant is an object property that does not change as a result of some transformation. For example, the area of a circle does not change if the circle is shifted to the left. Informally, an equivariant is a property that changes predictably under transformation. For example, the center of a circle moves by the same amount as the circle when shifted. A nonequivariant is a property whose value does not change predictably under a transformation. For example, transforming a circle into an ellipse means that its perimeter can no longer be computed as π times the diameter. In computer vision, the class of an object is expected to be an invariant over many transformations. I.e., a cat is still a cat if it is shifted, turned upside down or shrunken in size. However, many other properties are instead equivariant. The volume of a cat changes when it is scaled. Equivariant properties such as a spatial relationship are captured in a pose, data that describes an object's translation, rotation, scale and reflection. Translation is a change in location in one or more dimensions. Rotation is a change in orientation. Scale is a change in size. Reflection is a mirror image. Unsupervised capsnets learn a global linear manifold between an object and its pose as a matrix of weights. In other words, capsnets can identify an object independent of its pose, rather than having to learn to recognize the object while including its spatial relationships as part of the object. In capsnets, the pose can incorporate properties other than spatial relationships, e.g., color (cats can be of various colors). Multiplying the object by the manifold poses the object (for an object, in space). == Pooling == Capsnets reject the pooling layer strategy of conventional CNNs that reduces the amount of detail to be processed at the next higher layer. Pooling allows a degree of translational invariance (it can recognize the same object in a somewhat different location) and allows a larger number of feature types to be represented. Capsnet proponents argue that pooling: violates biological shape perception in that it has no intrinsic coordinate frame; provides invariance (discarding positional information) instead of equivariance (disentangling that information); ignores the linear manifold that underlies many variations among images; routes statically instead of communicating a potential "find" to the feature that can appreciate it; damages nearby feature detectors, by deleting the information they rely upon. == Capsules == A capsule is a set of neurons that individually activate for various properties of a type of object, such as position, size and hue. Formally, a capsule is a set of neurons that collectively produce an activity vector with one element for each neuron to hold that neuron's instantiation value (e.g., hue). Graphics programs use instantiation value to draw an object. Capsnets attempt to derive these from their input. The probability of the entity's presence in a specific input is the vector's length, while the vector's orientation quantifies the capsule's properties. Artificial neurons traditionally output a scalar, real-valued activation that loosely represents the probability of an observation. Capsnets replace scalar-output feature detectors with vector-output capsules and max-pooling with routing-by-agreement. Because capsules are independent, when multiple capsules agree, the probability of correct detection is much higher. A minimal cluster of two capsules considering a six-dimensional entity would agree within 10% by chance only once in a million trials. As the number of dimensions increase, the likelihood of a chance agreement across a larger cluster with higher dimensions decreases exponentially. Capsules in higher layers take outputs from capsules at lower layers, and accept those whose outputs cluster. A cluster causes the higher capsule to output a high probability of observation that an entity is present and also output a high-dimensional (20-50+) pose. Higher-level capsules ignore outliers, concentrating on clusters. This is similar to the Hough transform, the RHT and RANSAC from classic digital image processing. == Routing by agreement == The outputs from one capsule (child) are routed to capsules in the next layer (parent) according to the child's ability to predict the parents' outputs. Over the course of a few iterations, each parents' outputs may converge with the predictions of some children and diverge from those of others, meaning that that parent is present or absent from the scene. For each possible parent, each child computes a prediction vector by multiplying its output by a weight matrix (trained by backpropagation). Next the output of the parent is computed as the scalar product of a prediction with a coefficient representing the probability that this child belongs to that parent. A child whose predictions are relatively close to the resulting output successively increases the coefficient between that parent and child and decreases it for parents that it matches less well. This increases the contribution that that child makes to that parent, thus increasing the scalar product of the capsule's prediction with the parent's output. After a few iterations, the coefficients strongly connect a parent to its most likely children, indicating that the presence of the children imply the presence of the parent in the scene. The more children whose predictions are close to a parent's output, the more quickly the coefficients grow, driving convergence. The pose of the parent (reflected in its output) progressively becomes compatible with that of its children. The coefficients' initial logits are the log prior probabilities that a child belongs to a parent. The priors can be trained discriminatively along with the weights. The priors depend on the location and type of the child and parent capsules, but not on the current input. At each iteration, the coefficients are adjusted via a "routing" softmax so that they continue to sum to 1 (to express the probability that a given capsule is the parent of a given child.) Softmax amplifies larger values and diminishes smaller values beyond their proportion of the total. Similarly, the probability that a feature is present in the input is exaggerated by a nonlinear "squashing" function that reduces values (smaller ones drastically and larger ones such that they are less than 1). This dynamic routing mechanism provides the necessary deprecation of alternatives ("explaining away") that is needed for segmenting overlapped objects. This learned routing of signals has no clear biological equivalent. Some operations can be found in cortical layers, but they do not seem to relate this technique. === Math/code === The pose vector u i {\textstyle \mathbf {u} _{i}} is rotated and translated by a matrix W i j {\textstyle \mathbf {W} _{ij}} into a vector u ^ j | i {\textstyle \mathbf {\hat {u}} _{j|i}} that predicts the output of the parent capsule. u ^ j | i = W i j u i {\displaystyle \mathbf {
Read more →
Conflict resolution strategies are used in production systems in artificial intelligence, such as in rule-based expert systems, to help in choosing which production rule to fire. The need for such a strategy arises when the conditions of two or more rules are satisfied by the currently known facts. == Categories == Conflict resolution strategies fall into several main categories. They each have advantages which form their rationales. Specificity - If all of the conditions of two or more rules are satisfied, choose the rule according to how specific its conditions are. It is possible to favor either the more general or the more specific case. The most specific may be identified roughly as the one having the greatest number of preconditions. This usefully catches exceptions and other special cases before firing the more general (default) rules. Recency - When two or more rules could be chosen, favor the one that matches the most recently added facts, as these are most likely to describe the current situation. Not previously used - If a rule's conditions are satisfied, but previously the same rule has been satisfied by the same facts, ignore the rule. This helps to prevent the system from entering infinite loops. Order - Pick the first applicable rule in order of presentation. This is the strategy that Prolog interpreters use by default, but any strategy may be implemented by building suitable rules in a Prolog system. Arbitrary choice - Pick a rule at random. This has the merit of being simple to compute.
Read more →
CatBoost is an open-source software library developed by Yandex. It provides a gradient boosting framework which, among other features, attempts to solve for categorical features using a permutation-driven alternative to the classical algorithm. It works on Linux, Windows, macOS, and is available in Python, R, and models built using CatBoost can be used for predictions in C++, Java, C#, Rust, Core ML, ONNX, and PMML. The source code is licensed under Apache License and available on GitHub. InfoWorld magazine awarded the library "The best machine learning tools" in 2017. along with TensorFlow, Pytorch, XGBoost and 8 other libraries. Kaggle listed CatBoost as one of the most frequently used machine learning (ML) frameworks in the world. It was listed as the top-8 most frequently used ML framework in the 2020 survey and as the top-7 most frequently used ML framework in the 2021 survey. As of April 2022, CatBoost is installed about 100000 times per day from PyPI repository == Features == CatBoost has gained popularity compared to other gradient boosting algorithms primarily due to the following features Native handling for categorical features Fast GPU training Visualizations and tools for model and feature analysis Using oblivious trees or symmetric trees for faster execution Ordered boosting to overcome overfitting == History == In 2009 Andrey Gulin developed MatrixNet, a proprietary gradient boosting library that was used in Yandex to rank search results. Since 2009 MatrixNet has been used in different projects at Yandex, including recommendation systems and weather prediction. In 2014–2015 Andrey Gulin worked with a team of researchers to start a new project called Tensornet which was aimed at solving the problem of "how to work with categorical data". Their work resulted in several proprietary Gradient Boosting libraries with different approaches to handling categorical data. In 2016 the Machine Learning Infrastructure team led by Anna Dorogush started working on Gradient Boosting in Yandex, including Matrixnet and Tensornet. They implemented and open-sourced the next version of Gradient Boosting library called CatBoost, which has support for categorical and text data, GPU training, model analysis, and visualization tools. CatBoost was open-sourced in July 2017 and is under active development in Yandex and the open-source community. == Application == JetBrains uses CatBoost for code completion Cloudflare uses CatBoost for bot detection Careem uses CatBoost to predict future destinations of the rides
Read more →
RTX Video Super Resolution (RTX VSR) is a video scaling feature by Nvidia. It was released on February 28, 2023. == History == The feature was first unveiled during CES 2023 as RTX Video Super Resolution. It uses the on-board Tensor Cores to upscale browser video content in real time. Video Super Resolution was initially only available on RTX 30 and 40 series GPUs, while support for 20 series GPUs was added afterwards; it is now available on all Nvidia RTX-branded GPUs. The feature supports input resolutions from 360p to 1440p and a max output of 4K and comes without support for HDR content although that could be likely added in the future. Nvidia released RTX Video Super Resolution 1.5 with improved video quality and RTX 20 series support on October 17, 2023. == Reception == According to ComputerBase, although "the algorithm is not yet working flawlessly", the feature is "overall recommendable".
Read more →
Illia Polosukhin is a Ukrainian-born computer scientist and entrepreneur known for his work on the transformer architecture in machine learning and for co-founding the NEAR blockchain. == Early life and education == Polosukhin studied at the Kharkiv Polytechnic Institute, later relocating to San Diego and then moving to Silicon Valley. == Career == === Google and transformer research === Polosukhin worked at Google and was part of the team associated with research on self-attention that culminated in the 2017 paper Attention Is All You Need, widely credited with introducing the transformer architecture used in modern large language models. === NEAR Protocol === After his work in machine learning, Polosukhin became a co-founder of NEAR Protocol and later associated with the NEAR Foundation ecosystem. In 2023, Polosukhin publicly argued that increasingly capable A.I. systems should be more transparent and user-controlled, and expressed skepticism that conventional regulation alone would solve problems created by closed, corporate models, warning about risks such as regulatory capture. He has promoted “user-owned AI” concepts that combine open approaches with decentralized infrastructure aligned with the blockchain technology. In 2024, Polosukhin downplayed scenarios of A.I. independently causing human extinction, arguing that conflicts are driven by people and that misuse of AI would reflect human intent and incentives. Later this year, Polosukhin said the NEAR Foundation would reduce its workforce by about 40%. == Publications == Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Lukasz Kaiser, Illia Polosukhin; et al. (2017). "Attention Is All You Need". arXiv.{{cite journal}}: CS1 maint: multiple names: authors list (link)
Read more →
Visual hierarchy, in Gestalt psychology, describes how particular elements in a visual field stand out more than others in a pattern, creating a perceived order of importance. Although it can occur naturally, the term is most often used in design—especially graphic design and cartography—where elements are arranged to appear more important than others. This order is created by the visual contrast between forms in a field of perception. Objects with highest contrast to their surroundings are recognized first by the human mind. == Evidence == There is some scientific evidence for visual hierarchy using eye tracking. For example, one study found that when people agree that a graphic design is good, they exhibit more similar eye movements; measured by the Fréchet distance. == Theory == The concept of visual hierarchy is based in Gestalt psychological theory, an early 20th-century German theory that proposes that the human brain has innate organizing tendencies that “structure individual elements, shapes or forms into a coherent, organized whole,” especially when processing visual information. The German word Gestalt translates into “form,” “pattern,” or “shape” in English. When an element in a visual field disconnects from the ‘whole’ created by the brain's perceptual organization, it “stands out” to the viewer. The shapes that disconnect most severely from their surroundings stand out the most. This is commonly encapsulated as the Von Restorff effect, which states that isolation attracts attention. === Physical characteristics === The brain distinguishes objects based on differences in their physical appearances. These characteristics fall into four categories: color, size, alignment, and character. Each type of contrast can be used to construct a visual hierarchy. The same characteristics are also sometimes categorized (especially among cartographers) according to the visual variables of Jacques Bertin. Color encompasses the hue, saturation, value, and perceived texture of forms. Dark figures will stand out on a light background, light figures will stand out on a dark background, brightly colored figures will stand out on a muted background, and so on. The fluorescent colors used for tennis balls and other sports equipment is intended to make them instantly stand out against almost any natural visual field. Size has a strong influence on visual hierarchy. Large elements typically attract attention, provided that they can be recognized as figures. Alignment is the arrangement of forms relative to one another. For example, items in the upper left corner of a page are often seen first (at least for those readers accustomed to western languages), the center of the field has prominence. Negative space can also be employed: a figure isolated among large amounts of white space will stand out more than one amid other figures. Character includes several kinds of contrasts based on shape. For example, complex patterns attract more attention than simple or predictable patterns, intricate shapes attract more attention than generalized ones. Even large-scale patterns can attract attention if they contrast with the pattern in the remainder of the visual field. Camouflage is an example of eliminating contrast in character in color and/or character specifically to reduce visual hierarchy. The "squint test" is often suggested as a simple, if unscientific, method to evaluate the visual hierarchy of a graphical product like a map or web page. When viewed out of focus (or from a great distance), the viewer is not distracted by details, but can only see overall (gestalt) patterns such as visual hierarchy. All of the above patterns, except some aspects of character, are recognizable by this method. == Application == Visual hierarchy is an important concept in the field of graphic design, a field that specializes in visual organization. Designers attempt to control visual hierarchy to guide the eye to information in a specific order for a specific purpose. One could compare visual hierarchy in graphic design to grammatical structure in writing in terms of the importance of each principle to these fields. === Cartography === In cartographic design, visual hierarchy is used to emphasize certain important features on a map over less important features. Typically, a map has a purpose that dictates a conceptual hierarchy of what should be more or less important, so one of the goals of the choice of map symbols is to match the visual hierarchy to the conceptual hierarchy. The Visual hierarchy of a map may apply to individual geographic features (such as making a single country stand out), to map layers of related features (e.g., making lakes stand out more than roads), and to the entire layout of map and non-map elements (e.g., making the title look more important than the scale bar). Like the main map elements, such features have weight, and the properties that apply to visual hierarchy of map layers also apply to other elements on the page. Size and alignment are the two main determinants of the visual hierarchy for these features. Cartographers often utilize principles of negative space and figure-ground contrast to design an appropriate visual hierarchy by employing contrast between unused space and layout features. === User experience design and behavioral design === In user experience design and behavioural design, such as web design, visual hierarchy is used to prioritize navigational structures and content, so that audiences focus on elements that facilitate system usage, or increases the chance that they notice content that contains psychological nudges. Color is one of many factors used in the design of a visual hierarchy, and a key factor due to the high salience of color perception.
Read more →