AI Chatbots and Assistants

Explore the best AI Chatbots and Assistants — independent reviews, comparisons, pricing and step-by-step how-to guides, curated by Aizhi.

• 

  Zero-knowledge service

  In cloud computing, the term zero-knowledge (or occasionally no-knowledge or zero-access) is a commonly used term for online services that store, transfer or manipulate data with a high level of confidentiality, where the data is only accessible to the data's owner (the client), and not to the service provider. However, unlike "end-to-end encryption", the term "zero-knowledge" does not imply any specific threat model or security notion, and its use is commonly frowned-upon by the security community. The term "zero-knowledge" was popularized by backup service SpiderOak, which later switched to using the term "no knowledge", acknowledging that the previous terminology was not technically accurate. == Disadvantages == Most cloud storage services keep a copy of the client's password on their servers, allowing clients who have lost their passwords to retrieve and decrypt their data using alternative means of authentication; but since zero-knowledge services do not store copies of clients' passwords, if a client loses their password then their data cannot be decrypted, making it practically unrecoverable. Most of the most used cloud storage services, such as Google Drive, Dropbox, OneDrive or iCloud, are also able to furnish access requests from law enforcement agencies for similar reasons; zero-knowledge services, however, are unable to do so, since their systems are designed to make clients' data inaccessible without the client's explicit cooperation.

  Read more →
• 

  Causal AI

  Causal AI is a technique in artificial intelligence that builds a causal model and can thereby make inferences using causality rather than just correlation. One practical use for causal AI is for organisations to explain decision-making and the causes for a decision. Systems based on causal AI, by identifying the underlying web of causality for a behaviour or event, provide insights that solely predictive AI models might fail to extract from historical data. An analysis of causality may be used to supplement human decisions in situations where understanding the causes behind an outcome is necessary, such as quantifying the impact of different interventions, policy decisions or performing scenario planning. A 2024 paper from Google DeepMind demonstrated mathematically that "Any agent capable of adapting to a sufficiently large set of distributional shifts must have learned a causal model". The paper offers the interpretation that learning to generalise beyond the original training set requires learning a causal model, concluding that causal AI is necessary for artificial general intelligence. == History == The concept of causal AI and the limits of machine learning were raised by Judea Pearl, the Turing Award-winning computer scientist and philosopher, in 2018's The Book of Why: The New Science of Cause and Effect. Pearl asserted: “Machines' lack of understanding of causal relations is perhaps the biggest roadblock to giving them human-level intelligence.” In 2020, Columbia University established a Causal AI Lab under Director Elias Bareinboim. Professor Bareinboim's research focuses on causal and counterfactual inference and their applications to data-driven fields in the health and social sciences as well as artificial intelligence and machine learning. Technological research and consulting firm Gartner for the first time included causal AI in its 2022 Hype Cycle report, citing it as one of five critical technologies in accelerated AI automation. Causal AI is closely related to but distinct from fields such as causal inference, explainable AI and causal reasoning. While causal inference focuses on estimating cause-effect relationships (often from observational data), causal AI emphasises the integration of those causal models into AI systems for prediction, planning and adaptation.

  Read more →
• 

  Documentation

  Documentation is any communicable material that is used to describe, explain, or instruct regarding some attributes of an object, system, or procedure, such as its parts, assembly, installation, maintenance, and use. As a form of knowledge management and knowledge organization, documentation can be provided on paper, online, or on digital or analog media, such as audio tape or CDs. Examples of such resources include user guides, white papers, online help, and quick-reference guides. Paper or hard-copy documentation has become less common. Contemporary documentation is often distributed through websites, software products, and other online applications. Documentation, understood as a set of instructional materials, should not be confused with documentation science, which is the study of the recording and retrieval of information. == Principles for producing documentation == While associated International Organization for Standardization (ISO) standards are not easily available publicly, a guide from other sources for this topic may serve the purpose. Documentation development may involve document drafting, formatting, submitting, reviewing, approving, distributing, reposting and tracking, etc., and are convened by associated standard operating procedure in a regulatory industry. It could also involve creating content from scratch. Documentation should be easy to read and understand. If it is too long and too wordy, it may be misunderstood or ignored. Clear, concise words should be used, and sentences should be limited to a maximum of 15 words. Documentation intended for a general audience should avoid gender-specific terms and cultural biases. In a series of procedures, steps should be clearly numbered. == Producing documentation == Technical writers and corporate communicators are professionals whose field and work is documentation. Ideally, technical writers have a background in both the subject matter and also in writing, managing content, and information architecture. Technical writers more commonly collaborate with subject-matter experts, such as engineers, technical experts, medical professionals, etc. to define and then create documentation to meet the user's needs. Corporate communications includes other types of written documentation, for example: Market communications (MarCom): MarCom writers endeavor to convey the company's value proposition through a variety of print, electronic, and social media. This area of corporate writing is often engaged in responding to proposals. Technical communication (TechCom): Technical writers document a company's product or service. Technical publications can include user guides, installation and configuration manuals, and troubleshooting and repair procedures. Legal writing: This type of documentation is often prepared by attorneys or paralegals. Compliance documentation: This type of documentation codifies standard operating procedures, for any regulatory compliance needs, as for safety approval, taxation, financing, and technical approval. Healthcare documentation: This field of documentation encompasses the timely recording and validation of events that have occurred during the course of providing health care. == Documentation in computer science == === Types === The following are typical software documentation types: Request for proposal Requirements/statement of work/scope of work Software design and functional specification System design and functional specifications Change management, error and enhancement tracking User acceptance testing Manpages The following are typical hardware and service documentation types: Network diagrams Network maps Datasheet for IT systems (server, switch, e.g.) Service catalog and service portfolio (Information Technology Infrastructure Library) === Software Documentation Folder (SDF) tool === A common type of software document written in the simulation industry is the SDF. When developing software for a simulator, which can range from embedded avionics devices to 3D terrain databases by way of full motion control systems, the engineer keeps a notebook detailing the development "the build" of the project or module. The document can be a wiki page, Microsoft Word document or other environment. They should contain a requirements section, an interface section to detail the communication interface of the software. Often a notes section is used to detail the proof of concept, and then track errors and enhancements. Finally, a testing section to document how the software was tested. This documents conformance to the client's requirements. The result is a detailed description of how the software is designed, how to build and install the software on the target device, and any known defects and workarounds. This build document enables future developers and maintainers to come up to speed on the software in a timely manner, and also provides a roadmap to modifying code or searching for bugs. === Software tools for network inventory and configuration === These software tools can automatically collect data of your network equipment. The data could be for inventory and for configuration information. The Information Technology Infrastructure Library requests to create such a database as a basis for all information for the IT responsible. It is also the basis for IT documentation. Examples include XIA Configuration. == Documentation in criminal justice == "Documentation" is the preferred term for the process of populating criminal databases. Examples include the National Counterterrorism Center's Terrorist Identities Datamart Environment, sex offender registries, and gang databases. == Documentation in early childhood education == Documentation, as it pertains to the early childhood education field, is "when we notice and value children's ideas, thinking, questions, and theories about the world and then collect traces of their work (drawings, photographs of the children in action, and transcripts of their words) to share with a wider community". Thus, documentation is a process, used to link the educator's knowledge and learning of the child/children with the families, other collaborators, and even to the children themselves. Documentation is an integral part of the cycle of inquiry - observing, reflecting, documenting, sharing and responding. Pedagogical documentation, in terms of the teacher documentation, is the "teacher's story of the movement in children's understanding". According to Stephanie Cox Suarez in "Documentation - Transforming our Perspectives", "teachers are considered researchers, and documentation is a research tool to support knowledge building among children and adults". Documentation can take many different styles in the classroom. The following exemplifies ways in which documentation can make the research, or learning, visible: Documentation panels (bulletin-board-like presentation with multiple pictures and descriptions about the project or event). Daily log (a log kept every day that records the play and learning in the classroom) Documentation developed by or with the children (when observing children during documentation, the child's lens of the observation is used in the actual documentation) Individual portfolios (documentation used to track and highlight the development of each child) Electronic documentation (using apps and devices to share documentation with families and collaborators) Transcripts or recordings of conversations (using recording in documentation can bring about deeper reflections for both the educator and the child) Learning stories (a narrative used to "describe learning and help children see themselves as powerful learners") The classroom as documentation (reflections and documentation of the physical environment of a classroom). Documentation is certainly a process in and of itself, and it is also a process within the educator. The following is the development of documentation as it progresses for and in the educator themselves: Develop(s) habits of documentation Become(s) comfortable with going public with recounting of activities Develop(s) visual literacy skills Conceptualize(s) the purpose of documentation as making learning styles visible, and Share(s) visible theories for interpretation purposes and further design of curriculum.

  Read more →
• 

  Aidoc

  Aidoc Medical is an Israeli technology company that develops computer-aided simple triage and notification systems. Aidoc has obtained U.S. Food and Drug Administration and CE mark approval for its stroke, pulmonary embolism, cervical fracture, intracranial hemorrhage, intra-abdominal free gas, and incidental pulmonary embolism algorithms. Aidoc algorithms are in use in more than 900 hospitals and imaging centers, including Montefiore Nyack Hospital, LifeBridge Health, LucidHealth, Yale New Haven Hospital, Cedars-Sinai Medical Center, University of Rochester Medical Center, and Sheba Medical Center. == History == Aidoc was founded in 2016 by Elad Walach as the CEO, Michael Braginsky as the CTO and Guy Reiner as the VP. In April 2017, the company raised $7M, led by TLV Partners, and in April 2019, the company raised another $27M, led by Square Peg capital. There have been several additional rounds of funding as well, bringing Aidoc's total investment to $370M as of July 2025. In August 2018, Aidoc gained FDA clearance for its intracranial hemorrhage system, and in May 2019 it received clearance for the pulmonary embolism system. In January 2020, the system for detecting large-vessel occlusions (LVOs) in head CTA examinations obtained FDA clearance. In October 2024, it was reported that Aidoc is working with NVIDIA to develop a framework for deployment and integration of artificial intelligence tools in healthcare. The Blueprint for Resilient Integration and Deployment of Guided Excellence (BRIDGE) is a guideline to facilitate AI adoption in the healthcare industry. == Products and market == Aidoc has developed a suite of artificial intelligence products that flag both time-sensitive and time-consuming (for the radiologist) abnormalities across the body. The algorithms are developed with large quantities of data to provide diagnostic aid for a broad set of pathologies. The company offers an array of algorithms that span across the body, including for intracranial hemorrhage, spine fractures (C, T & L), free air in the abdomen, pulmonary embolism, and more. It developed "Always-on AI", a term coined by Elad Walach that refers to a type of artificial intelligence that is "Always-on—constantly running in the background and automatically analyzing medical imaging data, identifying urgent findings, and sparing radiologists from "drowning" in vast amounts of irrelevant data. Aidoc's solutions cover medical conditions prevalent in all settings (ED/inpatient/outpatient), including level 1 trauma centers, outpatient imaging centers, teleradiology groups and, are set up in over 200 medical centers worldwide. Notable customers include the University of Rochester Medical Center and Global Diagnostics Australia. Aidoc announced in 2024 that its new Clinical AI Reasoning Engine (CARE1) had been submitted for FDA approval. In September 2025 Aidoc received a "Breakthrough Device Designation" from the FDA for a new multi-triage solution that spans numerous acute findings in CT scans. Aidoc's CARE1 foundation model was the basis of the workflow on which the designation was made, enabling simultaneous coverage of multiple pathologies. This new designation allows parallel FDA review of multiple indications under a single submission. In April 2026, Aidoc raised million in a Series E funding round led by Growth Equity at Goldman Sachs Alternatives, with participation from General Catalyst and NVentures. The financing brought the company's total funding to over million. == Clinical Research == A clinical study on Aidoc’ accuracy of deep convolutional neural networks for the detection of pulmonary embolism (PE) on CT pulmonary angiograms (CTPAs) was performed by the University Hospital of Basel and presented at the European Congress of Radiology, showing that the Aidoc algorithm reached 93% sensitivity and 95% specificity. Clinical research has also been performed to test the diagnostic performance of Aidoc's deep learning-based triage system for the flagging of acute findings in abdominal computed tomography (CT) examinations. Overall, the algorithm achieved 93% sensitivity (91/98, 7 false negatives) and 97% specificity (93/96, 3 false-positive) in the detection of acute abdominal findings. Additional clinical research on Aidoc's Intracranial hemorrhage algorithm accuracy was presented at the European Congress of Radiology by Antwerp University Hospital, evaluating the use of its deep learning algorithm for the detection of intracranial hemorrhage on non-contrast enhanced CT of the brain. The University of Washington completed a study on the accuracy of Aidoc's intracranial hemorrhage algorithm.

  Read more →
• 

  Medical data breach

  Medical data, including patients' identity information, health status, disease diagnosis and treatment, and biogenetic information, not only involve patients' privacy but also have a special sensitivity and important value, which may bring physical and mental distress and property loss to patients and even negatively affect social stability and national security once leaked. However, the development and application of medical AI must rely on a large amount of medical data for algorithm training, and the larger and more diverse the amount of data, the more accurate the results of its analysis and prediction will be. However, the application of big data technologies such as data collection, analysis and processing, cloud storage, and information sharing has increased the risk of data leakage. In the United States, the rate of such breaches has increased over time, with 176 million records breached by the end of 2017. By 2024, the U.S. Department of Health and Human Services reported 725 large healthcare data breaches affecting approximately 275 million individual records in a single year, marking a significant escalation in both the frequency and scale of incidents. == Black market for health data == In February 2015 an NPR report claimed that organized crime networks had ways of selling health data in the black market. In 2015 a Beazley employee estimated that medical records could sell on the black market for US$40-50. == How data is lost == Theft, data loss, hacking, and unauthorized account access are ways in which medical data breaches happen. Among reported breaches of medical information in the United States networked information systems accounted for the largest number of records breached. There are many data breaches happening in the US health care system, among business associates of the health care providers that continuously gain access to patients' data. == List of data breaches == In February 2024, a ransomware attack on Change Healthcare, a subsidiary of UnitedHealth Group, compromised the protected health information of approximately 100 million individuals, making it the largest healthcare data breach in United States history. The attack disrupted claims processing for healthcare providers nationwide for several weeks. In May 2024, MediSecure suffered a cyberattack involving ransomware in Australia. In May 2021, the Health Service Executive in the Republic of Ireland was the victim of a cyberattack involving ransomware, in the Health Service Executive cyberattack, with admission records and test results present in a sample of the data reviewed by the Financial Times. In October 2018, the Centers for Medicare and Medicaid Services in the US reported that around 75,000 individual records had been affected by a data breach that took place through the ACA Agent and Broker Portal. In 2018, Social Indicators Research published the scientific evidence of 173,398,820 (over 173 million) individuals affected in USA from October 2008 (when the data were collected) to September 2017 (when the statistical analysis took place). In 2015, Anthem Inc. lost data for 37 million people in the Anthem medical data breach In 2014 4.5 million people using Complete Health Systems had their data stolen In 2013-14 1 million people using Montana Department of Public Health and Human Services had their data stolen In 2013 4 million people using Advocate Health and Hospitals Corporation had their data stolen In 2011 4.9 million users of Tricare services had their data stolen due to an employee error by Science Applications International Corporation In 2011 1.9 million people using Health Net had their data stolen In 2011 1 million people using Nemours Foundation had their data stolen In 2010 6800 people using New York-Presbyterian Hospital and Columbia University Medical Center had their data breached. In response, those organizations agreed to pay the United States Department of Health and Human Services a US$4.8 million dollar fine. In 2009 1 million people using BlueCross BlueShield of Tennessee had their data stolen == Regulation == In the United States, the Health Insurance Portability and Accountability Act and Health Information Technology for Economic and Clinical Health Act require companies to report data breaches to affected individuals and the federal government. Under the HIPAA Breach Notification Rule, covered entities must notify affected individuals without unreasonable delay and no later than 60 days after discovering a breach of unsecured protected health information. Breaches affecting 500 or more individuals must also be reported to the HHS Secretary and to prominent media outlets serving the affected state or jurisdiction within the same timeframe; HHS publicly lists these larger breaches on its breach portal, commonly known as the "wall of shame." Breaches affecting fewer than 500 individuals are reported to HHS annually, no later than 60 days after the end of the calendar year in which they were discovered. Health Information Privacy Health Insurance Portability and Accountability Act of 1996 (HIPAA). - 45 CFR Parts 160 and 164, Standards for Privacy of Individually Identifiable Health Information and Security Standards for the Protection of Electronic Protected Health Information. HIPAA includes provisions designed to save health care businesses money by encouraging electronic transactions, as well as regulations to protect the security and confidentiality of patient information. The Privacy Rule became effective April 14, 2001, and most covered entities (health plans, health care clearinghouses, and health care providers that conduct certain financial and administrative transactions electronically) had until April 2003 to comply. This security provision became effective April 21, 2003. The Health Insurance Portability and Accountability Act (HIPAA) is the baseline set of federal regulations governing medical information. It does three things: i. i. i.Establish a structure for how personal health information is disclosed and establish the rights of individuals with respect to health information; ii.Specify security standards for the retention and transmission of electronic patient information; iii.Need a common format and data structure for the electronic exchange of health information. California-Specific Laws California’s medical privacy laws, primarily the Confidentiality of Medical Information Act (CMIA), the data breach sections of the Civil Code, and sections of the Health and Safety Code, provide HIPAA-like protections, although the terminology is different. HIPAA establishes a federal "minimum standard" that applies where there are gaps in California law, and HIPAA also specifies that stricter state laws will override or supersede HIPAA. California's health care privacy laws apply to providers who provide personal health records (PHR), while HIPAA only applies when the provider providing the PHR is a business associate of a covered entity. Federal law does not grant individuals the right to file a lawsuit in the event of a data breach (only the Attorney General can file a lawsuit), but California law does. This means that California law sets a higher standard for medical privacy, and that individuals in California enjoy stronger legal protections and more ways to hold entities that violate their medical privacy accountable. In the UK, the legal framework for how patient data is cared for and processed is the Data Protection Act 2018 (DPA), which incorporates the EU General Data Protection Regulation (GDPR) into law, and the common law duty of confidentiality (CLDC). The data protection legislation requires that the collection and processing of personal data be fair, lawful and transparent. This means that the collection and processing of data as defined by data protection legislation must always have a valid lawful basis and must also meet the requirements of the CLDC. In the China, Article 18 of the "National Health Care Big Data Standards, Security and Services Management Measures (for Trial Implementation)" (National Health Planning and Development (2018) No. 23) promulgated by the National Health Care Commission in 2018 states, "The responsible unit shall adopt measures such as data classification, important data backup, and encryption authentication to guarantee the security of health care big data." However, the scope and definition of important data are not covered. Although the "Information Security Technology-Healthcare Data Security Guide" (the "Guide") issued by the National Standardization Committee also proposes that important data should be evaluated and approved in accordance with the regulations, there is likewise no definition of the connotation and definition of important data.

  Read more →
• 

  Artificial intelligence in industry

  Industrial artificial intelligence, or industrial AI, refers to the application of artificial intelligence to industrial business processes. Unlike general artificial intelligence which is a frontier research discipline to build computerized systems that perform tasks requiring human intelligence, industrial AI is more concerned with the application of such technologies to address industrial pain-points for customer value creation, productivity improvement, cost reduction, site optimization, predictive analysis and insight discovery. Artificial intelligence and machine learning have become key enablers to leverage data in production in recent years due to a number of different factors: More affordable sensors and the automated process of data acquisition; More powerful computation capability of computers to perform more complex tasks at a faster speed with lower cost; Faster connectivity infrastructure and more accessible cloud services for data management and computing power outsourcing. == Categories == Possible applications of industrial AI and machine learning in the production domain can be divided into seven application areas: Market and trend analysis Machinery and equipment Intralogistics Production process Supply chain Building Product Each application area can be further divided into specific application scenarios that describe concrete AI/ML scenarios in production. While some application areas have a direct connection to production processes, others cover production adjacent fields like logistics or the factory building. An example from the application scenario Process Design & Innovation are collaborative robots. Collaborative robotic arms are able to learn the motion and path demonstrated by human operators and perform the same task. Predictive and preventive maintenance through data-driven machine learning are application scenarios from the Machinery & Equipment application area. == Challenges == In contrast to entirely virtual systems, in which ML applications are already widespread today, real-world production processes are characterized by the interaction between the virtual and the physical world. Data is recorded using sensors and processed on computational entities and, if desired, actions and decisions are translated back into the physical world via actuators or by human operators. This poses major challenges for the application of ML in production engineering systems. These challenges are attributable to the encounter of process, data and model characteristics: The production domain's high reliability requirements, high risk and loss potential, the multitude of heterogeneous data sources and the non-transparency of ML model functionality impede a faster adoption of ML in real-world production processes. In particular, production data comprises a variety of different modalities, semantics and quality. Furthermore, production systems are dynamic, uncertain and complex, and engineering and manufacturing problems are data-rich but information-sparse. Besides that, due to the variety of use cases and data characteristics, problem-specific data sets are required, which are difficult to acquire, hindering both practitioners and academic researchers in this domain. === Process and industry characteristics === The domain of production engineering can be considered as a rather conservative industry when it comes to the adoption of advanced technology and their integration into existing processes. This is due to high demands on reliability of the production systems resulting from the potentially high economic harm of reduced process effectiveness due to e.g., additional unplanned downtime or insufficient product qualities. In addition, the specifics of machining equipment and products prevent area-wide adoptions across a variety of processes. Besides the technical reasons, the reluctant adoption of ML is fueled by a lack of IT and data science expertise across the domain. === Data characteristics === The data collected in production processes mainly stem from frequently sampling sensors to estimate the state of a product, a process, or the environment in the real world. Sensor readings are susceptible to noise and represent only an estimate of the reality under uncertainty. Production data typically comprises multiple distributed data sources resulting in various data modalities (e.g., images from visual quality control systems, time-series sensor readings, or cross-sectional job and product information). The inconsistencies in data acquisition lead to low signal-to-noise ratios, low data quality and great effort in data integration, cleaning and management. In addition, as a result from mechanical and chemical wear of production equipment, process data is subject to various forms of data drifts. === Machine learning model characteristics === ML models are considered as black-box systems given their complexity and intransparency of input-output relation. This reduces the comprehensibility of the system behavior and thus also the acceptance by plant operators. Due to the lack of transparency and the stochasticity of these models, no deterministic proof of functional correctness can be achieved, complicating the certification of production equipment. Given their inherent unrestricted prediction behavior, ML models are vulnerable against erroneous or manipulated data, further risking the reliability of the production system because of lacking robustness and safety. In addition to high development and deployment costs, the data drifts cause high maintenance costs, which is disadvantageous compared to purely deterministic programs. == Standard processes for data science in production == The development of ML applications – starting with the identification and selection of the use case and ending with the deployment and maintenance of the application – follows dedicated phases that can be organized in standard process models. The process models assist in structuring the development process and defining requirements that must be met in each phase to enter the next phase. The standard processes can be classified into generic and domain-specific ones. Generic standard processes (e.g., CRISP-DM, ASUM-DM, or knowledge discovery in databases (KDD)) describe a generally valid methodology and are thus independent of individual domains. Domain-specific processes on the other hand consider specific peculiarities and challenges of special application areas. The Machine Learning Pipeline in Production is a domain-specific data science methodology that is inspired by the CRISP-DM model and was specifically designed to be applied in fields of engineering and production technology. To address the core challenges of ML in engineering – process, data, and model characteristics – the methodology especially focuses on use-case assessment, achieving a common data and process understanding data integration, data preprocessing of real-world production data and the deployment and certification of real-world ML applications. == Industrial data sources == The foundation of most artificial intelligence and machine learning applications in industrial settings are comprehensive datasets from the respective fields. Those datasets act as the basis for training the employed models. In other domains, like computer vision, speech recognition or language models, extensive reference datasets (e.g. ImageNet, Librispeech, The People's Speech) and data scraped from the open internet are frequently used for this purpose. Such datasets rarely exist in the industrial context because of high confidentiality requirements and high specificity of the data. Industrial applications of artificial intelligence are therefore often faced with the problem of data availability. For these reasons, existing open datasets applicable to industrial applications, often originate from public institutions like governmental agencies or universities and data analysis competitions hosted by companies. In addition to this, data sharing platforms exist. However, most of these platforms have no industrial focus and offer limited filtering abilities regarding industrial data sources.

  Read more →
• 

  Randomized rounding

  In computer science and operations research, randomized rounding is a widely used approach for designing and analyzing approximation algorithms. Many combinatorial optimization problems are computationally intractable to solve exactly (to optimality). For such problems, randomized rounding can be used to design fast (polynomial time) approximation algorithms—that is, algorithms that are guaranteed to return an approximately optimal solution given any input. The basic idea of randomized rounding is to convert an optimal solution of a relaxation of the problem into an approximately-optimal solution to the original problem. The resulting algorithm is usually analyzed using the probabilistic method. == Overview == The basic approach has three steps: Formulate the problem to be solved as an integer linear program (ILP). Compute an optimal fractional solution x {\displaystyle x} to the linear programming relaxation (LP) of the ILP. Round the fractional solution x {\displaystyle x} of the LP to an integer solution x ′ {\displaystyle x'} of the ILP. (Although the approach is most commonly applied with linear programs, other kinds of relaxations are sometimes used. For example, see Goemans' and Williamson's semidefinite programming-based Max-Cut approximation algorithm.) In the first step, the challenge is to choose a suitable integer linear program. Familiarity with linear programming, in particular modelling using linear programs and integer linear programs, is required. For many problems, there is a natural integer linear program that works well, such as in the Set Cover example below. (The integer linear program should have a small integrality gap; indeed randomized rounding is often used to prove bounds on integrality gaps.) In the second step, the optimal fractional solution can typically be computed in polynomial time using any standard linear programming algorithm. In the third step, the fractional solution must be converted into an integer solution (and thus a solution to the original problem). This is called rounding the fractional solution. The resulting integer solution should (provably) have cost not much larger than the cost of the fractional solution. This will ensure that the cost of the integer solution is not much larger than the cost of the optimal integer solution. The main technique used to do the third step (rounding) is to use randomization, and then to use probabilistic arguments to bound the increase in cost due to the rounding (following the probabilistic method from combinatorics). Therein, probabilistic arguments are used to show the existence of discrete structures with desired properties. In this context, one uses such arguments to show the following: Given any fractional solution x {\displaystyle x} of the LP, with positive probability the randomized rounding process produces an integer solution x ′ {\displaystyle x'} that approximates x {\displaystyle x} according to some desired criterion. Finally, to make the third step computationally efficient, one either shows that x ′ {\displaystyle x'} approximates x {\displaystyle x} with high probability (so that the step can remain randomized) or one derandomizes the rounding step, typically using the method of conditional probabilities. The latter method converts the randomized rounding process into an efficient deterministic process that is guaranteed to reach a good outcome. == Example: the set cover problem == The following example illustrates how randomized rounding can be used to design an approximation algorithm for the set cover problem. Fix any instance ⟨ c , S ⟩ {\displaystyle \langle c,{\mathcal {S}}\rangle } of set cover over a universe U {\displaystyle {\mathcal {U}}} . === Computing the fractional solution === For step 1, let IP be the standard integer linear program for set cover for this instance. For step 2, let LP be the linear programming relaxation of IP, and compute an optimal solution x ∗ {\displaystyle x^{}} to LP using any standard linear programming algorithm. This takes time polynomial in the input size. The feasible solutions to LP are the vectors x {\displaystyle x} that assign each set s ∈ S {\displaystyle s\in {\mathcal {S}}} a non-negative weight x s {\displaystyle x_{s}} , such that, for each element e ∈ U {\displaystyle e\in {\mathcal {U}}} , x ′ {\displaystyle x'} covers e {\displaystyle e} —the total weight assigned to the sets containing e {\displaystyle e} is at least 1, that is, ∑ s ∋ e x s ≥ 1. {\displaystyle \sum _{s\ni e}x_{s}\geq 1.} The optimal solution x ∗ {\displaystyle x^{}} is a feasible solution whose cost ∑ s ∈ S c ( S ) x s ∗ {\displaystyle \sum _{s\in {\mathcal {S}}}c(S)x_{s}^{}} is as small as possible. Note that any set cover C {\displaystyle {\mathcal {C}}} for S {\displaystyle {\mathcal {S}}} gives a feasible solution x {\displaystyle x} (where x s = 1 {\displaystyle x_{s}=1} for s ∈ C {\displaystyle s\in {\mathcal {C}}} , x s = 0 {\displaystyle x_{s}=0} otherwise). The cost of this C {\displaystyle {\mathcal {C}}} equals the cost of x {\displaystyle x} , that is, ∑ s ∈ C c ( s ) = ∑ s ∈ S c ( s ) x s . {\displaystyle \sum _{s\in {\mathcal {C}}}c(s)=\sum _{s\in {\mathcal {S}}}c(s)x_{s}.} In other words, the linear program LP is a relaxation of the given set-cover problem. Since x ∗ {\displaystyle x^{}} has minimum cost among feasible solutions to the LP, the cost of x ∗ {\displaystyle x^{}} is a lower bound on the cost of the optimal set cover. === Randomized rounding step === In step 3, we must convert the minimum-cost fractional set cover x ∗ {\displaystyle x^{}} into a feasible integer solution x ′ {\displaystyle x'} (corresponding to a true set cover). The rounding step should produce an x ′ {\displaystyle x'} that, with positive probability, has cost within a small factor of the cost of x ∗ {\displaystyle x^{}} .Then (since the cost of x ∗ {\displaystyle x^{}} is a lower bound on the cost of the optimal set cover), the cost of x ′ {\displaystyle x'} will be within a small factor of the optimal cost. As a starting point, consider the most natural rounding scheme: For each set s ∈ S {\displaystyle s\in {\mathcal {S}}} in turn, take x s ′ = 1 {\displaystyle x'_{s}=1} with probability min ( 1 , x s ∗ ) {\displaystyle \min(1,x_{s}^{})} , otherwise take x s ′ = 0 {\displaystyle x'_{s}=0} . With this rounding scheme, the expected cost of the chosen sets is at most ∑ s c ( s ) x s ∗ {\displaystyle \sum _{s}c(s)x_{s}^{}} , the cost of the fractional cover. This is good. Unfortunately the coverage is not good. When the variables x s ∗ {\displaystyle x_{s}^{}} are small, the probability that an element e {\displaystyle e} is not covered is about ∏ s ∋ e 1 − x s ∗ ≈ ∏ s ∋ e exp ⁡ ( − x s ∗ ) = exp ⁡ ( − ∑ s ∋ e x s ∗ ) ≈ exp ⁡ ( − 1 ) . {\displaystyle \prod _{s\ni e}1-x_{s}^{}\approx \prod _{s\ni e}\exp(-x_{s}^{})=\exp {\Big (}-\sum _{s\ni e}x_{s}^{}{\Big )}\approx \exp(-1).} So only a constant fraction of the elements will be covered in expectation. To make x ′ {\displaystyle x'} cover every element with high probability, the standard rounding scheme first scales up the rounding probabilities by an appropriate factor λ > 1 {\displaystyle \lambda >1} . Here is the standard rounding scheme: Fix a parameter λ ≥ 1 {\displaystyle \lambda \geq 1} . For each set s ∈ S {\displaystyle s\in {\mathcal {S}}} in turn, take x s ′ = 1 {\displaystyle x'_{s}=1} with probability min ( λ x s ∗ , 1 ) {\displaystyle \min(\lambda x_{s}^{},1)} , otherwise take x s ′ = 0 {\displaystyle x'_{s}=0} . Scaling the probabilities up by λ {\displaystyle \lambda } increases the expected cost by λ {\displaystyle \lambda } , but makes coverage of all elements likely. The idea is to choose λ {\displaystyle \lambda } as small as possible so that all elements are provably covered with non-zero probability. Here is a detailed analysis. ==== Lemma (approximation guarantee for rounding scheme) ==== Fix λ = ln ⁡ ( 2 | U | ) {\displaystyle \lambda =\ln(2|{\mathcal {U}}|)} . With positive probability, the rounding scheme returns a set cover x ′ {\displaystyle x'} of cost at most 2 ln ⁡ ( 2 | U | ) c ⋅ x ∗ {\displaystyle 2\ln(2|{\mathcal {U}}|)c\cdot x^{}} (and thus of cost O ( log ⁡ | U | ) {\displaystyle O(\log |{\mathcal {U}}|)} times the cost of the optimal set cover). (Note: with care the O ( log ⁡ | U | ) {\displaystyle O(\log |{\mathcal {U}}|)} can be reduced to ln ⁡ ( | U | ) + O ( log ⁡ log ⁡ | U | ) {\displaystyle \ln(|{\mathcal {U}}|)+O(\log \log |{\mathcal {U}}|)} .) ==== Proof ==== The output x ′ {\displaystyle x'} of the random rounding scheme has the desired properties as long as none of the following "bad" events occur: the cost c ⋅ x ′ {\displaystyle c\cdot x'} of x ′ {\displaystyle x'} exceeds 2 λ c ⋅ x ∗ {\displaystyle 2\lambda c\cdot x^{}} , or for some element e {\displaystyle e} , x ′ {\displaystyle x'} fails to cover e {\displaystyle e} . The expectation of each x s ′ {\displaystyle x'_{s}} is at most λ x s ∗ {\displaystyle \lambda x_{s

  Read more →
• 

  AI-assisted reverse engineering

  AI-assisted reverse engineering (AIARE) is a branch of computer science that leverages artificial intelligence (AI), notably machine learning (ML) strategies, to augment and automate the process of reverse engineering. The latter involves breaking down a product, system, or process to comprehend its structure, design, and functionality. AIARE was primarily introduced in the early years of the 21st century, witnessing substantial advancements from the mid-2010s onwards. == Overview == Conventionally, reverse engineering is conducted by specialists who dismantle a system to grasp its working principles, often for the purposes of reproduction, modification, enhancement of compatibility, or forensic examination. This method, while efficient, can be laborious and time-intensive, particularly when dealing with intricate software or hardware systems. AIARE integrates machine learning algorithms to either partially automate or augment this process. It is capable of detecting patterns, relationships, structures, and potential vulnerabilities within the analyzed system, frequently surpassing human experts in speed and accuracy. This has rendered AIARE a critical tool in numerous fields, including cybersecurity, software development, and hardware design and analysis. == Techniques == AIARE encompasses several AI methodologies: === Supervised learning === Supervised learning employs tagged data to train models to recognize system components, their operations, and their interconnections. This method is particularly helpful in software analysis to discover vulnerabilities or enhance compatibility. === Unsupervised learning === Unsupervised learning is utilized to detect concealed patterns and structures in untagged data. It proves beneficial in comprehending complex systems where there's no evident labeling or mapping of components. === Reinforcement learning === Reinforcement learning is employed to build models that progressively refine their system understanding through a process of trial and error. This method is often implemented when deciphering a system's functionality under various circumstances or configurations. === Deep learning === Deep learning is employed for analysis of high-dimensional data. For instance, deep learning techniques can aid in examining the layout and connections of integrated circuits (ICs), substantially reducing the manual effort required for reverse engineering. == Benefits == === Usable Security === AIARE expands usable security as reverse engineering is traditionally slow and highly specialized as it produces dense, low-level information (usually in Assembly or C) when using tools like Ghidra. The use of multiple different methods to interface with models today (such as through chat bots like ChatGPT) greatly reduces the barrier to entry by providing a clear way to interact with the user and even providing meaningful decompiled source code. In addition, either done automatically or through prompt engineering, a model is capable of producing a high-level summary and explanation of its reverse engineering efforts in human-readable form that doesn't require much knowledge on code. === Speedup === AIARE is capable of processing data much faster than humans, providing a boost in speed when analyzing said data. In the context of computer security, this can greatly speed up incident management or response and malware detection as AIARE can be automated to drastically reduce the manual effort usually associated with reverse engineering. == Limitations == In an effort to improve readability for reverse engineering, AI-generated code may introduce erroneous bugs not present in the source. This compromises the correctness of the code if not carefully validated and will throw off reverse engineering efforts. Additionally, AIARE's weakness in zero-shot prompting makes gathering accurate data without reference data in the prompt more inconsistent, thus requiring a user to provide some quality data of their own that hurts its usability.

  Read more →
• 

  Pixelmator

  Pixelmator is a series of graphics editors developed by Apple for macOS, iOS, and iPadOS. Pixelmator apps leverage Apple-specific technologies such as CoreML and Metal. Pixelmator uses a proprietary format across their apps (.PXD), but supports editing a variety of file types including Photoshop, RAW, and WebP. == History == Pixelmator Team was founded in 2007 by Lithuanian brothers Saulius and Aidas Dailidė, and released Pixelmator (now Pixelmator Classic) 1.0 in September of the same year. The company resided in Vilnius, Lithuania. In November 2024, Pixelmator Team agreed to be acquired by Apple for an unknown monetary amount, which was completed on 11 February 2025, the company was later folded into Apple with its products coming under them fully. == Pixelmator Classic == Pixelmator Classic was the original version of Pixelmator released for Mac on 25 September 2007. It uses a palette-style interface with floating toolbars compared to Pixelmator Pro's single-window interface. It is no longer being updated and has been delisted from the Mac App Store. == Pixelmator iOS == Pixelmator for iOS launched on 23 October 2014 as an iPad-exclusive app with touch-optimized versions of Pixelmator's desktop features. In May 2015, Pixelmator for iOS 2.0 was released with support for the iPhone. Apple no longer updates Pixelmator for iOS as of 13 January 2026, shortly before the release of Pixelmator Pro for iPad. == Pixelmator Pro == Pixelmator Pro is an image, video, and vector editing software for macOS that launched on 29 November 2017. It was a paid upgrade for Pixelmator Classic users, featuring a redesigned interface, a graphics pipeline rewritten using Metal, Apple silicon support and a greater focus on ML/AI editing features. On 28 January 2026, Apple announced Apple Creator Studio, a subscription bundle for their professional software that contains Pixelmator Pro. They also brought Pixelmator Pro to iPad, shortly after discontinuing Pixelmator iOS. == Photomator == Photomator (formerly Pixelmator Photo) is a photo-oriented editing app which launched on iPad in 2019, on iOS in 2021, and macOS in 2022. After launching the macOS version, the app moved from a one-time purchase to a subscription; however, a lifetime license can still be purchased for $99. Photomator differentiates itself from other Pixelmator apps with features such as batch editing of full photoshoots and AI-powered color correction. Edits in Photomator are made on a single layer and are non-destructive.

  Read more →
• 

  Pointer jumping

  Pointer jumping or path doubling is a design technique for parallel algorithms that operate on pointer structures, such as linked lists and directed graphs. Pointer jumping allows an algorithm to follow paths with a time complexity that is logarithmic with respect to the length of the longest path. It does this by "jumping" to the end of the path computed by neighbors. The basic operation of pointer jumping is to replace each neighbor in a pointer structure with its neighbor's neighbor. In each step of the algorithm, this replacement is done for all nodes in the data structure, which can be done independently in parallel. In the next step when a neighbor's neighbor is followed, the neighbor's path already followed in the previous step is added to the node's followed path in a single step. Thus, each step effectively doubles the distance traversed by the explored paths. Pointer jumping is best understood by looking at simple examples such as list ranking and root finding. == List ranking == One of the simpler tasks that can be solved by a pointer jumping algorithm is the list ranking problem. This problem is defined as follows: given a linked list of N nodes, find the distance (measured in the number of nodes) of each node to the end of the list. The distance d(n) is defined as follows, for nodes n that point to their successor by a pointer called next: If n.next is nil, then d(n) = 0. For any other node, d(n) = d(n.next) + 1. This problem can easily be solved in linear time on a sequential machine, but a parallel algorithm can do better: given n processors, the problem can be solved in logarithmic time, O(log N), by the following pointer jumping algorithm: The pointer jumping occurs in the last line of the algorithm, where each node's next pointer is reset to skip the node's direct successor. It is assumed, as in common in the PRAM model of computation, that memory access are performed in lock-step, so that each n.next.next memory fetch is performed before each n.next memory store; otherwise, processors may clobber each other's data, producing inconsistencies. The following diagram follows how the parallel list ranking algorithm uses pointer jumping for a linked list with 11 elements. As the algorithm describes, the first iteration starts initialized with all ranks set to 1 except those with a null pointer for next. The first iteration looks at immediate neighbors. Each subsequent iteration jumps twice as far as the previous. Analyzing the algorithm yields a logarithmic running time. The initialization loop takes constant time, because each of the N processors performs a constant amount of work, all in parallel. The inner loop of the main loop also takes constant time, as does (by assumption) the termination check for the loop, so the running time is determined by how often this inner loop is executed. Since the pointer jumping in each iteration splits the list into two parts, one consisting of the "odd" elements and one of the "even" elements, the length of the list pointed to by each processor's n is halved in each iteration, which can be done at most O(log N) time before each list has a length of at most one. == Root finding == Following a path in a graph is an inherently serial operation, but pointer jumping reduces the total amount of work by following all paths simultaneously and sharing results among dependent operations. Pointer jumping iterates and finds a successor — a vertex closer to the tree root — each time. By following successors computed for other vertices, the traversal down each path can be doubled every iteration, which means that the tree roots can be found in logarithmic time. Pointer doubling operates on an array successor with an entry for every vertex in the graph. Each successor[i] is initialized with the parent index of vertex i if that vertex is not a root or to i itself if that vertex is a root. At each iteration, each successor is updated to its successor's successor. The root is found when the successor's successor points to itself. The following pseudocode demonstrates the algorithm. algorithm Input: An array parent representing a forest of trees. parent[i] is the parent of vertex i or itself for a root Output: An array containing the root ancestor for every vertex for i ← 1 to length(parent) do in parallel successor[i] ← parent[i] while true for i ← 1 to length(successor) do in parallel successor_next[i] ← successor[successor[i]] if successor_next = successor then break for i ← 1 to length(successor) do in parallel successor[i] ← successor_next[i] return successor The following image provides an example of using pointer jumping on a small forest. On each iteration the successor points to the vertex following one more successor. After two iterations, every vertex points to its root node. == History and examples == Although the name pointer jumping would come later, JáJá attributes the first uses of the technique in early parallel graph algorithms and list ranking. The technique has been described with other names such as shortcutting, but by the 1990s textbooks on parallel algorithms consistently used the term pointer jumping. Today, pointer jumping is considered a software design pattern for operating on recursive data types in parallel. As a technique for following linked paths, graph algorithms are a natural fit for pointer jumping. Consequently, several parallel graph algorithms utilizing pointer jumping have been designed. These include algorithms for finding the roots of a forest of rooted trees, connected components, minimum spanning trees, and biconnected components. However, pointer jumping has also shown to be useful in a variety of other problems including computer vision, image compression, and Bayesian inference.

  Read more →
• 

  Principles for a Data Economy

  The Principles for a Data Economy – Data Rights and Transactions is a transatlantic legal project carried out jointly by the American Law Institute (ALI) and the European Law Institute (ELI). The Principles for a Data Economy deals with a range of different legal questions that arise in the data economy. Since data is different from other tradeable items, the Principles draw up legal rules for data transactions and data rights that take into account the interests of different stakeholders involved in the data economy. The Principles are designed to facilitate contractual relations as well as the drafting of model agreements and can guide courts and legislators worldwide. The project proposes a set of principles that can be implemented in any legal system and is designed to work in conjunction with any kind of data privacy/data protection law, intellectual property law or trade secret law. The Principles do not address or seek to change any of the substantive rules of these bodies of law. The Project Team consists of Neil B Cohen and Christiane Wendehorst (as Project Reporters) and Lord John Thomas as well as Steven O. Weise (as Project Chairs). == Characteristics of data == The law governing trades in commerce has historically focused on trade in items that are tangible like goods or on intangible assets, such as shares or licenses. However, data does not fit into any of these traditional categories, nor does it qualify as a service. It is often unclear how traditional legal rules and doctrines can apply to data, as data is different from other assets in many ways. For example, data can be multiplied at basically no cost and can be used in parallel for a variety of different purposes by many different people at the same time (data is a “non-rivalrous” resource). Uncertainty regarding the applicable rules to govern the data economy may inhibit innovation and growth and trouble stakeholders like data-driven industries, start-ups, and consumers. == Stakeholders in the data economy == The Principles have taken the basic types of players and relations which can be found in data ecosystems as a starting point to provide guidance in different situations. The central actors in the data economy are data controllers (also called “data holders”). They are in a position to access the data and decide for which purposes and means this data should be processed. A controller may exercise control all by itself or share it with co-controllers, such as under a data pooling arrangement. Data processors provide the processing of data on a controller’s behalf as a service. Another important group of stakeholders includes those that contribute to the generation of data (e.g. data subjects). Other players in the data economy include data assemblers or data intermediaries (e.g. data trusts). == History of the project and timeline == Before the official adoption of the project by ALI and ELI bodies in 2018, the project team carried out a Feasibility Study from October 2016 to February 2018. In the following years, the project team produced a number of drafts (e.g. “Preliminary Drafts” No. 1 to 4, “Tentative Draft No. 1”) and project progress were regularly discussed with advisory bodies and members of both the ALI and the ELI. The project reporters also included feedback and insights from industry stakeholders and experts that was gained after several meetings and workshops, hosted, inter alia by UNCITRAL, UNIDROIT and several national governmental institutions. Tentative Draft No. 2 was presented at the ALI Annual Meeting in May 2021 and approved by ALI membership. The latest draft ("Final Council Draft") was also approved by the ELI Council and ELI Membership. The Principles for a Data Economy were presented at an international conference with representatives from institutions such as the Uniform Law Commission (ULC), the European Commission, UNIDROIT, the OECD, the International Chamber of Commerce (ICC) and the World Economic Forum (WEF) in October 2021. == Project structure == The current draft (“Tentative Draft No. 2”) of the Principles consists of five Parts that each governs different aspects of the data economy: General Provisions, Data Contracts, Data Rights, Third Party Aspects of Data Activities, and Multi-State Issues. === General Provisions === Part I includes general provisions that apply to all other Parts of the Principles for a Data Economy. This Part sets out the purpose of the Principles: they aim to make existing law in the field of the data economy more coherent and support the development of the law in this field by courts and legislators worldwide. It is also clarified that the Principles have a wide scope of application and can be used in a variety of ways by stakeholders in the data economy. The Principles may, for example, serve private parties as a basis for contract formation, guide the deliberations of arbitral tribunals or inspire national legislation. Part I then defines several key terms, such as ‘digital data’ and ‘data right’. The scope of the Principles is limited to matters where information is recorded as an asset, resource or tradeable commodity and where large amounts of data, rather than single pieces of information, are concerned. This Part also clarifies that remedies with respect to data contracts and data rights are left to the applicable national law. === Data Contracts === Part II lists different types of contracts that often occur in the data economy and establishes two broad categories, namely contracts for the supply and sharing of data and contracts for services with regard to data. Contracts for the supply and sharing of data include, e.g. data transfer contracts or data pooling arrangements, while contracts for services with regard to data cover contracts for the processing of data or data intermediary contracts. The Principles provide default terms for each contract type, on issues such as the manner in which data should supply or which characteristics the data supplied should meet. These default terms 'automatically' become part of the contract unless the parties agree otherwise. === Data Rights === Part III governs legally protected interests of players in the data economy that stem from the characteristics of data as a resource (e.g. its non-rivalrous nature) or from public interest considerations. Such data rights may include the right to data access, the right to require the controller to desist from data activities or to correct incorrect/incomplete data, or even to receive an economic share in profits derived from the use of data. For example, the Principles deal with data rights of stakeholders that had a share in the co-generation of data and identify different factors to be considered in determining whether to afford a party a data right. The underlying idea that parties who have contributed to the generation of data should have some rights in the utilization of the data is also recognized by governmental institutions, such as by the Japanese Ministry of Economy, Trade and Industry (METI), and the term co-generated data, which was coined by the Principles for a Data Economy, has been adopted, inter alia by the European Commission, the German Data Ethics Commission and the Global Partnership on Artificial Intelligence (GPAI). This Part also deals with data rights for the public interest, such as data sharing rights in the field of innovation. === Third Party Aspects === Part IV governs different situations in which data transactions interfere with the rights of third parties. Such rights include intellectual property rights or rights derived from data privacy or data protection law. This Part sets out under which circumstances data activities should be considered wrongful vis à vis another party. For example, a data activity (like data processing or the onward supply of data) could be considered wrongful, if a controller interferes with the rights of data subjects that are protected by data-protection law. A data activity could also be wrongful if the controller is non-compliant with contractual limitations on data activities, enforceable by the protected party (e.g. a controller may only process data for a certain purpose). If someone obtained access to data by unauthorized means (i.e. data “theft”) this could also be considered wrongful. The Part on Third-Party Aspects also takes a detailed look at the effects of the onward supply of data can have on third parties, while balancing the protection of third parties on the one hand, with the interests of data recipients and the desire to encourage data sharing on the other. === Multi-State Issues === As transactions in the data economy are international by nature and hardly occur within one legal system alone, the Part V of the Principles also briefly touches upon the applicability of the rules and doctrines of private international law to such transactions. == Links == Website of the “Principles for a Data Economy – Data Rights and Transaction

  Read more →
• 

  Mathematical knowledge management

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

  Read more →
• 

  Symbolic regression

  Symbolic regression (SR) is a type of regression analysis that searches the space of mathematical expressions to find the model that best fits a given dataset, both in terms of accuracy and simplicity. No particular model is provided as a starting point for symbolic regression. Instead, initial expressions are formed by randomly combining mathematical building blocks such as mathematical operators, analytic functions, constants, and state variables. Usually, a subset of these primitives will be specified by the person operating it, but that's not a requirement of the technique. The symbolic regression problem for mathematical functions has been tackled with a variety of methods, including recombining equations most commonly using genetic programming, as well as more recent methods utilizing Bayesian methods and neural networks. Another non-classical alternative method to SR is called Universal Functions Originator (UFO), which has a different mechanism, search-space, and building strategy. Further methods such as Exact Learning attempt to transform the fitting problem into a moments problem in a natural function space, usually built around generalizations of the Meijer-G function. By not requiring a priori specification of a model, symbolic regression isn't affected by human bias, or unknown gaps in domain knowledge. It attempts to uncover the intrinsic relationships of the dataset, by letting the patterns in the data itself reveal the appropriate models, rather than imposing a model structure that is deemed mathematically tractable from a human perspective. The fitness function that drives the evolution of the models takes into account not only error metrics (to ensure the models accurately predict the data), but also special complexity measures, thus ensuring that the resulting models reveal the data's underlying structure in a way that's understandable from a human perspective. This facilitates reasoning and favors the odds of getting insights about the data-generating system, as well as improving generalisability and extrapolation behaviour by preventing overfitting. Accuracy and simplicity may be left as two separate objectives of the regression—in which case the optimum solutions form a Pareto front—or they may be combined into a single objective by means of a model selection principle such as minimum description length. It has been proven that symbolic regression is an NP-hard problem. Nevertheless, if the sought-for equation is not too complex it is possible to solve the symbolic regression problem exactly by generating every possible function (built from some predefined set of operators) and evaluating them on the dataset in question. == Difference from classical regression == While conventional regression techniques seek to optimize the parameters for a pre-specified model structure, symbolic regression avoids imposing prior assumptions, and instead infers the model from the data. In other words, it attempts to discover both model structures and model parameters. This approach has the disadvantage of having a much larger space to search, because not only the search space in symbolic regression is infinite, but there are an infinite number of models which will perfectly fit a finite data set (provided that the model complexity isn't artificially limited). This means that it will possibly take a symbolic regression algorithm longer to find an appropriate model and parametrization, than traditional regression techniques. This can be attenuated by limiting the set of building blocks provided to the algorithm, based on existing knowledge of the system that produced the data; but in the end, using symbolic regression is a decision that has to be balanced with how much is known about the underlying system. Nevertheless, this characteristic of symbolic regression also has advantages: because the evolutionary algorithm requires diversity in order to effectively explore the search space, the result is likely to be a selection of high-scoring models (and their corresponding set of parameters). Examining this collection could provide better insight into the underlying process, and allows the user to identify an approximation that better fits their needs in terms of accuracy and simplicity. == Benchmarking == === SRBench === In 2021, SRBench was proposed as a large benchmark for symbolic regression. In its inception, SRBench featured 14 symbolic regression methods, 7 other ML methods, and 252 datasets from PMLB. The benchmark intends to be a living project: it encourages the submission of improvements, new datasets, and new methods, to keep track of the state of the art in SR. === SRBench Competition 2022 === In 2022, SRBench announced the competition Interpretable Symbolic Regression for Data Science, which was held at the GECCO conference in Boston, MA. The competition pitted nine leading symbolic regression algorithms against each other on a novel set of data problems and considered different evaluation criteria. The competition was organized in two tracks, a synthetic track and a real-world data track. ==== Synthetic Track ==== In the synthetic track, methods were compared according to five properties: re-discovery of exact expressions; feature selection; resistance to local optima; extrapolation; and sensitivity to noise. Rankings of the methods were: QLattice PySR (Python Symbolic Regression) uDSR (Deep Symbolic Optimization) ==== Real-world Track ==== In the real-world track, methods were trained to build interpretable predictive models for 14-day forecast counts of COVID-19 cases, hospitalizations, and deaths in New York State. These models were reviewed by a subject expert and assigned trust ratings and evaluated for accuracy and simplicity. The ranking of the methods was: uDSR (Deep Symbolic Optimization) QLattice geneticengine (Genetic Engine) == Non-standard methods == Most symbolic regression algorithms prevent combinatorial explosion by implementing evolutionary algorithms that iteratively improve the best-fit expression over many generations. Recently, researchers have proposed algorithms utilizing other tactics in AI. Silviu-Marian Udrescu and Max Tegmark developed the "AI Feynman" algorithm, which attempts symbolic regression by training a neural network to represent the mystery function, then runs tests against the neural network to attempt to break up the problem into smaller parts. For example, if f ( x 1 , . . . , x i , x i + 1 , . . . , x n ) = g ( x 1 , . . . , x i ) + h ( x i + 1 , . . . , x n ) {\displaystyle f(x_{1},...,x_{i},x_{i+1},...,x_{n})=g(x_{1},...,x_{i})+h(x_{i+1},...,x_{n})} , tests against the neural network can recognize the separation and proceed to solve for g {\displaystyle g} and h {\displaystyle h} separately and with different variables as inputs. This is an example of divide and conquer, which reduces the size of the problem to be more manageable. AI Feynman also transforms the inputs and outputs of the mystery function in order to produce a new function which can be solved with other techniques, and performs dimensional analysis to reduce the number of independent variables involved. The algorithm was able to "discover" 100 equations from The Feynman Lectures on Physics, while a leading software using evolutionary algorithms, Eureqa, solved only 71. AI Feynman, in contrast to classic symbolic regression methods, requires a very large dataset in order to first train the neural network and is naturally biased towards equations that are common in elementary physics.

  Read more →
• 

  Adaptive algorithm

  An adaptive algorithm is an algorithm that changes its behavior at the time it is run, based on information available and on a priori defined reward mechanism (or criterion). Such information could be the story of recently received data, information on the available computational resources, or other run-time acquired (or a priori known) information related to the environment in which it operates. Among the most used adaptive algorithms is the Widrow-Hoff’s least mean squares (LMS), which represents a class of stochastic gradient-descent algorithms used in adaptive filtering and machine learning. In adaptive filtering the LMS is used to mimic a desired filter by finding the filter coefficients that relate to producing the least mean square of the error signal (difference between the desired and the actual signal). For example, stable partition, using no additional memory is O(n lg n) but given O(n) memory, it can be O(n) in time. As implemented by the C++ Standard Library, stable_partition is adaptive and so it acquires as much memory as it can get (up to what it would need at most) and applies the algorithm using that available memory. Another example is adaptive sort, whose behavior changes upon the presortedness of its input. An example of an adaptive algorithm in radar systems is the constant false alarm rate (CFAR) detector. In machine learning and optimization, many algorithms are adaptive or have adaptive variants, which usually means that the algorithm parameters such as learning rate are automatically adjusted according to statistics about the optimisation thus far (e.g. the rate of convergence). Examples include adaptive simulated annealing, adaptive coordinate descent, adaptive quadrature, AdaBoost, Adagrad, Adadelta, RMSprop, and Adam. In data compression, adaptive coding algorithms such as Adaptive Huffman coding or Prediction by partial matching can take a stream of data as input, and adapt their compression technique based on the symbols that they have already encountered. In signal processing, the Adaptive Transform Acoustic Coding (ATRAC) codec used in MiniDisc recorders is called "adaptive" because the window length (the size of an audio "chunk") can change according to the nature of the sound being compressed, to try to achieve the best-sounding compression strategy.

  Read more →
• 

  Run-time algorithm specialization

  In computer science, run-time algorithm specialization is a methodology for creating efficient algorithms for costly computation tasks of certain kinds. The methodology originates in the field of automated theorem proving and, more specifically, in the Vampire theorem prover project. The idea is inspired by the use of partial evaluation in optimising program translation. Many core operations in theorem provers exhibit the following pattern. Suppose that we need to execute some algorithm a l g ( A , B ) {\displaystyle {\mathit {alg}}(A,B)} in a situation where a value of A {\displaystyle A} is fixed for potentially many different values of B {\displaystyle B} . In order to do this efficiently, we can try to find a specialization of a l g {\displaystyle {\mathit {alg}}} for every fixed A {\displaystyle A} , i.e., such an algorithm a l g A {\displaystyle {\mathit {alg}}_{A}} , that executing a l g A ( B ) {\displaystyle {\mathit {alg}}_{A}(B)} is equivalent to executing a l g ( A , B ) {\displaystyle {\mathit {alg}}(A,B)} . The specialized algorithm may be more efficient than the generic one, since it can exploit some particular properties of the fixed value A {\displaystyle A} . Typically, a l g A ( B ) {\displaystyle {\mathit {alg}}_{A}(B)} can avoid some operations that a l g ( A , B ) {\displaystyle {\mathit {alg}}(A,B)} would have to perform, if they are known to be redundant for this particular parameter A {\displaystyle A} . In particular, we can often identify some tests that are true or false for A {\displaystyle A} , unroll loops and recursion, etc. == Difference from partial evaluation == The key difference between run-time specialization and partial evaluation is that the values of A {\displaystyle A} on which a l g {\displaystyle {\mathit {alg}}} is specialised are not known statically, so the specialization takes place at run-time. There is also an important technical difference. Partial evaluation is applied to algorithms explicitly represented as codes in some programming language. At run-time, we do not need any concrete representation of a l g {\displaystyle {\mathit {alg}}} . We only have to imagine a l g {\displaystyle {\mathit {alg}}} when we program the specialization procedure. All we need is a concrete representation of the specialized version a l g A {\displaystyle {\mathit {alg}}_{A}} . This also means that we cannot use any universal methods for specializing algorithms, which is usually the case with partial evaluation. Instead, we have to program a specialization procedure for every particular algorithm a l g {\displaystyle {\mathit {alg}}} . An important advantage of doing so is that we can use some powerful ad hoc tricks exploiting peculiarities of a l g {\displaystyle {\mathit {alg}}} and the representation of A {\displaystyle A} and B {\displaystyle B} , which are beyond the reach of any universal specialization methods. == Specialization with compilation == The specialized algorithm has to be represented in a form that can be interpreted. In many situations, usually when a l g A ( B ) {\displaystyle {\mathit {alg}}_{A}(B)} is to be computed on many values of B {\displaystyle B} in a row, a l g A {\displaystyle {\mathit {alg}}_{A}} can be written as machine code instructions for a special abstract machine, and it is typically said that A {\displaystyle A} is compiled. The code itself can then be additionally optimized by answer-preserving transformations that rely only on the semantics of instructions of the abstract machine. The instructions of the abstract machine can usually be represented as records. One field of such a record, an instruction identifier (or instruction tag), would identify the instruction type, e.g. an integer field may be used, with particular integer values corresponding to particular instructions. Other fields may be used for storing additional parameters of the instruction, e.g. a pointer field may point to another instruction representing a label, if the semantics of the instruction require a jump. All instructions of the code can be stored in a traversable data structure such as an array, linked list, or tree. Interpretation (or execution) proceeds by fetching instructions in some order, identifying their type, and executing the actions associated with said type. In many programming languages, such as C and C++, a simple switch statement may be used to associate actions with different instruction identifiers. Modern compilers usually compile a switch statement with constant (e.g. integer) labels from a narrow range by storing the address of the statement corresponding to a value i {\displaystyle i} in the i {\displaystyle i} -th cell of a special array, as a means of efficient optimisation. This can be exploited by taking values for instruction identifiers from a small interval of values. == Data-and-algorithm specialization == There are situations when many instances of A {\displaystyle A} are intended for long-term storage and the calls of a l g ( A , B ) {\displaystyle {\mathit {alg}}(A,B)} occur with different B {\displaystyle B} in an unpredictable order. For example, we may have to check a l g ( A 1 , B 1 ) {\displaystyle {\mathit {alg}}(A_{1},B_{1})} first, then a l g ( A 2 , B 2 ) {\displaystyle {\mathit {alg}}(A_{2},B_{2})} , then a l g ( A 1 , B 3 ) {\displaystyle {\mathit {alg}}(A_{1},B_{3})} , and so on. In such circumstances, full-scale specialization with compilation may not be suitable due to excessive memory usage. However, we can sometimes find a compact specialized representation A ′ {\displaystyle A^{\prime }} for every A {\displaystyle A} , that can be stored with, or instead of, A {\displaystyle A} . We also define a variant a l g ′ {\displaystyle {\mathit {alg}}^{\prime }} that works on this representation and any call to a l g ( A , B ) {\displaystyle {\mathit {alg}}(A,B)} is replaced by a l g ′ ( A ′ , B ) {\displaystyle {\mathit {alg}}^{\prime }(A^{\prime },B)} , intended to do the same job faster.

  Read more →