AI Face Fixer

AI Face Fixer — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

• 

  Read the Docs

  Read the Docs is an open-sourced free software documentation hosting platform. It generates documentation written with the Sphinx documentation generator, MkDocs, or Jupyter Book. == History == The site was created in 2010 by Eric Holscher, Bobby Grace, and Charles Leifer. On March 9, 2011, the Python Software Foundation Board awarded a grant of US$840 to the Read the Docs project for one year of hosting fees. On November 13, 2017, the Linux Mint project announced that they were moving their documentation to Read the Docs. In 2020, Read the Docs received a $200,000 grant from the Chan Zuckerberg Initiative. For 2021, Read the Docs reported 700 million page views and 196 million unique visitors. In 2013, a "Write the Docs" conference for Read the Docs users was launched, which has since turned into a generic software-documentation community. As of 2024, it continues to hold annual global conferences, organize local meetups, and maintain a Slack channel for "people who care about documentation."

  Read more →
• 

  HAKMEM

  HAKMEM, alternatively known as AI Memo 239, is a February 1972 "memo" (technical report) of the MIT AI Lab containing a wide variety of hacks, including useful and clever algorithms for mathematical computation, some number theory and schematic diagrams for hardware – in Guy L. Steele's words, "a bizarre and eclectic potpourri of technical trivia". Contributors included about two dozen members and associates of the AI Lab. The title of the report is short for "hacks memo", abbreviated to six upper case characters that would fit in a single PDP-10 machine word (using a six-bit character set). == History == HAKMEM is notable as an early compendium of algorithmic technique, particularly for its practical bent, and as an illustration of the wide-ranging interests of AI Lab people of the time, which included almost anything other than AI research. HAKMEM contains original work in some fields, notably continued fractions. == Introduction == Compiled with the hope that a record of the random things people do around here can save some duplication of effort -- except for fun. Here is some little known data which may be of interest to computer hackers. The items and examples are so sketchy that to decipher them may require more sincerity and curiosity than a non-hacker can muster. Doubtless, little of this is new, but nowadays it's hard to tell. So we must be content to give you an insight, or save you some cycles, and to welcome further contributions of items, new or used.

  Read more →
• 

  Emotion recognition

  Emotion recognition is the process of identifying human emotion. People vary widely in their accuracy at recognizing the emotions of others. Use of technology to help people with emotion recognition is a relatively nascent research area. Generally, the technology works best if it uses multiple modalities in context. To date, the most work has been conducted on automating the recognition of facial expressions from video, spoken expressions from audio, written expressions from text, and physiology as measured by wearables. == Human == Humans show a great deal of variability in their abilities to recognize emotion. A key point to keep in mind when learning about automated emotion recognition is that there are several sources of "ground truth", or truth about what the real emotion is. Suppose we are trying to recognize the emotions of Alex. One source is "what would most people say that Alex is feeling?" In this case, the 'truth' may not correspond to what Alex feels, but may correspond to what most people would say it looks like Alex feels. For example, Alex may actually feel sad, but he puts on a big smile and then most people say he looks happy. If an automated method achieves the same results as a group of observers it may be considered accurate, even if it does not actually measure what Alex truly feels. Another source of 'truth' is to ask Alex what he truly feels. This works if Alex has a good sense of his internal state, and wants to tell you what it is, and is capable of putting it accurately into words or a number. However, some people are alexithymic and do not have a good sense of their internal feelings, or they are not able to communicate them accurately with words and numbers. In general, getting to the truth of what emotion is actually present can take some work, can vary depending on the criteria that are selected, and will usually involve maintaining some level of uncertainty. == Automatic == Decades of scientific research have been conducted developing and evaluating methods for automated emotion recognition. There is now an extensive literature proposing and evaluating hundreds of different kinds of methods, leveraging techniques from multiple areas, such as signal processing, machine learning, computer vision, and speech processing. Different methodologies and techniques may be employed to interpret emotion such as Bayesian networks. , Gaussian Mixture models and Hidden Markov Models and deep neural networks. === Approaches === The accuracy of emotion recognition is usually improved when it combines the analysis of human expressions from multimodal forms such as texts, physiology, audio, or video. Different emotion types are detected through the integration of information from facial expressions, body movement and gestures, and speech. The technology is said to contribute in the emergence of the so-called emotional or emotive Internet. The existing approaches in emotion recognition to classify certain emotion types can be generally classified into three main categories: knowledge-based techniques, statistical methods, and hybrid approaches. ==== Knowledge-based techniques ==== Knowledge-based techniques (sometimes referred to as lexicon-based techniques), utilize domain knowledge and the semantic and syntactic characteristics of text and potentially spoken language in order to detect certain emotion types. In this approach, it is common to use knowledge-based resources during the emotion classification process such as WordNet, SenticNet, ConceptNet, and EmotiNet, to name a few. One of the advantages of this approach is the accessibility and economy brought about by the large availability of such knowledge-based resources. A limitation of this technique on the other hand, is its inability to handle concept nuances and complex linguistic rules. Knowledge-based techniques can be mainly classified into two categories: dictionary-based and corpus-based approaches. Dictionary-based approaches find opinion or emotion seed words in a dictionary and search for their synonyms and antonyms to expand the initial list of opinions or emotions. Corpus-based approaches on the other hand, start with a seed list of opinion or emotion words, and expand the database by finding other words with context-specific characteristics in a large corpus. While corpus-based approaches take into account context, their performance still vary in different domains since a word in one domain can have a different orientation in another domain. ==== Statistical methods ==== Statistical methods commonly involve the use of different supervised machine learning algorithms in which a large set of annotated data is fed into the algorithms for the system to learn and predict the appropriate emotion types. Machine learning algorithms generally provide more reasonable classification accuracy compared to other approaches, but one of the challenges in achieving good results in the classification process, is the need to have a sufficiently large training set. Some of the most commonly used machine learning algorithms include Support Vector Machines (SVM), Naive Bayes, and Maximum Entropy. Deep learning, which is under the unsupervised family of machine learning, is also widely employed in emotion recognition. Well-known deep learning algorithms include different architectures of Artificial Neural Network (ANN) such as Convolutional Neural Network (CNN), Long Short-term Memory (LSTM), and Extreme Learning Machine (ELM). The popularity of deep learning approaches in the domain of emotion recognition may be mainly attributed to its success in related applications such as in computer vision, speech recognition, and Natural Language Processing (NLP). ==== Hybrid approaches ==== Hybrid approaches in emotion recognition are essentially a combination of knowledge-based techniques and statistical methods, which exploit complementary characteristics from both techniques. Some of the works that have applied an ensemble of knowledge-driven linguistic elements and statistical methods include sentic computing and iFeel, both of which have adopted the concept-level knowledge-based resource SenticNet. The role of such knowledge-based resources in the implementation of hybrid approaches is highly important in the emotion classification process. Since hybrid techniques gain from the benefits offered by both knowledge-based and statistical approaches, they tend to have better classification performance as opposed to employing knowledge-based or statistical methods independently. A downside of using hybrid techniques however, is the computational complexity during the classification process. === Datasets === Data is an integral part of the existing approaches in emotion recognition and in most cases it is a challenge to obtain annotated data that is necessary to train machine learning algorithms. For the task of classifying different emotion types from multimodal sources in the form of texts, audio, videos or physiological signals, the following datasets are available: HUMAINE: provides natural clips with emotion words and context labels in multiple modalities Belfast database: provides clips with a wide range of emotions from TV programs and interview recordings SEMAINE: provides audiovisual recordings between a person and a virtual agent and contains emotion annotations such as angry, happy, fear, disgust, sadness, contempt, and amusement IEMOCAP: provides recordings of dyadic sessions between actors and contains emotion annotations such as happiness, anger, sadness, frustration, and neutral state eNTERFACE: provides audiovisual recordings of subjects from seven nationalities and contains emotion annotations such as happiness, anger, sadness, surprise, disgust, and fear DEAP: provides electroencephalography (EEG), electrocardiography (ECG), and face video recordings, as well as emotion annotations in terms of valence, arousal, and dominance of people watching film clips DREAMER: provides electroencephalography (EEG) and electrocardiography (ECG) recordings, as well as emotion annotations in terms of valence, dominance of people watching film clips MELD: is a multiparty conversational dataset where each utterance is labeled with emotion and sentiment. MELD provides conversations in video format and hence suitable for multimodal emotion recognition and sentiment analysis. MELD is useful for multimodal sentiment analysis and emotion recognition, dialogue systems and emotion recognition in conversations. MuSe: provides audiovisual recordings of natural interactions between a person and an object. It has discrete and continuous emotion annotations in terms of valence, arousal and trustworthiness as well as speech topics useful for multimodal sentiment analysis and emotion recognition. UIT-VSMEC: is a standard Vietnamese Social Media Emotion Corpus (UIT-VSMEC) with about 6,927 human-annotated sentences with six emotion labels, contributing to emotion recognition research in Vietnamese

  Read more →
• 

  Flajolet–Martin algorithm

  The Flajolet–Martin algorithm is an algorithm for approximating the number of distinct elements in a stream with a single pass and space-consumption logarithmic in the maximal number of possible distinct elements in the stream (the count-distinct problem). The algorithm was introduced by Philippe Flajolet and G. Nigel Martin in their 1984 article "Probabilistic Counting Algorithms for Data Base Applications". Later it has been refined in "LogLog counting of large cardinalities" by Marianne Durand and Philippe Flajolet, and "HyperLogLog: The analysis of a near-optimal cardinality estimation algorithm" by Philippe Flajolet et al. In their 2010 article "An optimal algorithm for the distinct elements problem", Daniel M. Kane, Jelani Nelson and David P. Woodruff give an improved algorithm, which uses nearly optimal space and has optimal O(1) update and reporting times. == The algorithm == Assume that we are given a hash function h a s h ( x ) {\displaystyle \mathrm {hash} (x)} that maps input x {\displaystyle x} to integers in the range [ 0 ; 2 L − 1 ] {\displaystyle [0;2^{L}-1]} , and where the outputs are sufficiently uniformly distributed. Note that the set of integers from 0 to 2 L − 1 {\displaystyle 2^{L}-1} corresponds to the set of binary strings of length L {\displaystyle L} . For any non-negative integer y {\displaystyle y} , define b i t ( y , k ) {\displaystyle \mathrm {bit} (y,k)} to be the k {\displaystyle k} -th bit in the binary representation of y {\displaystyle y} , such that: y = ∑ k ≥ 0 b i t ( y , k ) 2 k . {\displaystyle y=\sum _{k\geq 0}\mathrm {bit} (y,k)2^{k}.} We then define a function ρ ( y ) {\displaystyle \rho (y)} that outputs the position of the least-significant set bit in the binary representation of y {\displaystyle y} , and L {\displaystyle L} if no such set bit can be found as all bits are zero: ρ ( y ) = { min { k ≥ 0 ∣ b i t ( y , k ) ≠ 0 } y > 0 L y = 0 {\displaystyle \rho (y)={\begin{cases}\min\{k\geq 0\mid \mathrm {bit} (y,k)\neq 0\}&y>0\\L&y=0\end{cases}}} Note that with the above definition we are using 0-indexing for the positions, starting from the least significant bit. For example, ρ ( 13 ) = ρ ( 1101 2 ) = 0 {\displaystyle \rho (13)=\rho (1101_{2})=0} , since the least significant bit is a 1 (0th position), and ρ ( 8 ) = ρ ( 1000 2 ) = 3 {\displaystyle \rho (8)=\rho (1000_{2})=3} , since the least significant set bit is at the 3rd position. At this point, note that under the assumption that the output of our hash function is uniformly distributed, then the probability of observing a hash output ending with 2 k {\displaystyle 2^{k}} (a one, followed by k {\displaystyle k} zeroes) is 2 − ( k + 1 ) {\displaystyle 2^{-(k+1)}} , since this corresponds to flipping k {\displaystyle k} heads and then a tail with a fair coin. Now the Flajolet–Martin algorithm for estimating the cardinality of a multiset M {\displaystyle M} is as follows: Initialize a bit-vector BITMAP to be of length L {\displaystyle L} and contain all 0s. For each element x {\displaystyle x} in M {\displaystyle M} : Calculate the index i = ρ ( h a s h ( x ) ) {\displaystyle i=\rho (\mathrm {hash} (x))} . Set B I T M A P [ i ] = 1 {\displaystyle \mathrm {BITMAP} [i]=1} . Let R {\displaystyle R} denote the smallest index i {\displaystyle i} such that B I T M A P [ i ] = 0 {\displaystyle \mathrm {BITMAP} [i]=0} . Estimate the cardinality of M {\displaystyle M} as 2 R / ϕ {\displaystyle 2^{R}/\phi } , where ϕ ≈ 0.77351 {\displaystyle \phi \approx 0.77351} . The idea is that if n {\displaystyle n} is the number of distinct elements in the multiset M {\displaystyle M} , then B I T M A P [ 0 ] {\displaystyle \mathrm {BITMAP} [0]} is accessed approximately n / 2 {\displaystyle n/2} times, B I T M A P [ 1 ] {\displaystyle \mathrm {BITMAP} [1]} is accessed approximately n / 4 {\displaystyle n/4} times and so on. Consequently, if i ≫ log 2 ⁡ n {\displaystyle i\gg \log _{2}n} , then B I T M A P [ i ] {\displaystyle \mathrm {BITMAP} [i]} is almost certainly 0, and if i ≪ log 2 ⁡ n {\displaystyle i\ll \log _{2}n} , then B I T M A P [ i ] {\displaystyle \mathrm {BITMAP} [i]} is almost certainly 1. If i ≈ log 2 ⁡ n {\displaystyle i\approx \log _{2}n} , then B I T M A P [ i ] {\displaystyle \mathrm {BITMAP} [i]} can be expected to be either 1 or 0. The correction factor ϕ ≈ 0.77351 {\displaystyle \phi \approx 0.77351} (OEIS: A244256) is found by calculations, which can be found in the original article. == Improving accuracy == A problem with the Flajolet–Martin algorithm in the above form is that the results vary significantly. A common solution has been to run the algorithm multiple times with k {\displaystyle k} different hash functions and combine the results from the different runs. One idea is to take the mean of the k {\displaystyle k} results together from each hash function, obtaining a single estimate of the cardinality. The problem with this is that averaging is very susceptible to outliers (which are likely here). A different idea is to use the median, which is less prone to be influences by outliers. The problem with this is that the results can only take form 2 R / ϕ {\displaystyle 2^{R}/\phi } , where R {\displaystyle R} is integer. A common solution is to combine both the mean and the median: Create k ⋅ l {\displaystyle k\cdot l} hash functions and split them into k {\displaystyle k} distinct groups (each of size l {\displaystyle l} ). Within each group use the mean for aggregating together the l {\displaystyle l} results, and finally take the median of the k {\displaystyle k} group estimates as the final estimate. The 2007 HyperLogLog algorithm splits the multiset into subsets and estimates their cardinalities, then it uses the harmonic mean to combine them into an estimate for the original cardinality.

  Read more →
• 

  Apertus (LLM)

  Apertus is a public large language model, developed by the Swiss AI Initiative (a collaboration between EPFL, ETH Zurich, and the Swiss National Supercomputing Centre). It was released on September 2, 2025, under the free and open-source Apache 2.0 license. Designed initially for business and research use cases around the world, Apertus was trained on over 1800 languages, and comes in 8 billion or 70 billion parameter versions and is available on Hugging Face for download. The model was developed aiming to adhere to European copyright law, and is one of the first examples of AI as a public good in the vein of AI Sovereignty. It is also the first large model to comply with the European Union's Artificial Intelligence Act. At its launch, the model creators emphasized multilinguality, transparency, and auditability as priorities in contrast to commercial frontier model. While international reception was largely positive, the first iteration was significantly behind the capabilities of frontier models and needs adaptation for many use cases with chatbots being a secondary but not a primary use case. As of late 2025, it was considered the largest and most capable fully open model. The capability of future models will depend in part on how much more funding can be secured.

  Read more →
• 

  Pointer algorithm

  In computer science, a pointer algorithm (sometimes called a pointer machine, or a reference machine; see the article Pointer machine for a close but non-identical concept) is a type of algorithm that manages a linked data structure. This concept is used as a model for lower-bound proofs and specific restrictions on the linked data structure and on the algorithm's access to the structure vary. This model has been used extensively with problems related to the disjoint-set data structure. Thus, Tarjan and La Poutré used this model to prove lower bounds on the amortized complexity of a disjoint-set data structure (La Poutré also addressed the interval split-find problem). Blum used this model to prove a lower bound on the single operation worst-case time of disjoint set data structure. Blum and Rochow proved a worst-case lower bound for the interval union-find problem. == Example == In Tarjan's lower bound for the disjoint set union problem, the assumptions on the algorithm are: The algorithm maintains a linked structure of nodes. Each element of the problem is associated with a node. Each set is represented by a node. The nodes of each set constitute a distinct connected component in the structure (this property is called separability). The find operation is performed by following links from the element node to the set node. Under these assumptions, the lower bound of Ω ( m α ( m , n ) ) {\displaystyle \Omega (m\alpha (m,n))} on the cost of a sequence of m operations is proven.

  Read more →
• 

  Bisection (software engineering)

  Bisection is a method used in software development to identify change sets that result in a specific behavior change. It is mostly employed for finding the patch that introduced a bug. Another application area is finding the patch that indirectly fixed a bug. == Overview == The process of locating the changeset that introduced a specific regression was described as "source change isolation" in 1997 by Brian Ness and Viet Ngo of Cray Research. Regression testing was performed on Cray's compilers in editions comprising one or more changesets. Editions with known regressions could not be validated until developers addressed the problem. Source change isolation narrowed the cause to a single changeset that could then be excluded from editions, unblocking them with respect to this problem, while the author of the change worked on a fix. Ness and Ngo outlined linear search and binary search methods of performing this isolation. Code bisection has the goal of minimizing the effort to find a specific change set. It employs a divide and conquer algorithm that depends on having access to the code history which is usually preserved by revision control in a code repository. == Bisection method == === Code bisection algorithm === Code history has the structure of a directed acyclic graph which can be topologically sorted. This makes it possible to use a divide and conquer search algorithm which: splits up the search space of candidate revisions tests for the behavior in question reduces the search space depending on the test result re-iterates the steps above until a range with at most one bisectable patch candidate remains === Algorithmic complexity === Bisection is in LSPACE having an algorithmic complexity of O ( log ⁡ N ) {\displaystyle O(\log N)} with N {\displaystyle N} denoting the number of revisions in the search space, and is similar to a binary search. === Desirable repository properties === For code bisection it is desirable that each revision in the search space can be built and tested independently. === Monotonicity === For the bisection algorithm to identify a single changeset which caused the behavior being tested to change, the behavior must change monotonically across the search space. For a Boolean function such as a pass/fail test, this means that it only changes once across all changesets between the start and end of the search space. If there are multiple changesets across the search space where the behavior being tested changes between false and true, then the bisection algorithm will find one of them, but it will not necessarily be the root cause of the change in behavior between the start and the end of the search space. The root cause could be a different changeset, or a combination of two or more changesets across the search space. To help deal with this problem, automated tools allow specific changesets to be ignored during a bisection search. == Automation support == Although the bisection method can be completed manually, one of its main advantages is that it can be easily automated. It can thus fit into existing test automation processes: failures in exhaustive automated regression tests can trigger automated bisection to localize faults. Ness and Ngo focused on its potential in Cray's continuous delivery-style environment in which the automatically isolated bad changeset could be automatically excluded from builds. The revision control systems Fossil, Git and Mercurial have built-in functionality for code bisection. The user can start a bisection session with a specified range of revisions from which the revision control system proposes a revision to test, the user tells the system whether the revision tested as "good" or "bad", and the process repeats until the specific "bad" revision has been identified. Other revision control systems, such as Bazaar or Subversion, support bisection through plugins or external scripts. Phoronix Test Suite can do bisection automatically to find performance regressions.

  Read more →
• 

  Artificial intelligence industry in Taiwan

  The artificial intelligence (AI) industry in Taiwan refers to the development, application, and commercialization of artificial intelligence technologies within Taiwan. The industry has grown alongside Taiwan's established strengths in semiconductor manufacturing and information and communications technology (ICT), and is supported by government policy, research institutions, and private sector participation. AI development in Taiwan has focused on integrating hardware capabilities with software applications across sectors such as manufacturing, healthcare, and smart infrastructure. Artificial intelligence has been identified as a strategic area of development in Taiwan since the late 2010s. While Taiwan has historically played a limited role in early theoretical and expert-system phases of AI development, its position in global electronics manufacturing has provided a foundation for participation in the contemporary era of machine learning and data-driven AI systems. Taiwan's AI industry is characterized by a strong hardware base, particularly in semiconductor production and AI server manufacturing, combined with increasing investment in software, data infrastructure, and applied AI services. The sector has been shaped by global demand for computing power, advances in deep learning, and the expansion of AI applications in industrial and commercial contexts. == Government policy and development == The Taiwanese government has promoted AI development through a series of national strategies. In 2017, the Ministry of Science and Technology launched the "AI Grand Strategy for a Small Country" initiative, investing approximately US$517 million between 2017 and 2021 to support research, infrastructure, and talent development. This initiative aimed to build a domestic AI ecosystem by funding research centers, expanding data infrastructure, and supporting industrial adoption. The Executive Yuan also introduced the AI Taiwan Action Plan 1.0 (2018–2021), which focused on integrating AI technologies into existing industries and strengthening research and development capabilities. A subsequent plan, AI Taiwan Action Plan 2.0 (2023–2026), expanded the focus to include ethical governance, regulatory frameworks, and risk management in response to the growth of generative AI technologies. In 2023, the Taiwan AI Center of Excellence (Taiwan AICoE), a government-backed hub, was established by the National Science and Technology Council to accelerate AI development, foster international collaboration, and train talent in Taiwan. It acts as a specialized think tank focusing on creating a "smart technology island" by integrating AI resources and developing trusted, human-centric AI technologies. In 2024, the Taiwan Chip-based Industrial Innovation Program (CbI) was launched by the Executive Yuan as a 10-year, NT$300 billion (US$9.3 billion) initiative to leverage Taiwan's semiconductor dominance, driving innovation in AI, smart mobility, manufacturing, and healthcare. It aims to combine generative AI with IC technology, cultivate talent, and attract global startups to build a "Silicon Island". In parallel, the Taiwanese government has explored legislative frameworks such as a proposed Artificial Intelligence Fundamental Act in December 2025, addressing issues including data protection, safety standards, and intellectual property. == Industrial structure == === Semiconductor and hardware foundation === Taiwan's AI industry is closely linked to its semiconductor sector. In 2020, Taiwan accounted for approximately 77.3% of the global wafer foundry market and 57.7% of packaging and testing, with a 20.1% share in integrated circuit (IC) design. These capabilities provide critical infrastructure for AI systems, which rely on high-performance computing hardware. Taiwanese firms are also involved in the production of AI servers and related components, contributing significantly to global supply chains for data centers and cloud computing. The integration of chip design, manufacturing, and assembly has enabled Taiwan to play a central role in providing the computational resources required for AI development. On 20 November 2025, Google established the "Google Taiwan AI Infrastructure R&D Center", second only to its US headquarters and largest AI hardware infrastructure engineering center outside of the United States. === Software and services === Compared to its hardware capabilities, Taiwan's AI software sector is less developed. The absence of large-scale global AI platform companies has been noted as a structural limitation. As a result, much of Taiwan's AI industry focuses on applied solutions, including customization of existing AI models for specific industries. Therefore, efforts to strengthen software capabilities have included investment in research institutions, startup ecosystems, and collaborations between academia and industry. == Applications == === Smart manufacturing === AI has been widely applied in Taiwan's manufacturing sector, which is a major component of the economy. Applications include process automation, predictive maintenance, quality control, and fault detection. AI-enabled smart manufacturing systems aim to improve efficiency, reduce production costs, and enhance product quality. Taiwan's manufacturing industry has incorporated AI technologies into production lines, particularly in electronics and machinery sectors. === Healthcare === The use of AI in healthcare in Taiwan has expanded in areas such as medical imaging, diagnostics, and drug development. AI systems are used to analyze CT scans, MRI data, and other clinical information to support diagnosis and treatment planning. Taiwan's healthcare sector, which includes medical devices, pharmaceuticals, and medical services, has benefited from the integration of AI technologies, particularly in precision medicine and clinical decision support systems. A notable example of AI healthcare deployment in Taiwan is the collaboration between Siemens Healthineers, Ever Fortune AI, and Asia University Hospital. === Edge computing and IoT === AI applications in Taiwan increasingly involve edge computing, where data processing occurs near the source rather than in centralized cloud systems. This approach reduces latency and bandwidth requirements and is used in smart devices, sensors, and industrial equipment. Edge AI technologies are applied in areas such as smart appliances, industrial automation, and transportation systems. == Education and talent development == Human capital development has been a key focus of Taiwan's AI strategy. The Taiwan AI Academy, established in 2018 with support from Academia Sinica and industry partners, provides training programs for professionals and students aimed at accelerating the adoption of artificial intelligence technologies across industries. The academy offers a range of courses, including executive-level programs, technical training, and specialized tracks in areas such as smart manufacturing, smart healthcare, and edge AI. These programs are designed to provide intensive and practical instruction over relatively short periods. A notable component of the curriculum is project-based learning, in which participants are required to complete proof-of-concept (POC) projects addressing real-world industrial problems. These projects are often developed further for implementation within companies, facilitating technology transfer and commercialization. Between 2018 and 2021, more than 8,000 individuals completed AI training programs across campuses in Taipei, Hsinchu, Taichung, and Tainan. Graduates of the academy have contributed to the introduction of AI systems in sectors such as manufacturing, healthcare, and finance, supporting broader industrial transformation efforts. In addition to the Taiwan AI Academy, universities and research institutions in Taiwan play a significant role in AI education and research. Leading universities have expanded programs in computer science, data science, and machine learning, while research institutes conduct applied and fundamental studies in artificial intelligence. Collaboration between academia, government, and industry is a common feature of Taiwan's AI ecosystem, with joint research projects, internship programs, and technology incubation initiatives supporting talent development. Government-supported initiatives have also sought to attract and retain AI talent, including funding for graduate education, international collaboration programs, and incentives for industry–academic partnerships. These efforts aim to address talent shortages and strengthen Taiwan's capacity in both applied and foundational AI research. == Regulation and governance == Taiwan has developed guidelines and policy frameworks to address the risks associated with AI technologies. In 2023, the Executive Yuan issued guidelines for the use of generative AI in government agencies, focusing on data security and privacy. Ongoing policy discussions hav

  Read more →
• 

  Concept mining

  Concept mining is an activity that results in the extraction of concepts from artifacts. Solutions to the task typically involve aspects of artificial intelligence and statistics, such as data mining and text mining. Because artifacts are typically a loosely structured sequence of words and other symbols (rather than concepts), the problem is nontrivial, but it can provide powerful insights into the meaning, provenance and similarity of documents. == Methods == Traditionally, the conversion of words to concepts has been performed using a thesaurus, and for computational techniques the tendency is to do the same. The thesauri used are either specially created for the task, or a pre-existing language model, usually related to Princeton's WordNet. The mappings of words to concepts are often ambiguous. Typically each word in a given language will relate to several possible concepts. Humans use context to disambiguate the various meanings of a given piece of text, where available machine translation systems cannot easily infer context. For the purposes of concept mining, however, these ambiguities tend to be less important than they are with machine translation, for in large documents the ambiguities tend to even out, much as is the case with text mining. There are many techniques for disambiguation that may be used. Examples are linguistic analysis of the text and the use of word and concept association frequency information that may be inferred from large text corpora. Recently, techniques that base on semantic similarity between the possible concepts and the context have appeared and gained interest in the scientific community. == Applications == === Detecting and indexing similar documents in large corpora === One of the spin-offs of calculating document statistics in the concept domain, rather than the word domain, is that concepts form natural tree structures based on hypernymy and meronymy. These structures can be used to generate simple tree membership statistics, that can be used to locate any document in a Euclidean concept space. If the size of a document is also considered as another dimension of this space then an extremely efficient indexing system can be created. This technique is currently in commercial use locating similar legal documents in a 2.5 million document corpus. === Clustering documents by topic === Standard numeric clustering techniques may be used in "concept space" as described above to locate and index documents by the inferred topic. These are numerically far more efficient than their text mining cousins, and tend to behave more intuitively, in that they map better to the similarity measures a human would generate.

  Read more →
• 

  AI-driven design automation

  AI-driven design automation is the use of artificial intelligence (AI) to automate and improve different parts of the electronic design automation (EDA) process. It is particularly important in the design of integrated circuits (chips) and complex electronic systems, where it can potentially increase productivity, decrease costs, and speed up design cycles. AI Driven Design Automation uses several methods, including machine learning, expert systems, and reinforcement learning. These are used for many tasks, from planning a chip's architecture and logic synthesis to its physical design and final verification. == History == === 1980s–1990s: Expert systems and early experiments === The use of AI for design automation originated in the 1980s and 1990s, mainly with the creation of expert systems. These systems tried to capture the knowledge and practical rules used by human design experts, and used these rules, along with reasoning engines, to direct the design process. A notable early project was the ULYSSES system from Carnegie Mellon University. ULYSSES was a CAD tool integration environment that let expert designers turn their design methods into scripts that could be run automatically. It treated design tools as sources of knowledge that a scheduler could manage. Another example was the ADAM (Advanced Design AutoMation) system at the University of Southern California, which used an expert system called the Design Planning Engine. This engine figured out design strategies on the fly and handled different design jobs by organizing specialized knowledge into structured formats called frames. Other systems like DAA (Design Automation Assistant) used a rule-based approach for specific jobs, such as register transfer level (RTL) design for systems like the IBM 370. Researchers at Carnegie Mellon University also created TALIB, an expert system for mask layout that used over 1200 rules, and EMUCS/DAA for CPU architectural design which used about 70 rules. These projects showed that AI worked better for problems where relatively few rules were required to describe much larger amounts of data. At the same time, there was a surge of tools called silicon compilers like MacPitts, Arsenic, and Palladio. They used algorithms and search techniques to explore different design paradigms. This was another way to automate design, even if it was not always based on expert systems. Early tests with neural networks in VLSI design also happened during this time, although they were not as common as systems based on rules. === 2000s: Introduction of machine learning === In the 2000s, interest in AI for design automation increased. This was mostly because of better machine learning (ML) algorithms and more available data from design and manufacturing. For example, they were used to model and reduce the effects of small manufacturing differences in semiconductor devices. This became very important as the size of components on chips became smaller. The large amount of data created during chip design provided the foundation needed to train smarter ML models. This allowed for predicting outcomes and optimizing in areas that were hard to automate before. === 2016–2020: Reinforcement learning and large scale initiatives === A major turning point happened in the mid to late 2010s, sparked by successes in other areas of AI. The success of DeepMind's AlphaGo in mastering the game of Go inspired researchers. They began to apply reinforcement learning (RL) to difficult EDA problems. These problems often require searching through many options and making a series of decisions. In 2018, the U.S. DARPA started the Intelligent Design of Electronic Assets (IDEA) program. A main goal of IDEA was to create a fully automated layout generator that required no human intervention, able to produce a chip design ready for manufacturing from RTL specifications in 24 hours. Another big initiative was the OpenROAD project, a large effort under IDEA led by UC San Diego with industry and university partners, aimed to build an open source, independent toolchain. It used machine learning, parallelization and divide and conquer approaches. A much-publicized but controversial demonstration of RL's potential came from Google researchers between 2020 and 2021. They created a deep reinforcement learning method for planning the layout of a chip, known as floorplanning. They reported that this method created layouts that were as good as or better than those made by human experts, and it did so in less than six hours. This method used a type of network called a graph convolutional neural network. It showed that it could learn general patterns that could be applied to new problems, getting better as it saw more chip designs. The technology was later used to design Google's Tensor Processing Unit (TPU) accelerators. However, in the original paper, the improvement (if any) from AI was not demonstrated. There was no comparison with existing non-AI tools performing the same task, and since the data is proprietary, no ability for anyone else to perform this comparison. Various efforts to reproduce the AI algorithm, and compare its results with various commercial and academic tools, have yielded mixed results with no conclusive advantage to AI. === 2020s: Autonomous systems and agents === Entering the 2020s, the industry saw the commercial launch of autonomous AI driven EDA systems. For example, Synopsys launched DSO.ai (Design Space Optimization AI) in early 2020, calling it the first autonomous artificial intelligence application for chip design in the industry. This system uses reinforcement learning to search for the best ways to optimize a design within the huge number of possible solutions, trying to improve power, performance, and area (PPA). By 2023, DSO.ai had been used to produce over 100 commercial chips, showing mainstream adoption. Synopsys later grew its AI tools into a suite called Synopsys.ai. The goal was to use AI in the entire EDA workflow, including verification and testing. These advancements, which combine modern AI methods with cloud computing and large data resources, have led to talks about a new phase in EDA. Industry experts and participants sometimes call this 'EDA 4.0'. This new era is defined by the widespread use of AI and machine learning to deal with growing design complexity, automate more of the design process, and help engineers handle the huge amounts of data that EDA tools create. The purpose of EDA 4.0 is to optimize product performance, get products to market faster and make development and manufacturing smoother through intelligent automation. == Applications == Artificial intelligence (AI) is now used in many stages of the electronic design workflow. It aims to improve productivity, get better results, and handle the growing complexity of modern integrated circuits. AI helps designers from the very first ideas about architecture all the way to manufacturing and testing. === High level synthesis and architectural exploration === In the first phases of chip design, AI helps with High Level Synthesis (HLS) and exploring different system level design options (DSE). These processes are key for turning general ideas into detailed hardware plans. AI algorithms, often using supervised learning, are used to build simpler, substitute models. These models can quickly guess important design measurements like area, performance, and power for many different architectural options or HLS settings. For example, the Ithemal tool uses deep neural networks to estimate how fast basic code blocks will run, which helps in making processor architecture decisions. Similarly, PRIMAL uses machine learning estimate power use at the register transfer level (RTL), giving early information about how much power the chip will use. Reinforcement learning (RL) and Bayesian optimization are also used to guide the DSE process. They help search through the many parameters to find the best HLS settings or architectural details like cache sizes. LLMs are also being tested for creating architectural plans or initial C code for HLS, as seen with GPT4AIGChip. === Logic synthesis and optimization === Logic synthesis starts from a high level hardware description and generates an optimized list of electronic gates, known as a gate level netlist, that is ready for placement, routing, and then construction in a specific manufacturing process. AI methods help with different parts of this process, including logic optimization, technology mapping, and making improvements after mapping. Supervised learning, especially with Graph Neural Networks (GNNs), is good at handling data or problems that can be represented as graphs. Since circuit diagrams are instances of directed graphs, supervised learning can help create models that predict design properties like power or error rates in circuits. In logic synthesis and optimization reinforcement learning is used to perform logic optimization directly. In some cases ag

  Read more →
• 

  Synthetic data

  Synthetic data are artificially generated data not produced by real-world events. Typically created using algorithms, synthetic data can be deployed to validate mathematical models and to train machine learning models. Data generated by a computer simulation can be seen as synthetic data. This encompasses most applications of physical modeling, such as music synthesizers or flight simulators. The output of such systems approximates the real thing, but is fully algorithmically generated. Synthetic data is used in a variety of fields as a filter for information that would otherwise compromise the confidentiality of particular aspects of the data. In many sensitive applications, datasets theoretically exist but cannot be released to the general public; synthetic data sidesteps the privacy issues that arise from using real consumer information without permission or compensation. == Usefulness == Synthetic data is generated to meet specific needs or certain conditions that may not be found in the original, real data. One of the hurdles in applying up-to-date machine learning approaches for complex scientific tasks is the scarcity of labeled data, a gap effectively bridged by the use of synthetic data, which closely replicates real experimental data. This can be useful when designing many systems, from simulations based on theoretical value, to database processors, etc. This helps detect and solve unexpected issues such as information processing limitations. Synthetic data are often generated to represent the authentic data and allows a baseline to be set. Another benefit of synthetic data is to protect the privacy and confidentiality of authentic data, while still allowing for use in testing systems. Computer security experts claim generated synthetic data "... enables us to create realistic behavior profiles for users and attackers. The data is used to train the fraud detection system itself, thus creating the necessary adaptation of the system to a specific environment." In defense and military contexts, synthetic data is seen as a potentially valuable tool to develop and improve complex AI systems, particularly in contexts where high-quality real-world data is scarce. At the same time, synthetic data together with the testing approach can give the ability to model real-world scenarios. == History == Scientific modelling of physical systems has a long history that runs concurrent with the history of physics. For example, research into synthesis of audio and voice can be traced back to the 1930s and before, driven forward by the developments of the telephone and audio recording technologies. Digitization gave rise to software synthesizers from the 1970s onwards. In the context of privacy-preserving statistical analysis, in 1993, the idea of original fully synthetic data was created by Donald Rubin. Rubin originally designed this to synthesize the Decennial Census long form responses for the short form households. He then released samples that did not include any actual long form records - in this he preserved anonymity of the household. Later that year, the idea of original partially synthetic data was created by Little. Little used this idea to synthesize the sensitive values on the public use file. A 1993 work fitted a statistical model to 60,000 MNIST digits, then it was used to generate over 1 million examples. Those were used to train a LeNet-4 to reach state of the art performance. In 1994, Stephen Fienberg introduced 'critical refinement', in which a parametric posterior predictive distribution (instead of a Bayes bootstrap) is used to do the sampling. Later, other important contributors to the development of synthetic data generation were Trivellore Raghunathan, Jerry Reiter, Donald Rubin, John M. Abowd, and Jim Woodcock. Collectively they came up with a solution for how to treat partially synthetic data with missing data. Similarly, they developed the technique of Sequential Regression Multivariate Imputation. == Calculations == Researchers test the framework on synthetic data, which is "the only source of ground truth on which they can objectively assess the performance of their algorithms". Synthetic data can be generated through the use of random lines, having different orientations and starting positions. Datasets can get fairly complicated. A more complicated dataset can be generated by using a synthesizer build. To create a synthesizer build, first use the original data to create a model or equation that fits the data the best. This model or equation will be called a synthesizer build. This build can be used to generate more data. Constructing a synthesizer build involves constructing a statistical model. In a linear regression line example, the original data can be plotted, and a best fit linear line can be created from the data. This line is a synthesizer created from the original data. The next step will be generating more synthetic data from the synthesizer build or from this linear line equation. In this way, the new data can be used for studies and research, and it protects the confidentiality of the original data. David Jensen from the Knowledge Discovery Laboratory explains how to generate synthetic data: "Researchers frequently need to explore the effects of certain data characteristics on their data model." To help construct datasets exhibiting specific properties, such as auto-correlation or degree disparity, proximity can generate synthetic data having one of several types of graph structure: random graphs that are generated by some random process; lattice graphs having a ring structure; lattice graphs having a grid structure, etc. In all cases, the data generation process follows the same process: Generate the empty graph structure. Generate attribute values based on user-supplied prior probabilities. Since the attribute values of one object may depend on the attribute values of related objects, the attribute generation process assigns values collectively. == Applications == === Fraud detection and confidentiality systems === Testing and training fraud detection and confidentiality systems are devised using synthetic data. Specific algorithms and generators are designed to create realistic data, which then assists in teaching a system how to react to certain situations or criteria. For example, intrusion detection software is tested using synthetic data. This data is a representation of the authentic data and may include intrusion instances that are not found in the authentic data. The synthetic data allows the software to recognize these situations and react accordingly. If synthetic data was not used, the software would only be trained to react to the situations provided by the authentic data and it may not recognize another type of intrusion. === Scientific research === Researchers doing clinical trials or any other research may generate synthetic data to aid in creating a baseline for future studies and testing. Real data can contain information that researchers may not want released, so synthetic data is sometimes used to protect the privacy and confidentiality of a dataset. Using synthetic data reduces confidentiality and privacy issues since it holds no personal information and cannot be traced back to any individual. Beyond privacy protection, synthetic data is also being explored for methodological innovation in drug development. For instance, synthetic data may be used to construct synthetic control arms as an alternative to conventional external control arms based on real-world data (RWD) or randomized controlled trials (RCTs). Collectively, regulatory agencies such as the FDA and EMA appear to be at various stages of recognizing and integrating AI-generated synthetic data into their methodologies. While there is growing consensus on the potential of such data to support model development and the broader lifecycle of medicinal products, to date no drug or medical device has been approved using solely or predominantly synthetic data—particularly not as a comparator arm generated entirely via data-driven algorithms. The quality and statistical handling of synthetic data are expected to become more prominent in future regulatory discussions, particularly in contexts such as predictive modeling (e.g., digital twins), where innovative approaches have already been referenced. === Machine learning === Synthetic data is increasingly being used for machine learning applications: a model is trained on a synthetically generated dataset with the intention of transfer learning to real data. Efforts have been made to enable more data science experiments via the construction of general-purpose synthetic data generators, such as the Synthetic Data Vault. In general, synthetic data has several natural advantages: once the synthetic environment is ready, it is fast and cheap to produce as much data as needed; synthetic data can have perfectly accurate labels, including labeling that may be very expensive or impo

  Read more →
• 

  Weak stability boundary

  Weak stability boundary (WSB), including low-energy transfer, is a concept introduced by Edward Belbruno in 1987. The concept explained how a spacecraft could change orbits using very little fuel. Weak stability boundary is defined for the three-body problem. This problem considers the motion of a particle P of negligible mass moving with respect to two larger bodies, P1, P2, modeled as point masses, where these bodies move in circular or elliptical orbits with respect to each other, and P2 is smaller than P1. The force between the three bodies is the classical Newtonian gravitational force. For example, P1 is the Earth, P2 is the Moon and P is a spacecraft; or P1 is the Sun, P2 is Jupiter and P is a comet, etc. This model is called the restricted three-body problem. The weak stability boundary defines a region about P2 where P is temporarily captured. This region is in position-velocity space. Capture means that the Kepler energy between P and P2 is negative. This is also called weak capture. == Background == This boundary was defined for the first time by Edward Belbruno of Princeton University in 1987. He described a Low-energy transfer which would allow a spacecraft to change orbits using very little fuel. It was for motion about Moon (P2) with P1 = Earth. It is defined algorithmically by monitoring cycling motion of P about the Moon and finding the region where cycling motion transitions between stable and unstable after one cycle. Stable motion means P can completely cycle about the Moon for one cycle relative to a reference section, starting in weak capture. P needs to return to the reference section with negative Kepler energy. Otherwise, the motion is called unstable, where P does not return to the reference section within one cycle or if it returns, it has non-negative Kepler energy. The set of all transition points about the Moon comprises the weak stability boundary, W. The motion of P is sensitive or chaotic as it moves about the Moon within W. A mathematical proof that the motion within W is chaotic was given in 2004. This is accomplished by showing that the set W about an arbitrary body P2 in the restricted three-body problem contains a hyperbolic invariant set of fractional dimension consisting of the infinitely many intersections Hyperbolic manifolds. The weak stability boundary was originally referred to as the fuzzy boundary. This term was used since the transition between capture and escape defined in the algorithm is not well defined and limited by the numerical accuracy. This defines a "fuzzy" location for the transition points. It is also due the inherent chaos in the motion of P near the transition points. It can be thought of as a fuzzy chaos region. As is described in an article in Discover magazine, the WSB can be roughly viewed as the fuzzy edge of a region, referred to as a gravity well, about a body (the Moon), where its force of gravity becomes small enough to be dominated by force of gravity of another body (the Earth) and the motion there is chaotic. A much more general algorithm defining W was given in 2007. It defines W relative to n-cycles, where n = 1,2,3,..., yielding boundaries of order n. This gives a much more complex region consisting of the union of all the weak stability boundaries of order n. This definition was explored further in 2010. The results suggested that W consists, in part, of the hyperbolic network of invariant manifolds associated to the Lyapunov orbits about the L1, L2 Lagrange points near P2. The explicit determination of the set W about P2 = Jupiter, where P1 is the Sun, is described in "Computation of Weak Stability Boundaries: Sun-Jupiter Case". It turns out that a weak stability region can also be defined relative to the larger mass point, P1. A proof of the existence of the weak stability boundary about P1 was given in 2012, but a different definition is used. The chaos of the motion is analytically proven in "Geometry of Weak Stability Boundaries". The boundary is studied in "Applicability and Dynamical Characterization of the Associated Sets of the Algorithmic Weak Stability Boundary in the Lunar Sphere of Influence". == Applications == There are a number of important applications for the weak stability boundary (WSB). Since the WSB defines a region of temporary capture, it can be used, for example, to find transfer trajectories from the Earth to the Moon that arrive at the Moon within the WSB region in weak capture, which is called ballistic capture for a spacecraft. No fuel is required for capture in this case. This was numerically demonstrated in 1987. This is the first reference for ballistic capture for spacecraft and definition of the weak stability boundary. The boundary was operationally demonstrated to exist in 1991 when it was used to find a ballistic capture transfer to the Moon for Japan's Hiten spacecraft. Other missions have used the same transfer type as Hiten, including Grail, Capstone, Danuri, Hakuto-R Mission 1 and SLIM. The WSB for Mars is studied in "Earth-Mars Transfers with Ballistic Capture" and ballistic capture transfers to Mars are computed. The BepiColombo mission of ESA should achieve ballistic capture at the WSB of Mercury in November 2026. The WSB region can be used in the field of Astrophysics. It can be defined for stars within open star clusters. This is done in "Chaotic Exchange of Solid Material Between Planetary Systems: Implications for the Lithopanspermia Hypothesis" to analyze the capture of solid material that may have arrived on the Earth early in the age of the Solar System to study the validity of the lithopanspermia hypothesis. Numerical explorations of trajectories for P starting in the WSB region about P2 show that after the particle P escapes P2 at the end of weak capture, it moves about the primary body, P1, in a near resonant orbit, in resonance with P2 about P1. This property was used to study comets that move in orbits about the Sun in orbital resonance with Jupiter, which change resonance orbits by becoming weakly captured by Jupiter. An example of such a comet is 39P/Oterma. This property of change of resonance of orbits about P1 when P is weakly captured by the WSB of P2 has an interesting application to the field of quantum mechanics to the motion of an electron about the proton in a hydrogen atom. The transition motion of an electron about the proton between different energy states described by the Schrödinger equation is shown to be equivalent to the change of resonance of P about P1 via weak capture by P2 for a family of transitioning resonance orbits. This gives a classical model using chaotic dynamics with Newtonian gravity for the motion of an electron.

  Read more →
• 

  Outline of deep learning

  The following outline is provided as an overview of, and topical guide to, deep learning: Deep learning is a subfield of machine learning and artificial intelligence based on artificial neural networks with multiple processing layers. It emphasizes representation learning and is widely used in areas such as computer vision, natural language processing, speech recognition, recommender systems, robotics, and generative artificial intelligence. == Ways to categorize deep learning == A field of study A branch of artificial intelligence A subfield of machine learning A subfield of computer science A form of representation learning A class of methods based on artificial neural networks An approach used in computational statistics == History == === Precursors === Cybernetics Perceptron Connectionism Neocognitron Backpropagation === Milestones === LeNet Long short-term memory Deep belief network AlexNet Sequence to sequence learning Generative adversarial network Residual neural network Transformer BERT Generative pre-trained transformer Diffusion model === Related histories === History of artificial intelligence History of machine learning Timeline of machine learning == Core concepts == == Learning settings == Supervised learning Unsupervised learning Self-supervised learning Semi-supervised learning Reinforcement learning Transfer learning Multitask learning Multimodal learning Online machine learning Continual learning == Common tasks == Image classification Object detection Image segmentation Automatic speech recognition Neural machine translation Question answering Automatic summarization Text-to-image model Protein structure prediction == Architectures == === Feedforward and convolutional architectures === Feedforward neural network Multilayer perceptron Convolutional neural network Radial basis function network Residual neural network U-Net === Recurrent and sequence architectures === Recurrent neural network Long short-term memory Gated recurrent unit Sequence to sequence learning Recursive neural network === Representation-learning architectures === Autoencoder Denoising autoencoder Sparse autoencoder Variational autoencoder Restricted Boltzmann machine Deep belief network === Attention and transformer architectures === Attention (machine learning) Transformer BERT Generative pre-trained transformer Vision transformer === Generative and probabilistic architectures === Autoregressive model Diffusion model Energy-based model Generative adversarial network Mixture of experts === Graph and memory architectures === Graph neural network Graph convolutional network Siamese network Neural Turing machine Memory network Echo state network Capsule neural network == Neural network components and techniques == Artificial neuron Activation function Rectified linear unit Sigmoid function Softmax function Embedding Convolution Pooling layer Attention Batch normalization Layer normalization Residual connections == Training and optimization == Backpropagation Gradient descent Stochastic gradient descent Adam optimization Learning rate Loss function Cross-entropy Mean squared error Regularization Dropout Early stopping Batch normalization Data augmentation Transfer learning Knowledge distillation Ensemble learning Curriculum learning == Datasets and benchmarks == CIFAR-10 ImageNet MNIST database Common Objects in Context (COCO) General Language Understanding Evaluation (GLUE) benchmark LibriSpeech SQuAD == Applications == === Computer vision === Computer vision Facial recognition system Image classification Image segmentation Medical imaging Object detection Optical character recognition === Natural language processing === Automatic summarization Chatbot Information retrieval Large language model Natural language processing Neural machine translation Question answering Sentiment analysis === Speech and audio === Automatic speech recognition Music information retrieval Speaker recognition Speech synthesis === Science and medicine === Bioinformatics Computational biology Drug discovery Medical diagnosis Protein structure prediction === Robotics and control === Autonomous car Computer game bot Control theory Robotics === Recommendation, search, and forecasting === Anomaly detection Forecasting Fraud detection Recommender system Search engine === Generative artificial intelligence === Deepfake Generative artificial intelligence Large language model Speech synthesis Text-to-image model === Computer graphics and video games === Deep Learning Anti-Aliasing (DLAA) Deep Learning Super Sampling (DLSS) == Hardware == AMD Instinct AMD XDNA Application-specific integrated circuit Deep learning processor, Neural processing unit (NPU), or Neural Engine Field-programmable gate array General-purpose computing on graphics processing units (GPGPU) Graphics processing unit NVIDIA Deep Learning Accelerator (NVDLA) Tensor processing unit Vision processing unit Wafer-scale integration === Supporting software platforms === CUDA Metal ROCm == Software == === Open-source frameworks and libraries === === Neural network software === EDLUT Emergent Encog JOONE Neuroph NeuroSolutions OpenNN Peltarion Synapse SNNS === Platforms, tools, and deployment === Amazon SageMaker Google Colab Hugging Face Kaggle Kubeflow MLflow ONNX OpenVINO TensorFlow Hub == Algorithms for deep learning and neural networks == Backpropagation Conjugate gradient method Generalized Hebbian algorithm Gradient descent Levenberg–Marquardt algorithm Perceptron Quasi-Newton method Wake-sleep algorithm == Methods and related topics == === Representation and metric learning === Contrastive learning Embedding Feature learning Manifold learning Metric learning === Generative modeling === Autoregressive model Diffusion model Generative adversarial network Generative model Variational inference === Efficient and scalable deep learning === Knowledge distillation Low-rank approximation Mixture of experts Quantization Sparsity === Reliability, safety, and interpretability === Adversarial machine learning AI alignment Algorithmic bias Catastrophic forgetting Differential privacy Explainable artificial intelligence Federated learning Hallucination (artificial intelligence) == Conferences and workshops == Annual Meeting of the Association for Computational Linguistics Conference on Computer Vision and Pattern Recognition Conference on Neural Information Processing Systems International Conference on Computer Vision International Conference on Learning Representations International Conference on Machine Learning == Organizations == === Research laboratories and institutions === Allen Institute for AI Alberta Machine Intelligence Institute European Laboratory for Learning and Intelligent Systems Google DeepMind Meta AI Mila Microsoft Research Vector Institute === Companies === Anthropic Cerebras Cohere DeepSeek Mistral AI OpenAI Stability AI xAI == Publications == === Books === Deep Learning – Ian Goodfellow and Yoshua Bengio Neural Networks and Deep Learning – Michael Nielsen Perceptrons – Marvin Minsky and Seymour Papert === Journals === IEEE Transactions on Neural Networks and Learning Systems Neural Networks Neural Computation == Influential persons ==

  Read more →
• 

  Algorithmic mechanism design

  Algorithmic mechanism design (AMD) lies at the intersection of economic game theory, optimization, and computer science. The prototypical problem in mechanism design is to design a system for multiple self-interested participants, such that the participants' self-interested actions at equilibrium lead to good system performance. Typical objectives studied include revenue maximization and social welfare maximization. Algorithmic mechanism design differs from classical economic mechanism design in several respects. It typically employs the analytic tools of theoretical computer science, such as worst case analysis and approximation ratios, in contrast to classical mechanism design in economics which often makes distributional assumptions about the agents. It also considers computational constraints to be of central importance: mechanisms that cannot be efficiently implemented in polynomial time are not considered to be viable solutions to a mechanism design problem. This often, for example, rules out the classic economic mechanism, the Vickrey–Clarke–Groves auction. == History == Noam Nisan and Amir Ronen first coined "Algorithmic mechanism design" in a research paper published in 1999.

  Read more →
• 

  Storage area network

  A storage area network (SAN) or storage network is a computer network which provides access to consolidated, block-level data storage. SANs are primarily used to access data storage devices, such as disk arrays and tape libraries from servers so that the devices appear to the operating system as direct-attached storage. A SAN typically is a dedicated network of storage devices not accessible through the local area network (LAN). Although a SAN provides only block-level access, file systems built on top of SANs do provide file-level access and are known as shared-disk file systems. Newer SAN configurations enable hybrid SAN and allow traditional block storage that appears as local storage but also object storage for web services through APIs. == Storage architectures == Storage area networks (SANs) are sometimes referred to as network behind the servers and historically developed out of a centralized data storage model, but with its own data network. A SAN is, at its simplest, a dedicated network for data storage. In addition to storing data, SANs allow for the automatic backup of data, and the monitoring of the storage as well as the backup process. A SAN is a combination of hardware and software. It grew out of data-centric mainframe architectures, where clients in a network can connect to several servers that store different types of data. To scale storage capacities as the volumes of data grew, direct-attached storage (DAS) was developed, where disk arrays or just a bunch of disks (JBODs) were attached to servers. In this architecture, storage devices can be added to increase storage capacity. However, the server through which the storage devices are accessed is a single point of failure, and a large part of the LAN network bandwidth is used for accessing, storing and backing up data. To solve the single point of failure issue, a direct-attached shared storage architecture was implemented, where several servers could access the same storage device. DAS was the first network storage system and is still widely used where data storage requirements are not very high. Out of it developed the network-attached storage (NAS) architecture, where one or more dedicated file server or storage devices are made available in a LAN. Therefore, the transfer of data, particularly for backup, still takes place over the existing LAN. If more than a terabyte of data was stored at any one time, LAN bandwidth became a bottleneck. Therefore, SANs were developed, where a dedicated storage network was attached to the LAN, and terabytes of data are transferred over a dedicated high speed and bandwidth network. Within the SAN, storage devices are interconnected. Transfer of data between storage devices, such as for backup, happens behind the servers and is meant to be transparent. In a NAS architecture data is transferred using the TCP and IP protocols over Ethernet. Distinct protocols were developed for SANs, such as Fibre Channel, iSCSI, Infiniband. Therefore, SANs often have their own network and storage devices, which have to be bought, installed, and configured. This makes SANs inherently more expensive than NAS architectures. == Components == SANs have their own networking devices, such as SAN switches. To access the SAN, so-called SAN servers are used, which in turn connect to SAN host adapters. Within the SAN, a range of data storage devices may be interconnected, such as SAN-capable disk arrays, JBODs and tape libraries. === Host layer === Servers that allow access to the SAN and its storage devices are said to form the host layer of the SAN. Such servers have host adapters, which are cards that attach to slots on the server motherboard (usually PCI slots) and run with a corresponding firmware and device driver. Through the host adapters the operating system of the server can communicate with the storage devices in the SAN. In Fibre channel deployments, a cable connects to the host adapter through the gigabit interface converter (GBIC). GBICs are also used on switches and storage devices within the SAN, and they convert digital bits into light impulses that can then be transmitted over the Fibre Channel cables. Conversely, the GBIC converts incoming light impulses back into digital bits. The predecessor of the GBIC was called gigabit link module (GLM). === Fabric layer === The fabric layer consists of SAN networking devices that include SAN switches, routers, protocol bridges, gateway devices, and cables. SAN network devices move data within the SAN, or between an initiator, such as an HBA port of a server, and a target, such as the port of a storage device. When SANs were first built, hubs were the only devices that were Fibre Channel capable, but Fibre Channel switches were developed and hubs are now rarely found in SANs. Switches have the advantage over hubs that they allow all attached devices to communicate simultaneously, as a switch provides a dedicated link to connect all its ports with one another. When SANs were first built, Fibre Channel had to be implemented over copper cables, these days multimode optical fibre cables are used in SANs. SANs are usually built with redundancy, so SAN switches are connected with redundant links. SAN switches connect the servers with the storage devices and are typically non-blocking allowing transmission of data across all attached wires at the same time. SAN switches are for redundancy purposes set up in a meshed topology. A single SAN switch can have as few as 8 ports and up to 32 ports with modular extensions. So-called director-class switches can have as many as 128 ports. In switched SANs, the Fibre Channel switched fabric protocol FC-SW-6 is used under which every device in the SAN has a hardcoded World Wide Name (WWN) address in the host bus adapter (HBA). If a device is connected to the SAN its WWN is registered in the SAN switch name server. In place of a WWN, or worldwide port name (WWPN), SAN Fibre Channel storage device vendors may also hardcode a worldwide node name (WWNN). The ports of storage devices often have a WWN starting with 5, while the bus adapters of servers start with 10 or 21. === Storage layer === The serialized Small Computer Systems Interface (SCSI) protocol is often used on top of the Fibre Channel switched fabric protocol in servers and SAN storage devices. The Internet Small Computer Systems Interface (iSCSI) over Ethernet and the Infiniband protocols may also be found implemented in SANs, but are often bridged into the Fibre Channel SAN. However, Infiniband and iSCSI storage devices, in particular, disk arrays, are available. The various storage devices in a SAN are said to form the storage layer. It can include a variety of hard disk and magnetic tape devices that store data. In SANs, disk arrays are joined through a RAID which makes a lot of hard disks look and perform like one big storage device. Every storage device, or even partition on that storage device, has a logical unit number (LUN) assigned to it. This is a unique number within the SAN. Every node in the SAN, be it a server or another storage device, can access the storage by referencing the LUN. The LUNs allow for the storage capacity of a SAN to be segmented and for the implementation of access controls. A particular server, or a group of servers, may, for example, be only given access to a particular part of the SAN storage layer, in the form of LUNs. When a storage device receives a request to read or write data, it will check its access list to establish whether the node, identified by its LUN, is allowed to access the storage area, also identified by a LUN. LUN masking is a technique whereby the host bus adapter and the SAN software of a server restrict the LUNs for which commands are accepted. In doing so LUNs that should never be accessed by the server are masked. Another method to restrict server access to particular SAN storage devices is fabric-based access control, or zoning, which is enforced by the SAN networking devices and servers. Under zoning, server access is restricted to storage devices that are in a particular SAN zone. == Network protocols == A mapping layer to other protocols is used to form a network: ATA over Ethernet (AoE), mapping of AT Attachment (ATA) over Ethernet Fibre Channel Protocol (FCP), a mapping of SCSI over Fibre Channel Fibre Channel over Ethernet (FCoE) ESCON over Fibre Channel (FICON), used by mainframe computers HyperSCSI, mapping of SCSI over Ethernet iFCP or SANoIP mapping of FCP over IP iSCSI, mapping of SCSI over TCP/IP iSCSI Extensions for RDMA (iSER), mapping of iSCSI over InfiniBand Network block device, mapping device node requests on UNIX-like systems over stream sockets like TCP/IP SCSI RDMA Protocol (SRP), another SCSI implementation for remote direct memory access (RDMA) transports Storage networks may also be built using Serial Attached SCSI (SAS) and Serial ATA (SATA) technologies. SAS evolved from SCSI direct-attached storage. SATA evolved from Para

  Read more →