No Thanks is a Palestinian boycott-awareness mobile application developed by Palestinian software engineer Ahmed Bashbash, created to assist consumers in identifying and boycotting products associated with companies linked to Israel. Launched in 13 November 2023, the app gained significant attention amid the Gaza–Israel conflict. == History == No Thanks is a mobile application developed by Ahmed Bashbash, a Palestinian software engineer from Gaza residing in Hungary. The app was conceived in October 2023 following the death of Bashbash's brother in an Israeli airstrike on October 31, 2023. His sister had previously died in 2020 due to delayed medical treatment. The app was officially launched on November 13, 2023, and quickly gained traction, got over 100,000 downloads within its first month of release. On November 30, 2023, Google removed the app from its Play Store due to a violation of its content policies. The app's home page included a description: "Welcome to No Thanks, here you can see if the product in your hand supports killing children in Palestine or not," which was deemed to contravene Google's guidelines on hate speech and sensitive content. On December 3, 2023, following changes to the app's description, Google reinstated the app.
Apps to analyse COVID-19 sounds
Apps to analyse COVID-19 sounds are mobile software applications designed to collect respiratory sounds and aid diagnosis in response to the COVID-19 pandemic. Numerous applications are in development, with different institutions and companies taking various approaches to privacy and data collection. Current efforts are aimed at gathering data. In a later stage, it is possible that sound apps will have the capacity (and ethical approvals) to provide information back to users. In order to develop and train signal analysis approaches, large datasets are required. == History == The COVID-19 outbreak was announced as a global pandemic by the World Health Organization in March 2020 and has affected a growing number of people globally. In this context, advanced artificial intelligence techniques are being considered as tools in aiding our response to global health crisis. Other COVID-19 apps which offer solutions for user tracking have been developed. At the same time a number of approaches which tries to use respiratory sounds and artificial intelligence to understand if the disease can be diagnosed have been proposed. A few studies are available as preprints (i.e. not yet peer-reviewed) documents. == Methodologies == The potential for using speech and sound analysis by artificial intelligence to help in this scenario, by surveying which types of related or contextually significant phenomena can be automatically assessed from speech or sound has been recently overviewed. These include the automatic recognition and monitoring of breathing, dry and wet coughing or sneezing sounds, speech under cold, eating behaviour, sleepiness, or pain. Additionally, the potential use-cases of intelligent speech analysis for COVID-19 diagnosed patients has also been presented. In particular, by analysing speech recordings from these patients, an audio-only-based model to automatically categorise the health state of patients from four aspects, including the severity of illness, sleep quality, fatigue, and anxiety, is constructed. This work shows promise in estimating the severity of illness. Machine learning methods have been explored to recognize and diagnose coughs from different diseases. These included a low complexity, automated recognition and diagnostic tool for screening respiratory infections that utilizes convolutional neural networks (CNNs) to detect cough within environment audio and diagnose three potential illnesses (i.e. bronchitis, bronchiolitis and pertussis) based on their unique cough audio features. A large-scale crowdsourced dataset of respiratory sounds has been collected to aid diagnosis of COVID-19: coughs and breathing sounds are sufficient to distinguish users affected by COVID-19 versus those affected by asthma or healthy controls. Behind these studies is the ambition that automated systems to screen for respiratory diseases based on voice, raw cough or other sound data would have positive medical applications in both clinical and public health arenas. == List of apps to analyse COVID-19 sounds ==
Korean Decimal Classification
The Korean Decimal Classification (KDC) is a system of library classification used in South Korea. The structure and main level classes of the KDC are based on the Dewey Decimal Classification. The KDC is maintained and published by the Classification Committee of the Korean Library Association. The first edition of the classification was published in 1964; the most recent edition is the sixth edition published in 2013. Almost all school and public libraries in South Korea use the KDC to organize their collections, as well as the National Library of Korea and some university libraries. == History == Multiple library classification systems had been developed for Korean libraries before the publication of the KDC. These included the Railway Bureau Library Classification(1920), the Korean Decimal Classification edited by Bong-Suk Park(known as KDCP, 1947), the Han-Un Decimal Classification(1954), and the Kuk-Yeon Decimal Classification(1958). After the disappearance of editor Bong-Suk Park in the 1950s, the KDCP system decreased in use. Korean librarians considered adopting the Dewey Decimal Classification (DDC), especially after it was implemented at Yonsei University in 1957, but struggled to apply it to East Asian and Korean-focused works in their collections. In February 1963, members of the Korean Library Association's Classification were appointed to create a national classification; they decided to make revisions to the order of the main classes of the DDC, for example bringing together the class Language(700) together with the class for Literature(800). Committee members prepared draft classes and indexes and the first edition of the KDC was published in May 1964. Both the text and the index were written in Korean Hangul characters and Chinese characters. The second edition was published just two years later, in 1966, correcting errors and omissions found in the first edition. The third edition was published in 1980, maintaining the basic framework of the previous editions while expanding significantly. The fourth edition, published in 1996, made considerable changes, including increasing the number of representatives on the Classification Committee. The committee sought feedback from the library community and implemented revisions included in the recently published edition 20 of the DDC and edition 9 of the Nippon Decimal Classification. New policies applied to the fourth edition included principles suggesting the main classes should remain as static as possible, with focus shown to expanding classes devoted to technology and science. Likewise, many subject specialists were consulted for the publication of the fifth edition in 2009. The publication of the 23rd edition of the DDC in 2011 provided opportunity for a new revision of the KDC, and the sixth edition was published in July 2013. Greater numbers of classes provided number building capacity in the sixth edition, allowing for more specificity. == Description == The KDC classifies resources primarily by discipline, though some classes are collocated by subject. There are eight auxiliary mnemonic tables used to expand class numbers. The main classes of the KDC are the same as the main classes of the Dewey Decimal Classification, but four of those main classes are in a different order: Natural sciences (400), Technology and engineering (500), Arts (600), and Language 700. Though the structure is heavily influenced by the DDC, aspects of multiple library classifications have been invoked in the creation of the KDC, including the Library of Congress Classification for the arrangement of the social sciences (300), the Universal Decimal Classification for medical sciences (510), the KDCP for Korean and Oriental subjects, the Nippon Decimal Classification for those of Japan and Oriental subjects. === Classes of the KDC 6th edition === 000 General works 000 General works 010 Books, Bibliography 020 Library & information science 030 General encyclopedias 040 General collected essays 050 General serial publications 060 General societies 070 Newspapers, journalism 080 General collected works 090 Materials of province 100 Philosophy 100 Philosophy 110 Metaphysics 120 Epistemology, etc. 130 Systems of philosophy 140 Chinese classics 150 Oriental philosophy and thought 160 Western philosophy 170 Logic 180 Psychology 190 Ethics, moral philosophy 200 Religion 200 Religion 210 Comparative religion 220 Buddhism 230 Christian religion 240 Taoism 250 Chondoism 260 [Unassigned] 270 Hinduism, Brahmanism 280 Islam, Mohammedianism 290 Other religions 300 Social sciences 300 Social sciences 310 Statistics 320 Economics 330 Sociology and social problems 340 Political sciences 350 Public administration 360 Law 370 Education 380 Customs, Etiquette, Folklore 390 Military science 400 Natural sciences 400 Natural sciences 410 Mathematics 420 Physics 430 Chemistry 440 Astronomy 450 Earth science 460 Mineralogy 470 Life science 480 Botany 490 Zoological science 500 Technology 500 Technology 510 Medical science 520 Agriculture 530 Engineering, technology, etc. 540 Construction and architecture 550 Mechanical engineering 560 Electrical, comm. & electric engineering 570 Chemical engineering 580 Manufactures 590 Human ecology 600 Arts 600 Arts 610 [Unassigned] 620 Sculpture, plastic art 630 Crafts 640 Calligraphy 650 Painting, design 660 Photography 670 Music 680 Stage performance, museum arts 690 Amusements, sports & physical training 700 Language 700 Language 710 Korean language 720 Chinese language 730 Japanese & other Asian languages 740 English 750 German 760 French languages 770 Spanish languages & Portuguese language 780 Italian languages 790 Other languages 800 Literature 800 Literature 810 Korean literature 820 Chinese literature 830 Japanese & other Asian literature 840 English & American literature 850 German literature 860 French literature 870 Spanish & Portuguese literature 880 Italian literature 890 Other literatures 900 History 900 History 910 Asia 920 Europe 930 Africa 940 North America 950 South America 960 Oceania and Polar regions 970 [Unassigned] 980 Geography 990 Biography === Expansion tables === Table 1. Standard subdivisions Table 2. Geographic Areas Table 3. Korean geographic areas Table 4. Korean historical period Table 5. Languages Table 6. Subdivisions of individual languages Table 7. Subdivisions of individual literatures Table 8. Subdivisions of individual religions == Usage == KDC is used by a wide range of libraries within Korea, including by the National Library of Korea and most school and public libraries in the country, along with some university libraries, such as the one at Keimyung University.
Demis Hassabis
Sir Demis Hassabis (/ˈdɛ.mɪs/ DE-mis /hɑːˈsɑː.bis/ hah-SAH-bees; born Dimitrios Hassapis, Greek: Δημήτριος Χασάπης, 27 July 1976) is a British artificial intelligence (AI) researcher and entrepreneur. He is the chief executive officer and co-founder of Google DeepMind and Isomorphic Labs, and a UK Government AI Adviser. In 2024, Hassabis and John M. Jumper were jointly awarded the Nobel Prize in Chemistry for their AI research contributions to protein structure prediction. Hassabis is a Fellow of the Royal Society and has won awards for his research efforts, including the Breakthrough Prize, the Canada Gairdner International Award and the Lasker Award. He was appointed a CBE in 2017, and knighted in 2024 for his work on AI. He was also listed among the Time 100 most influential people in the world in 2017 and 2025, and was one of the "Architects of AI" collectively chosen as Time's 2025 Person of the Year. == Early life and education == Hassabis was born to Costas and Angela Hassapis. His father is a Greek Cypriot and his mother is a Chinese Singaporean. Demis grew up in North London. His original surname was "Hassapis" (Greek: Χασάπης), meaning "butcher" in Greek, but he later, according to Ingo Althöfer, "executed a point mutation by changing ‘p’ to ‘b’". One of his younger brothers still carries the original surname. In his early career, he was a video game AI programmer and designer, and an expert board games player. A child prodigy in chess from the age of four, when he first learnt chess by watching his father playing against his uncle, Hassabis reached master standard at the age of 13 with an Elo rating of 2300 and captained many of the England junior chess teams. He represented the University of Cambridge in the Oxford–Cambridge varsity chess matches of 1995, 1996 and 1997, winning a half blue. He first got interested in technology after buying his first computer in 1984, a ZX Spectrum 48K, funded from chess winnings. He taught himself how to program from books. He subsequently wrote his first AI program on a Commodore Amiga to play the reversi board game. Between 1988 and 1990, Hassabis was educated at Queen Elizabeth's School, Barnet, a boys' grammar school in North London. He was subsequently home-schooled by his parents for a year, before studying at the comprehensive school of Christ's College in East Finchley. He completed his A-level exams two years early at 16. === Bullfrog Productions === Asked by Cambridge University to take a gap year owing to his young age, Hassabis began his computer games career at Bullfrog Productions after entering an Amiga Power "Win-a-job-at-Bullfrog" competition. He began by playtesting on Syndicate and then at 17 co-designing and lead-programming on the 1994 game Theme Park, with the game's designer Peter Molyneux. Theme Park, a simulation video game, sold several million copies and inspired a whole genre of simulation sandbox games. Despite being offered a seven-figure sum to remain in the games industry, he turned it down. He earned enough from his gap year to pay his own way through university. === University of Cambridge === Hassabis left Bullfrog to study at Queens' College of the University of Cambridge, where he completed the Computer Science Tripos and graduated in 1997 with a double first. == Career and research == === Lionhead === After graduating from Cambridge, Hassabis worked at Lionhead Studios. Games designer Peter Molyneux, with whom Hassabis had worked at Bullfrog Productions, had recently founded the company. At Lionhead, Hassabis worked as lead AI programmer on the 2001 god game Black & White. === Elixir Studios === Hassabis left Lionhead in 1998 to found Elixir Studios, a London-based independent games developer, signing publishing deals with Eidos Interactive, Vivendi Universal and Microsoft. In addition to managing the company, Hassabis served as executive designer of the games Republic: The Revolution and Evil Genius. Each received BAFTA nominations for their interactive music scores, created by James Hannigan. The release of Elixir's first game, Republic: The Revolution, a highly ambitious and unusual political simulation game, was delayed due to its huge scope, which involved an AI simulation of the workings of an entire fictional country. The final game was reduced from its original vision and greeted with lukewarm reviews, receiving a Metacritic score of 62/100. Evil Genius, a tongue-in-cheek Austin Powers parody, fared much better with a score of 75/100. In April 2005 the intellectual property and technology rights were sold to various publishers and the studio was closed. === Neuroscience research === Following Elixir Studios, Hassabis returned to academia to obtain his PhD in cognitive neuroscience from UCL Queen Square Institute of Neurology in 2009 supervised by Eleanor Maguire. He sought to find inspiration in the human brain for new AI algorithms. He continued his neuroscience and artificial intelligence research as a visiting scientist jointly at Massachusetts Institute of Technology (MIT), in the lab of Tomaso Poggio, and Harvard University, before earning a Henry Wellcome postdoctoral research fellowship to the Gatsby Computational Neuroscience Unit at UCL in 2009 working with Peter Dayan. Working in the field of imagination, memory, and amnesia, he co-authored several influential papers published in Nature, Science, Neuron, and PNAS. His very first academic work, published in PNAS, was a landmark paper that showed systematically for the first time that patients with damage to their hippocampus, known to cause amnesia, were also unable to imagine themselves in new experiences. The finding established a link between the constructive process of imagination and the reconstructive process of episodic memory recall. Based on this work and a follow-up functional magnetic resonance imaging (fMRI) study, Hassabis developed a new theoretical account of the episodic memory system identifying scene construction, the generation and online maintenance of a complex and coherent scene, as a key process underlying both memory recall and imagination. This work received widespread coverage in the mainstream media and was listed in the top 10 scientific breakthroughs of the year by the journal Science. He later generalised these ideas to advance the notion of a 'simulation engine of the mind' whose role it was to imagine events and scenarios to aid with better planning. === DeepMind === Hassabis is the CEO and co-founder of DeepMind, a machine learning AI startup, founded in London in 2010 with Shane Legg and Mustafa Suleyman. Hassabis met Legg when both were postdocs at the Gatsby Computational Neuroscience Unit, and he and Suleyman had been friends through family. Hassabis also recruited his university friend and Elixir partner David Silver. DeepMind's mission is to "solve intelligence" and then use intelligence "to solve everything else". More concretely, DeepMind aims to combine insights from systems neuroscience with new developments in machine learning and computing hardware to unlock increasingly powerful general-purpose learning algorithms that will work towards the creation of an artificial general intelligence (AGI). The company has focused on training learning algorithms to master games, and in December 2013 it announced that it had made a pioneering breakthrough by training an algorithm called a Deep Q-Network (DQN) to play Atari games at a superhuman level by using only the raw pixels on the screen as inputs. DeepMind's early investors included several high-profile tech entrepreneurs. In 2014, Google purchased DeepMind for £400 million. Although most of the company has remained an independent entity based in London, DeepMind Health has since been directly incorporated into Google Health. Since the Google acquisition, the company has notched up a number of significant achievements, perhaps the most notable being the creation of AlphaGo, a program that defeated world champion Lee Sedol at the complex game of Go. Go had been considered a holy grail of AI, for its high number of possible board positions and resistance to existing programming techniques. However, AlphaGo beat European champion Fan Hui 5–0 in October 2015 before winning 4–1 against former world champion Lee Sedol in March 2016 and winning 3–0 against the world's top-ranked player Ke Jie in 2017. Additional DeepMind accomplishments include creating a neural Turing machine, reducing the energy used by the cooling systems in Google's data centres by 40%, and advancing research on AI safety. DeepMind has also been responsible for technical advances in machine learning, having produced a number of award-winning papers. In particular, the company has made significant advances in deep learning and reinforcement learning, and pioneered the field of deep reinforcement learning which combines these two methods. Hassabis has predicted that artificial intelligence will be "one of the most beneficial techn
National Security Commission on Artificial Intelligence
The National Security Commission on Artificial Intelligence (NSCAI) was an independent commission of the United States of America from 2018 to 2021. Its mission was to make recommendations to the President and Congress to "advance the development of artificial intelligence, machine learning, and associated technologies to comprehensively address the national security and defense needs of the United States". The commission's 15 members were nominated by the United States Congress. The NSCAI was dissolved on 1 October 2021. == History and reporting == The NSCAI began working in March 2019 and by November 2019 it had received more than 200 classified and unclassified briefings to help with the creation of its final report due in 2021.On 4 November 2019, the NSCAI shared its interim report with Congress, where it explained the 27 initial judgements to base its ongoing work. In the interim report the commission also agreed on seven principles: Global leadership in AI technology is a national security priority AI adoption is an urgent imperative for national security A shared sense of responsibility for the American peoples security must be created from government officials and private sector leaders. It needs to find local AI talent and use it to attract the world’s best minds Actions used for the protection of America’s AI leadership against foreign threats needs to follow the principles of free enterprise, free inquiry and free flow of ideas. The technical limitations of AI are universally known, however, a strong desire remains for powerful, dependable, and secure AI systems. United States used AI must follow American values including the rule of law Fundamental areas of effort for the preservation of U.S. advantages were also agreed upon in the interim report of 2019. The NSCAI released its first report of recommendations in March 2020, most of which were included in the 2021 National Defense Authorization Act. In July 2020, the commission published the second report to Congress. It identified 35 actions for both Executive and Legislative branches, which were focused on six fundamental areas. This report was available to the public. In January 2021, a draft of the final report was presented at a panel led by Schmidt. The report recommended the US to use AI technology for military use and development. It issued its final report in March 2021, saying that the U.S. is not sufficiently prepared to defend or compete against China in the AI era. It was broken up into two parts, the first titled “Defending America in the AI Era”, and the second “Winning the Technology Competition”. The report spoke about China’s efforts and investments into integration and that it could very well take the lead in AI in the next few years. Additional suggestions were made to concentrate on AI in everything we do and to implement it into US national security on multiple levels, as well as focus on bringing in new talent to develop AI and to introduce it to the working force on both civilian and military levels. Another recommendation of the NSCAI report was to develop and provide China and Russia with alternative models that are based on norms and democratic values. The final report also included a proposed $40 billion budget for government spending. On 14 April 2021, NSCAI executive director Ylli Bajraktari and director of Research and Analysis Justin Lynch participated in an event held by the Center for Security and Emerging Technology (CSET) to discuss the final report findings. In October 2021, NSCAI chair Eric Schmidt founded the bipartisan, non-profit Special Competitive Studies Project (SCSP) through his family led non-profit Eric & Wendy Schmidt Fund for Strategic Innovation in order to carry on the NSCAI’s efforts and expand beyond national security. The Foundation for Defense of Democracies held an event in June 2023, called “Thinking Forward After the NSCAI and CSC: A Discussion on AI and Cyber Policy”, with former members of NSCAI on the moderation panel, including Eric Schmidt and Ylli Bajraktari. == Members == Members of the National Security Commission on Artificial Intelligence: Eric Schmidt (chair), former CEO of Google Robert Work (Vice Chair), former Deputy Secretary of Defense Mignon Clyburn, former Commissioner of the Federal Communications Commission Chris Darby, CEO of In-Q-Tel Kenneth M. Ford, CEO of the Florida Institute for Human and Machine Cognition Jose-Marie Griffiths, President of Dakota State University Eric Horvitz, Technical Fellow at Microsoft Katrina G. McFarland, former Assistant Secretary of Defense for Acquisition Jason Matheny, Director of the Center for Security and Emerging Technology at Georgetown University Gilman Louie, partner at Alsop Louie Partners William Mark, vice president at SRI International Andy Jassy, CEO of Amazon Web Services (AWS) Safra Catz, CEO of Oracle Steve Chien, Technical Fellow at Jet Propulsion Laboratory (JPL) Andrew Moore, Google/Alphabet == Recommendations == The report's recommendations include: Dramatically increasing non-defense federal spending on AI research and development, doubling every year from $2 billion in 2022, to $32 billion in 2026. That would bring it up to a level similar to spending on biomedical research A dramatic increase in undergraduate scholarship and graduate studies fellowships in AI Creation of a Digital Corps to bring skilled tech workers into government Founding of a Digital Service Academy: an accredited university providing subsidized education in exchange for a commitment to work for a time in government Include civil rights and civil liberty reports for new AI systems or major updates to existing systems Expanding allocations of employment-based green cards, and giving them to every AI PhD graduate from an accredited U.S. university Reforming the acquisition management system Department of Defense to make it faster and easier to introduce new technologies == Transparency == In December 2019, a ruling was made under the Freedom of Information Act (FOIA) that the NSCAI must also provide historical documents upon request. The Electronic Privacy Information Center (EPIC) filed the lawsuit against the NSCAI in September 2019 after being refused information about the upcoming meetings and prepared records of the commission under FOIA and the Federal Advisory Committee Act (FACA). The U.S. District Court for the District of Columbia ruled in June 2020 that the NSCAI must comply with FACA and therefore hold open meetings and provide records to the public. The lawsuit was also filed by EPIC.
Weight initialization
In deep learning, weight initialization or parameter initialization describes the initial step in creating a neural network. A neural network contains trainable parameters that are modified during training: weight initialization is the pre-training step of assigning initial values to these parameters. The choice of weight initialization method affects the speed of convergence, the scale of neural activation within the network, the scale of gradient signals during backpropagation, and the quality of the final model. Proper initialization is necessary for avoiding issues such as vanishing and exploding gradients and activation function saturation. Note that even though this article is titled "weight initialization", both weights and biases are used in a neural network as trainable parameters, so this article describes how both of these are initialized. Similarly, trainable parameters in convolutional neural networks (CNNs) are called kernels and biases, and this article also describes these. == Constant initialization == We discuss the main methods of initialization in the context of a multilayer perceptron (MLP). Specific strategies for initializing other network architectures are discussed in later sections. For an MLP, there are only two kinds of trainable parameters, called weights and biases. Each layer l {\displaystyle l} contains a weight matrix W ( l ) ∈ R n l − 1 × n l {\displaystyle W^{(l)}\in \mathbb {R} ^{n_{l-1}\times n_{l}}} and a bias vector b ( l ) ∈ R n l {\displaystyle b^{(l)}\in \mathbb {R} ^{n_{l}}} , where n l {\displaystyle n_{l}} is the number of neurons in that layer. A weight initialization method is an algorithm for setting the initial values for W ( l ) , b ( l ) {\displaystyle W^{(l)},b^{(l)}} for each layer l {\displaystyle l} . The simplest form is zero initialization: W ( l ) = 0 , b ( l ) = 0 {\displaystyle W^{(l)}=0,b^{(l)}=0} Zero initialization is usually used for initializing biases, but it is not used for initializing weights, as it leads to symmetry in the network, causing all neurons to learn the same features. In this page, we assume b = 0 {\displaystyle b=0} unless otherwise stated. Recurrent neural networks typically use activation functions with bounded range, such as sigmoid and tanh, since unbounded activation may cause exploding values. (Le, Jaitly, Hinton, 2015) suggested initializing weights in the recurrent parts of the network to identity and zero bias, similar to the idea of residual connections and LSTM with no forget gate. In most cases, the biases are initialized to zero, though some situations can use a nonzero initialization. For example, in multiplicative units, such as the forget gate of LSTM, the bias can be initialized to 1 to allow good gradient signal through the gate. For neurons with ReLU activation, one can initialize the bias to a small positive value like 0.1, so that the gradient is likely nonzero at initialization, avoiding the dying ReLU problem. == Random initialization == Random initialization means sampling the weights from a normal distribution or a uniform distribution, usually independently. === LeCun initialization === LeCun initialization, popularized in (LeCun et al., 1998), is designed to preserve the variance of neural activations during the forward pass. It samples each entry in W ( l ) {\displaystyle W^{(l)}} independently from a distribution with mean 0 and variance 1 / n l − 1 {\displaystyle 1/n_{l-1}} . For example, if the distribution is a continuous uniform distribution, then the distribution is U ( ± 3 / n l − 1 ) {\displaystyle {\mathcal {U}}(\pm {\sqrt {3/n_{l-1}}})} . === Glorot initialization === Glorot initialization (or Xavier initialization) was proposed by Xavier Glorot and Yoshua Bengio. It was designed as a compromise between two goals: to preserve activation variance during the forward pass and to preserve gradient variance during the backward pass. For uniform initialization, it samples each entry in W ( l ) {\displaystyle W^{(l)}} independently and identically from U ( ± 6 / ( n l + 1 + n l − 1 ) ) {\displaystyle {\mathcal {U}}(\pm {\sqrt {6/(n_{l+1}+n_{l-1})}})} . In the context, n l − 1 {\displaystyle n_{l-1}} is also called the "fan-in", and n l + 1 {\displaystyle n_{l+1}} the "fan-out". When the fan-in and fan-out are equal, then Glorot initialization is the same as LeCun initialization. === He initialization === As Glorot initialization performs poorly for ReLU activation, He initialization (or Kaiming initialization) was proposed by Kaiming He et al. for networks with ReLU activation. It samples each entry in W ( l ) {\displaystyle W^{(l)}} from N ( 0 , 2 / n l − 1 ) {\displaystyle {\mathcal {N}}(0,2/n_{l-1})} . === Orthogonal initialization === (Saxe et al. 2013) proposed orthogonal initialization: initializing weight matrices as uniformly random (according to the Haar measure) semi-orthogonal matrices, multiplied by a factor that depends on the activation function of the layer. It was designed so that if one initializes a deep linear network this way, then its training time until convergence is independent of depth. Sampling a uniformly random semi-orthogonal matrix can be done by initializing X {\displaystyle X} by IID sampling its entries from a standard normal distribution, then calculate ( X X ⊤ ) − 1 / 2 X {\displaystyle \left(XX^{\top }\right)^{-1/2}X} or its transpose, depending on whether X {\displaystyle X} is tall or wide. For CNN kernels with odd widths and heights, orthogonal initialization is done this way: initialize the central point by a semi-orthogonal matrix, and fill the other entries with zero. As an illustration, a kernel K {\displaystyle K} of shape 3 × 3 × c × c ′ {\displaystyle 3\times 3\times c\times c'} is initialized by filling K [ 2 , 2 , : , : ] {\displaystyle K[2,2,:,:]} with the entries of a random semi-orthogonal matrix of shape c × c ′ {\displaystyle c\times c'} , and the other entries with zero. (Balduzzi et al., 2017) used it with stride 1 and zero-padding. This is sometimes called the Orthogonal Delta initialization. Related to this approach, unitary initialization proposes to parameterize the weight matrices to be unitary matrices, with the result that at initialization they are random unitary matrices (and throughout training, they remain unitary). This is found to improve long-sequence modelling in LSTM. Orthogonal initialization has been generalized to layer-sequential unit-variance (LSUV) initialization. It is a data-dependent initialization method, and can be used in convolutional neural networks. It first initializes weights of each convolution or fully connected layer with orthonormal matrices. Then, proceeding from the first to the last layer, it runs a forward pass on a random minibatch, and divides the layer's weights by the standard deviation of its output, so that its output has variance approximately 1. === Fixup initialization === In 2015, the introduction of residual connections allowed very deep neural networks to be trained, much deeper than the ~20 layers of the previous state of the art (such as the VGG-19). Residual connections gave rise to their own weight initialization problems and strategies. These are sometimes called "normalization-free" methods, since using residual connection could stabilize the training of a deep neural network so much that normalizations become unnecessary. Fixup initialization is designed specifically for networks with residual connections and without batch normalization, as follows: Initialize the classification layer and the last layer of each residual branch to 0. Initialize every other layer using a standard method (such as He initialization), and scale only the weight layers inside residual branches by L − 1 2 m − 2 {\displaystyle L^{-{\frac {1}{2m-2}}}} . Add a scalar multiplier (initialized at 1) in every branch and a scalar bias (initialized at 0) before each convolution, linear, and element-wise activation layer. Similarly, T-Fixup initialization is designed for Transformers without layer normalization. === Others === Instead of initializing all weights with random values on the order of O ( 1 / n ) {\displaystyle O(1/{\sqrt {n}})} , sparse initialization initialized only a small subset of the weights with larger random values, and the other weights zero, so that the total variance is still on the order of O ( 1 ) {\displaystyle O(1)} . Random walk initialization was designed for MLP so that during backpropagation, the L2 norm of gradient at each layer performs an unbiased random walk as one moves from the last layer to the first. Looks linear initialization was designed to allow the neural network to behave like a deep linear network at initialization, since W R e L U ( x ) − W R e L U ( − x ) = W x {\displaystyle W\;\mathrm {ReLU} (x)-W\;\mathrm {ReLU} (-x)=Wx} . It initializes a matrix W {\displaystyle W} of shape R n 2 × m {\displaystyle \mathbb {R} ^{{\frac {n}{2}}\times m}} by any method, such as orthogonal initialization, t
Comet (browser)
Comet is an AI-powered web browser based on Chromium. It was released by Perplexity AI for Microsoft Windows and macOS on July 9, 2025, for Android on November 20, 2025, and for iOS on March 18, 2026. Initial access to the browser was limited to users subscribed to Perplexity's most expensive tier, with broader availability expected over time. The browser was released for free download in October 2025. == Features == Comet is integrated with Perplexity's AI-assisted search engine. The browser features an assistant which enables users to perform a variety of tasks such as generating article summaries, sending emails, or buying products. == Security concerns == Researchers at LayerX Security identified a malicious attack vector which they call CometJacking. The exploit could possibly exfiltrate a user's personal sensitive data to a remote server controlled by the attacker. LayerX attempted to responsibly disclose their findings to Comet's developer Perplexity AI in August 2025. Perplexity responded that they saw no security impact and marked the disclosure as not applicable.