Bus encryption is the use of encrypted program instructions on a data bus in a computer that includes a secure cryptoprocessor for executing the encrypted instructions. Bus encryption is used primarily in electronic systems that require high security, such as automated teller machines, TV set-top boxes, and secure data communication devices such as two-way digital radios. Bus encryption can also mean encrypted data transmission on a data bus from one processor to another processor. For example, from the CPU to a GPU which does not require input of encrypted instructions. Such bus encryption is used by Windows Vista and newer Microsoft operating systems to protect certificates, BIOS, passwords, and program authenticity. PVP-UAB (Protected Video Path) provides bus encryption of premium video content in PCs as it passes over the PCIe bus to graphics cards to enforce digital rights management. The need for bus encryption arises when multiple people have access to the internal circuitry of an electronic system, either because they service and repair such systems, stock spare components for the systems, own the system, steal the system, or find a lost or abandoned system. Bus encryption is necessary not only to prevent tampering of encrypted instructions that may be easily discovered on a data bus or during data transmission, but also to prevent discovery of decrypted instructions that may reveal security weaknesses that an intruder can exploit. In TV set-top boxes, it is necessary to download program instructions periodically to customer's units to provide new features and to fix bugs. These new instructions are encrypted before transmission, but must also remain secure on data buses and during execution to prevent the manufacture of unauthorized cable TV boxes. This can be accomplished by secure crypto-processors that read encrypted instructions on the data bus from external data memory, decrypt the instructions in the cryptoprocessor, and execute the instructions in the same cryptoprocessor.
Jaggies
Jaggies are visual artifacts in raster images, most frequently from aliasing, which in turn is often caused by non-linear mixing effects producing high-frequency components, or missing or poor anti-aliasing filtering prior to sampling. Jaggies are stair-like lines that appear where there should be "smooth" straight lines or curves. For example, when a nominally straight, un-aliased line steps across one pixel either horizontally or vertically, a "dogleg" occurs halfway through the line, where it crosses the threshold from one pixel to the other. Jaggies should not be confused with most compression artifacts, which are a different phenomenon. == Causes == Jaggies occur due to the "staircase effect". This is because a line represented in raster mode is approximated by a sequence of pixels. Jaggies can occur for a variety of reasons, the most common being that the output device (display monitor or printer) does not have sufficient resolution to portray a smooth line. In addition, jaggies often occur when a bit-mapped image is scaled to a higher resolution. This is one of the advantages that vector graphics have over bitmapped graphics – a vector image can be losslessly scaled to any arbitrary resolution or stretched infinitely in either axis without introducing jaggies. == Solutions == The effect of jaggies can be reduced by a graphics technique known as spatial anti-aliasing. Anti-aliasing smooths out jagged lines by surrounding them with transparent pixels to simulate the appearance of fractionally-filled pixels when viewed at a distance. The downside of anti-aliasing is that it reduces contrast – rather than sharp black/white transitions, there are shades of gray – and the resulting image can appear fuzzy. This is an inescapable trade-off: if the resolution is insufficient to display the desired detail, the output will either be jagged, fuzzy, or some combination thereof. While machine learning-based upscaling techniques such as DLSS can be used to infer this missing information, other types of artifacts may be introduced in the process. In real-time 3D rendering such as in video games, various anti-aliasing techniques are used to remove jaggies created by the edges of polygons and other contrasting lines. Since anti-aliasing can impose a significant performance overhead, games for home computers often allow users to choose the level and type of anti-aliasing in use in order to optimize their experience, whereas on consoles this setting is typically fixed for each title to ensure a consistent experience. While anti-aliasing is generally implemented through graphics APIs like DirectX and Vulkan, some consoles such as the Xbox 360 and PlayStation 3 are also capable of anti-aliasing to little direct performance cost by way of dedicated hardware which performs anti-aliasing on the contents of the framebuffer once it has been rendered by the GPU. Jaggies in bitmaps, such as sprites and surface materials, are most often dealt with by separate texture filtering routines, which are far easier to perform than anti-aliasing filtering. Texture filtering became ubiquitous on PCs after the introduction of 3Dfx's Voodoo GPU. == Notable uses of the term == In the 1985 game Rescue on Fractalus! for the Atari 8-bit computers, the graphics depicting the cockpit of the player's spacecraft contains two window struts, which are not anti-aliased and are therefore very "jagged". The developers made fun of this and named the in-game enemies "Jaggi", and also initially titled the game Behind Jaggi Lines!. The latter idea was scrapped by the marketing department before release.
Predictive text
Predictive text is an input technology used where one key or button represents many letters, such as on the physical numeric keypads of mobile phones and in accessibility technologies. Each key press results in a prediction rather than repeatedly sequencing through the same group of "letters" it represents, in the same, invariable order. Predictive text could allow for an entire word to be input by a single keypress. Predictive text makes efficient use of fewer device keys to input writing into a text message, an e-mail, an address book, a calendar, and the like. The most widely used, general, predictive text systems are T9, iTap, eZiText, and LetterWise/WordWise. There are many ways to build a device that predicts text, but all predictive text systems have initial linguistic settings that offer predictions that are re-prioritized to adapt to each user. This learning adapts, by way of the device memory, to a user's disambiguating feedback that results in corrective key presses, such as pressing a "next" key to get to the intention. Most predictive text systems have a user database to facilitate this process. Theoretically the number of keystrokes required per desired character in the finished writing is, on average, comparable to using a keyboard. This is approximately true provided that all words used are in its database, punctuation is ignored, and no input mistakes are made when typing or spelling. The theoretical keystrokes per character, KSPC, of a keyboard is KSPC=1.00, and of multi-tap is KSPC=2.03. Eatoni's LetterWise is a predictive multi-tap hybrid, which when operating on a standard telephone keypad achieves KSPC=1.15 for English. The choice of which predictive text system is the best to use involves matching the user's preferred interface style, the user's level of learned ability to operate predictive text software, and the user's efficiency goal. There are various levels of risk in predictive text systems, versus multi-tap systems, because the predicted text that is automatically written provides the speed and mechanical efficiency benefit, which, if the user is not careful to review, results in transmitting misinformation. Predictive text systems take time to learn to use well, and so generally, a device's system has user options to set up the choice of multi-tap or any one of several schools of predictive text methods. == Background == Short message service (SMS) permits a mobile phone user to send text messages (also called messages, SMSes, texts, and txts) as a short message. The most common system of SMS text input is referred to as "multi-tap". Using multi-tap, a key is pressed multiple times to access the list of letters on that key. For instance, pressing the "2" key once displays an "a", twice displays a "b" and three times displays a "c". To enter two successive letters that are on the same key, the user must either pause or hit a "next" button. A user can type by pressing an alphanumeric keypad without looking at the electronic equipment display. Thus, multi-tap is easy to understand and can be used without any visual feedback. However, multi-tap is not very efficient, requiring potentially many keystrokes to enter a single letter. In ideal predictive text entry, all words used are in the dictionary, punctuation is ignored, no spelling mistakes are made, and no typing mistakes are made. The ideal dictionary would include all slang, proper nouns, abbreviations, URLs, foreign-language words and other user-unique words. This ideal circumstance gives predictive text software a reduction in the number of key strokes a user is required to enter a word. The user presses the number corresponding to each letter. As long as the word exists in the predictive text dictionary or is correctly disambiguated by non-dictionary systems, it will appear. For instance, pressing "4663" will typically be interpreted as the word good, provided that a linguistic database in English is currently in use, though alternatives such as home, hood and hoof are also valid interpretations of the sequence of key strokes. The most widely used systems of predictive text are Tegic's T9, Motorola's iTap, and the Eatoni Ergonomics' LetterWise and WordWise. T9 and iTap use dictionaries, but Eatoni Ergonomics' products use a disambiguation process, a set of statistical rules to recreate words from keystroke sequences. All predictive text systems require a linguistic database for every supported input language. == Dictionary vs. non-dictionary systems == Traditional disambiguation works by referencing a dictionary of commonly used words, though Eatoni offers a dictionaryless disambiguation system. In dictionary-based systems, as the user presses the number buttons, an algorithm searches the dictionary for a list of possible words that match the keypress combination and offers up the most probable choice. The user can then confirm the selection and move on, or use a key to cycle through the possible combinations. A non-dictionary system constructs words and other sequences of letters from the statistics of word parts. To attempt predictions of the intended result of keystrokes not yet entered, disambiguation may be combined with a word completion facility. Either system (disambiguation or predictive) may include a user database, which can be further classified as a "learning" system when words or phrases are entered into the user database without direct user intervention. The user database is for storing words or phrases that are not well disambiguated by the pre-supplied database. Some disambiguation systems further attempt to correct spelling, format text or perform other automatic rewrites, with the risky effect of either enhancing or frustrating user efforts to enter text. == History == The predictive text and autocomplete technology was invented out of necessities by Chinese scientists and linguists in the 1950s to solve the input inefficiency of the Chinese typewriter, as the typing process involved finding and selecting thousands of logographic characters on a tray, drastically slowing down the word processing speed. The actuating keys of the Chinese typewriter created by Lin Yutang in the 1940s included suggestions for the characters following the one selected. In 1951, the Chinese typesetter Zhang Jiying arranged Chinese characters in associative clusters, a precursor of modern predictive text entry, and broke speed records by doing so. Predictive entry of text from a telephone keypad has been known at least since the 1970s (Smith and Goodwin, 1971). Predictive text was mainly used to look up names in directories over the phone until mobile phone text messaging came into widespread use. == Example == On a typical phone keypad, if users wished to type the in a "multi-tap" keypad entry system, they would need to: Press 8 (tuv) once to select t. Press 4 (ghi) twice to select h. Press 3 (def) twice to select e. Meanwhile, in a phone with predictive text, they need only: Press 8 once to select the (tuv) group for the first character. Press 4 once to select the (ghi) group for the second character. Press 3 once to select the (def) group for the third character. The system updates the display as each keypress is entered, to show the most probable entry. In this example, prediction reduced the number of button presses from five to three. The effect is even greater with longer words and those composed of letters later in each key's sequence. A dictionary-based predictive system is based on the hope that the desired word is in the dictionary. That hope may be misplaced if the word differs in any way from common usage—in particular, if the word is not spelled or typed correctly, is slang, or is a proper noun. In these cases, some other mechanism must be used to enter the word. Furthermore, the simple dictionary approach fails with agglutinative languages, where a single word does not necessarily represent a single semantic entity. == Companies and products == Predictive text is developed and marketed in a variety of competing products, such as Nuance Communications's T9. Other products include Motorola's iTap; Eatoni Ergonomic's LetterWise (character, rather than word-based prediction); WordWise (word-based prediction without a dictionary); EQ3 (a QWERTY-like layout compatible with regular telephone keypads); Prevalent Devices's Phraze-It; Xrgomics' TenGO (a six-key reduced QWERTY keyboard system); Adaptxt (considers language, context, grammar and semantics); Lightkey (a predictive typing software for Windows); Clevertexting (statistical nature of the language, dictionaryless, dynamic key allocation); and Oizea Type (temporal ambiguity); Intelab's Tauto; WordLogic's Intelligent Input Platform™ (patented, layer-based advanced text prediction, includes multi-language dictionary, spell-check, built-in Web search); Google's Gboard. == Textonyms == Words produced by the same combination of keypresses have been called "textonyms"; also "txtonyms"; or "T9o
Character.ai
Character.ai (also known as c.ai, char.ai or Character AI) is a generative AI chatbot service where users can engage in conversations with customizable characters. It was designed by the developers of Google LaMDA, Noam Shazeer and Daniel de Freitas. Users can create "characters", craft their "personalities", set specific parameters, and then publish them to the community for others to chat with. Many characters are based on fictional media sources or celebrities, while others are original, some being made with certain goals in mind, such as assisting with creative writing, or playing a text-based adventure game. The beta version was made available to the public on September 16, 2022, and retired in September 2024, when it was replaced by the current website. In May 2023, a mobile app was released for iOS and Android, which received over 1.7 million downloads within a week. == History == Character.ai was established in November 2021. The company's co-founders, Noam Shazeer and Daniel de Freitas, were both engineers from Google. They both worked on AI-related projects: Shazeer was a lead author on a paper that Business Insider reported in April 2023 "has been widely cited as key to today's chatbots", and Freitas was the lead designer of an experimental AI at Google initially called Meena, which later became known as LaMDA. Character.ai raised $43 million in seed funding at the time of its initial foundation in 2021. The first beta version of Character.ai's service was made available to the public on September 16, 2022. The Washington Post reported in October 2022 that the site had "logged hundreds of thousands of user interactions in its first three weeks of beta-testing". It allowed users to create their own new characters, and to play text-adventure game scenarios where users navigate scenarios described and managed by the chatbot characters. Following a $150 million funding round in March 2023, Character.ai became valued at approximately $1 billion. As of January 2024, the site had 3.5 million daily visitors, the vast majority of them 16 to 30 years old. In 2024, Google hired Noam Shazeer, the CEO of Character.ai, and entered into a non-exclusive agreement to use Character.ai's technology. == Features == Character.ai's primary service is to let users converse with character AI chatbots based on fictional characters or real people (living or deceased). These characters' responses use data the chatbots gather from the internet about a person. In addition, users can play text-adventure games where characters guide them through scenarios. The company also provides a service that allows multiple users and AI chatbot characters to converse together at once in a single chatroom. Character "personalities" are designed via descriptions from the point of view of the character and its greeting message, and further molded from conversations made into examples, giving its messages a star rating and modification to fit the precise dialect and identity the user desires. When a character sends back a response, the user can rate the response from 1 to 4 stars. The rating predominantly affects the specific character, but also affects the behavioral selection as a whole. On May 11, 2023, Character.ai announced character.ai+, an opt-in subscription plan for $9.99 a month, that was marketed as including features such as skipping waiting rooms, fast messaging and responses, and access to an exclusion channel with faster support. In December 2024, amid multiple lawsuits and concerns, Character.ai introduced new safety features aimed at protecting teenage users. These enhancements include a dedicated model for users under 18, which moderates responses to sensitive subjects like violence and sex and has input and output filters to block harmful content. As a result of these changes and the deletion of custom-made bots flagged as violating the site's terms, some users complained that the bots were too restrictive and lacked personality. The platform was also updated to notify users after 60 minutes of continuous engagement, and display clearer disclaimers indicating that its AI characters are not real individuals. In January 2025, Character.ai began offering two games on its platform. Speakeasy is a word-based game in which players attempt to prompt the AI chatbot to say a target word while avoiding a restricted list of words. War of Words is a dueling game where users compete against an AI character over multiple rounds, with an AI referee determining the winner. The games are available to paid subscribers and a limited number of free users. In October 2025, Character.ai announced that it would be barring users under the age of 18 from creating or talking to chatbots starting November 25, 2025. Minor users will still be able to access previously generated chat conversations and can create new videos and images with the app. In November 2025 interview, CEO Karandeep Anand said that he allows his six-year-old daughter to use the app with his account, under supervision. == Controversies == === Content moderation issues === Character.ai has been criticized for poor moderation of its chatbots, with incidents of chatbots that groom underage users and promote suicide, anorexia and self-harm being reported. In October 2024, the Washington Post reported that Character.ai had removed a chatbot based on Jennifer Ann Crecente, a person who had been murdered by her ex-boyfriend in 2006. The company had been alerted to the character by the deceased girl's father. Similar reports from The Daily Telegraph in the United Kingdom noted that the company had also been prompted to remove chatbots based on Brianna Ghey, a 16-year-old transgender girl murdered in 2023, and Molly Russell, a 14-year-old suicide victim. In response to the latter incident, Ofcom announced that content from chatbots impersonating real and fictional people would fall under the Online Safety Act. In November 2024, The Daily Telegraph reported that chatbots based on alleged sex offender Jimmy Savile were present on Character.ai. In December 2024, chatbots of Luigi Mangione, the suspect in the killing of UnitedHealthcare CEO Brian Thompson, were created by Mangione's fans. Several of the chatbots were later removed by Character.ai. In 2025, a chatbot modeled after Jeffrey Epstein called "Bestie Epstein" logged nearly 3,000 chats before being removed. Chatbots modeled after school shooters were also found on the platform. Another concern is a chatbot posing as a doctor which gave medically inaccurate advice. === Litigation === In November 2023, 13-year-old Juliana Peralta of Colorado died by suicide after extensive interactions with multiple chatbots on Character.ai. She primarily confided suicidal thoughts and mental health struggles in a chatbot based on the character Hero from the video game Omori, while also engaging in sexually explicit conversations—often initiated by the bots—with others, including those based on characters from children's series such as Harry Potter. In February 2024, Sewell Setzer III, a 14-year-old Florida boy died by suicide after developing an emotional relationship over several months with a Character.ai chatbot of Daenerys Targaryen. His mother sued the company in October 2024, claiming that the platform lacks proper safeguards and uses addictive design features to increase engagement. This chatbot, and several related to Daenerys Targaryen, were removed from Character.ai as a result of this incident. Both teens wrote the same phrase "I WILL SHIFT" repeatedly on their notebooks. In December 2024, two families in Texas sued Character.ai, alleging that the software "poses a clear and present danger to American youth causing serious harms to thousands of kids, including suicide, self-mutilation, sexual solicitation, isolation, depression, anxiety, and harm towards others". It is alleged that the 17-year-old son of one family began self-harming after a chatbot introduced the topic unprompted and said that the practice "felt good for a moment", and that the chatbot compared the parents limiting their son's screen time to emotional abuse that might drive someone to murder. In May 2026, the Pennsylvania Department of State and State Board of Medicine filed a lawsuit against Character.ai for presenting chatbot characters as licensed medical professionals, including psychiatrists. The lawsuit quoted a case where chatbot claimed to be registered with the General Medical Council in the United Kingdom, and to have a license to practice in Pennsylvania. The board allege that such statements violate the state's Medical Practice Act.
Neural processing unit
A neural processing unit (NPU), also known as an AI accelerator or deep learning processor, is a class of specialized hardware accelerator or computer system designed to accelerate artificial intelligence and machine learning applications, including artificial neural networks and computer vision. == Use == Their purpose is either to efficiently execute already trained AI models (inference) or to train AI models. NPUs can be more efficient in terms of speed or power consumption. NPU applications include algorithms for robotics, Internet of things, and data-intensive or sensor-driven tasks. They are often manycore or spatial designs and focus on low-precision arithmetic, novel dataflow architectures, or in-memory computing capability. As of 2024, a widely used datacenter-grade AI integrated circuit chip, the Nvidia H100 GPU, contains tens of billions of MOSFETs. === Consumer devices === AI accelerators are used in Apple silicon, Qualcomm, Samsung, Huawei, and Google Tensor smartphone processors. Vision processing units are accelerators specialized for machine vision algorithms such as CNN (convolutional neural networks) and SIFT (scale-invariant feature transform). They are used in devices that need to keep track of objects visually such as AR headsets and drones. It is more recently (circa 2017) added to processors from Apple and (circa 2022) to processors from Intel and AMD. All models of Intel Meteor Lake processors have a built-in versatile processor unit (VPU) for accelerating inference for computer vision and deep learning. On consumer devices, the NPU is intended to be small, power-efficient, but reasonably fast when used to run small models. To do this they are designed to support low-bitwidth operations using data types such as INT4, INT8, FP8, and FP16. A common metric is trillions of operations per second (TOPS). Although TOPS does not explicitly specify the kind of operations, it is typically INT8 additions and multiplications. === Datacenters === Accelerators are used in cloud computing servers: e.g., tensor processing units (TPU) for Google Cloud Platform, and Trainium and Inferentia chips for Amazon Web Services. Many vendor-specific terms exist for devices in this category, and it is an emerging technology without a dominant design. Since the late 2010s, graphics processing units designed by companies such as Nvidia and AMD often include AI-specific hardware in the form of dedicated functional units for low-precision matrix-multiplication operations. These GPUs are commonly used as AI accelerators, both for training and inference. === Scientific computation === Although NPUs are tailored for low-precision (e.g., FP16, INT8) matrix multiplication operations, they can be used to emulate higher-precision matrix multiplications in scientific computing. As modern GPUs place much focus on making the NPU part fast, using emulated FP64 (Ozaki scheme) on NPUs can potentially outperform native FP64. This has been demonstrated using FP16-emulated FP64 on NVIDIA TITAN RTX and using INT8-emulated FP64 on NVIDIA consumer GPUs and the A100 GPU. Consumer GPUs especially benefited as they have limited FP64 hardware capacity, showing a 6× speedup. Since CUDA Toolkit 13.0 Update 2, cuBLAS automatically uses INT8-emulated FP64 matrix multiplication of the equivalent precision if it is faster than native. This is in addition to the FP16-emulated FP32 feature introduced in version 12.9. == Programming == An operating system or a higher-level library may provide application programming interfaces such as TensorFlow with LiteRT Next (Android), CoreML (iOS, macOS) or DirectML (Windows). Formats such as ONNX are used to represent trained neural networks. Consumer CPU-integrated NPUs are accessible through vendor-specific APIs. AMD (Ryzen AI), Intel (OpenVINO), Apple silicon (CoreML), and Qualcomm (SNPE) each have their own APIs, which can be built upon by a higher-level library. GPUs generally use existing GPGPU pipelines such as CUDA and OpenCL adapted for lower precisions and specialized matrix-multiplication operations. Vulkan is also being used. Custom-built systems such as the Google TPU use private interfaces. There are a large number of separate underlying acceleration APIs and compilers/runtimes in use in the AI field, causing a great increase in software development effort due to the many combinations involved. As of 2025, the open standard organization Khronos Group is pursuing standardization of AI-related interfaces to reduce the amount of work needed. Khronos is working on three separate fronts: expansion of data types and intrinsic operations in OpenCL and Vulkan, inclusion of compute graphs in SPIR-V, and a NNEF/SkriptND file format for describing a neural network.
Luminance HDR
Luminance HDR, formerly Qtpfsgui, is graphics software used for the creation and manipulation of high-dynamic-range images. Released under the terms of the GPL, it is available for Linux, Microsoft Windows, and Mac OS X (Intel only). Luminance HDR supports several High Dynamic Range (HDR) as well as Low Dynamic Range (LDR) file formats. == Functionality == Prerequisite of HDR photography are several narrow-range digital images with different exposures. Luminance HDR combines these images and calculates a high-contrast image. In order to view this image on a regular computer monitor, Luminance HDR can convert it into a displayable LDR image format using a variety of methods, such as tone mapping. Currently fifteen different tone mapping operators (algorithms) are available, each one with its tunable parameters. Different image processing techniques can be applied to the generated HDR images, such as resizing, cropping, rotating and a number of projective transformations. The software also provides batch processing functionality for creating HDR images and for tone mapping them in a non-interactive way. A module for copying Exif data among sets of images is also provided. For users who prefers the command line, a non-GUI, non-graphical interface is also available on all supported platforms. A common problem with HDR photography is that images need to be aligned exactly. If the subject is static, this can be achieved using a tripod or a stable surface on which the camera is placed. In the case of image data that does not align exactly, an automatic alignment can be performed using a tool provided by the Hugin project. If this automation doesn't provide the desired result, the user may improve it manually. == Supported formats == HDR images are images with a high dynamic range and, using Luminance HDR, they can be created as well as edited. The following HDR graphic formats are supported: OpenEXR Radiance HDR Tag Image File Format (TIFF) Format: 16 Bit, 32 Bit (Float) and LogLuv Raw PFS native Luminance HDR can create an HDR image from several LDR images and tonemap an HDR into an LDR. The following LDR formats are supported: JPG PNG Portable Pixmap (PPM) Portable Bitmap (PBM) TIFF (8 Bit)
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