Wiki surveys or wikisurveys are a software-based survey method that crowdsource discussions and help participants to find areas of agreement. Other names include bridging systems and collective response systems. The approach, inspired by Wikipedia, is to open up surveys where participants can shape the questions, instead of traditional 'closed' surveys where participants can only respond to the questions asked. Wiki surveys have been used for purposes including facilitating deliberative democracy, crowdsourcing opinions from experts and figuring out common beliefs on a given topic. A notable usage of wiki surveys is in Taiwan's government system, where citizens can participate in crowdsourced lawmaking through Pol.is wiki surveys. == Implementations == === All Our Ideas === All Our Ideas founders coined the term "wiki survey," explaining how they took inspiration from the organic evolution of Wikipedia and hoped to create something similar for surveys. They hosted 5000 surveys between 2010 and 2014. A 2020 survey using the tool found 3 of its top 10 findings were user-generated. === Decidim === Decidim has been used by governments throughout Spain and Europe to help with participatory budgeting and other public policy decisions. === Polis === Polis (also known as Pol.is) was developed in 2012. The focus of Polis is to project participants into an 'opinion space' where they can see how their voting behavior compares to other participants. The opinion space clusters participants into groups of similar opinion and is designed in a way to avoid tyranny of the majority by being able to include groups that have small numbers of participants. The questions participants are presented with are agree/disagree/pass on a single 'comment' submitted by a participant. The code for Polis is free and open-source software under the GNU AGPL. === Remesh === Remesh was founded in 2013 and has partnered with the United Nations and Alliance for Middle East Peace efforts to bring peaceful resolutions to conflicts. Participants are anonymous and the algorithm can be fine-tuned to better understand local dialects in specific regions. == Examples == PlaNYC used All Our Ideas to gather ideas on how to establish New York City's sustainability plan vTaiwan, a citizen-lead government process in Taiwan, uses Polis for enabling large amounts of citizens to deliberate and consequently provide input on Taiwan's legislative decisions OECD used All Our Ideas to gather ideas from the public prior to meeting for a forum and meeting on which skills are most important to invest in for the 21st century March On, an offshoot of the Women's March Movement, used Polis to understand the opinions of people wanting to support the movement Residents of Harrogate use Polis to debate issues in their community, with the results being released publicly to everyone == Characteristics == Wiki surveys often have these three characteristics: === Collaborativeness === Wiki surveys allow participants to contribute questions, as well as answer questions created by its participants. === Adaptivity === Wiki surveys adapt to elicit the most useful information from its participants. One example involves changing the ordering of questions based on the voting behavior of previous participants so as to maximize consensus. The heuristic determining the ordering of questions highly values showing the comments that have been voted on the least. === 'Greediness' === In the context of wiki surveys, 'greediness' simply means making full use of information that participants are willing to provide. Wiki surveys do not require participants to answer a fixed amount of questions, so participants can answer as little or as much as they want. This is intended to be more efficient in capturing participants' preferences by allowing more organic sharing of their perspectives. == Traditional survey methods vs. wiki surveys == Questions in traditional survey methods fall into two categories: Open and closed questions. Open questions ask the person taking the survey to write an open response while closed questions give a fixed set of responses to select from. Wiki surveys are like a hybrid of the two, enabling insightful consensus in certain situations where traditional survey methods may lack. Closed questions are easy to analyze quantitively, but the limited options to select from for a given question may cause bias. Open questions are not as subject to bias, but are difficult to analyze quantitatively at scale. Wiki surveys allow for open responses by the users' contribution of survey questions (also called 'items'), and uses machine learning techniques to (at least partially) automate the quantitative analysis of the responses to those questions.
Site-specific browser
A site-specific browser (SSB) is a software application dedicated to accessing pages from a single source (site) on a computer network such as the Internet or a private intranet. SSBs typically simplify the more complex functions of a web browser by excluding the menus, toolbars and browser graphical user interface associated with functions that are external to the workings of a single site. Modern site-specific browsers range from simple browser windows without navigation controls to sophisticated desktop applications built with frameworks like Electron that bundle entire browser engines. This evolution has enabled many popular desktop applications to be built using web technologies, effectively making them advanced site-specific browsers. == History == === Early development === One of the earliest examples of an SSB was MacDICT, a Mac OS 9 application that accessed various websites to define, translate, or find synonyms for words typed into a text box. However, the first general-purpose SSB is considered to be Bubbles, which launched in late 2005 on the Windows platform. Bubbles introduced the term "Site Specific Extensions" for SSB userscripts and created the first SSB JavaScript API. In 2007, Mozilla announced Prism (originally called WebRunner), a project to integrate web applications with the desktop. That same year, Todd Ditchendorf, a former Apple Dashboard engineer, released Fluid for macOS. On 2 September 2008, Google Chrome was released with a built-in "Create application shortcut" feature, bringing SSB functionality to mainstream users. This feature allowed any website to be launched in a separate window without the browser interface. === Modern era === The landscape of site-specific browsers changed dramatically with the introduction of Electron in 2013 (originally called Atom Shell). Electron combined Chromium and Node.js into a single runtime, enabling developers to build desktop applications using web technologies. This framework has since powered applications used by hundreds of millions of users, including Visual Studio Code, Slack, Discord, and Microsoft Teams. In 2015, the concept of Progressive Web Apps (PWAs) was introduced by Google engineers Alex Russell and Frances Berriman, representing a parallel evolution in web-to-desktop technology. While PWAs share similar goals with SSBs, they follow web standards and can be installed directly from browsers. More recently, alternative frameworks like Tauri have emerged, offering significantly smaller application sizes by using the system's native web renderer instead of bundling Chromium. == Technical implementation == Site-specific browsers can be implemented through various approaches: === Browser-based SSBs === The simplest form of SSB is created through browser features that allow websites to run in separate windows without the standard browser interface. Modern Chromium-based browsers offer "Install as app" or "Create shortcut" functionality that creates a dedicated window for a specific website. These SSBs share the browser's underlying engine and resources but operate in isolated windows. === Framework-based SSBs === More sophisticated SSBs are built using application frameworks: Electron: Bundles a complete Chromium browser with Node.js, resulting in applications of 85MB or larger. Each Electron application runs its own browser instance, providing full access to system APIs but consuming significant resources. Tauri: Uses the operating system's native web rendering engine (WebView2 on Windows, WebKit on macOS, and WebKitGTK on Linux), resulting in applications typically 2.5-10MB in size. Other frameworks: Include Neutralino.js (ultra-lightweight using system browser), Wails (Go-based), and the Chromium Embedded Framework (CEF). == Comparison with Progressive Web Apps == While site-specific browsers and Progressive Web Apps (PWAs) share the goal of bringing web content to the desktop, they differ in several key aspects: == Applications == Site-specific browsers have become the foundation for many popular desktop applications: Communication and collaboration: Many modern communication tools are built as SSBs, including Slack, Discord, Microsoft Teams, and WhatsApp Desktop. These applications benefit from web-based development while providing desktop integration. Development tools: Visual Studio Code, used by 73.6% of developers according to Stack Overflow's 2024 survey, is built with Electron, as are Atom and GitHub Desktop. Productivity software: Applications like Notion, Obsidian, and various project management tools use SSB technology to provide consistent experiences across platforms. Security and Privacy: Web browsers can be modified to only have access to a single site, in order to protect the security and privacy of the user via compartmentalization == Security and performance == === Memory usage === Framework-based SSBs, particularly those using Electron, are known for high memory consumption. Studies show Electron applications typically use 120-300MB at baseline, with complex applications consuming significantly more. This is approximately 5-10 times more memory than equivalent native applications. === Security considerations === SSBs can provide security benefits through process isolation, where each application runs in its own sandboxed environment. However, bundling an entire browser engine also means each application must be updated independently to patch security vulnerabilities. Research presented at the Network and Distributed System Security (NDSS) Symposium has identified various security challenges specific to Electron applications. === Bundle sizes === The choice of framework significantly impacts application size: Electron applications: 85MB+ (includes full Chromium) Tauri applications: 2.5-10MB (uses system WebView) Browser-based SSBs: No additional download (uses existing browser) == Software == === Browser support === Most modern browsers provide some form of SSB functionality: Chromium-based browsers (Google Chrome, Microsoft Edge, Brave, Opera, Vivaldi): "Install as app" or "Create shortcut" feature Safari: "Add to Dock" feature in macOS Sonoma (2023) Firefox: Removed SSB support in December 2020 (version 85) GNOME Web: "Install Site as Web Application" feature === Standalone tools === ==== Active ==== WebCatalog (Windows, macOS, Linux) – Manages multiple SSBs with isolated storage Fluid (macOS) – Pioneering SSB creator for Mac Unite (macOS) – Creates SSBs with customization options Coherence X (macOS) – Advanced SSB creation tool Pake (cross-platform) – Open-source SSB creator Wavebox (cross-platform) – Workspace browser with SSB features ==== Discontinued ==== Mozilla Prism – Cross-platform SSB creator (discontinued 2011) Nativefier – Command-line SSB creator (discontinued 2023) Epichrome – macOS SSB creator (discontinued 2021) === Development frameworks === Electron – Most popular framework, bundles Chromium and Node.js Tauri – Rust-based framework using system WebView Chromium Embedded Framework (CEF) – C++ library for embedding Chromium Neutralino.js – Lightweight framework using system browser Wails – Go-based framework for web frontends
Neural architecture search
Neural architecture search (NAS) is a technique for automating the design of artificial neural networks (ANN), a widely used model in the field of machine learning. NAS has been used to design networks that are on par with or outperform hand-designed architectures. Methods for NAS can be categorized according to the search space, search strategy and performance estimation strategy used: The search space defines the type(s) of ANN that can be designed and optimized. The search strategy defines the approach used to explore the search space. The performance estimation strategy evaluates the performance of a possible ANN from its design (without constructing and training it). NAS is closely related to hyperparameter optimization and meta-learning and is a subfield of automated machine learning (AutoML). == Reinforcement learning == Reinforcement learning (RL) can underpin a NAS search strategy. Barret Zoph and Quoc Viet Le applied NAS with RL targeting the CIFAR-10 dataset and achieved a network architecture that rivals the best manually-designed architecture for accuracy, with an error rate of 3.65, 0.09 percent better and 1.05x faster than a related hand-designed model. On the Penn Treebank dataset, that model composed a recurrent cell that outperforms LSTM, reaching a test set perplexity of 62.4, or 3.6 perplexity better than the prior leading system. On the PTB character language modeling task it achieved bits per character of 1.214. Learning a model architecture directly on a large dataset can be a lengthy process. NASNet addressed this issue by transferring a building block designed for a small dataset to a larger dataset. The design was constrained to use two types of convolutional cells to return feature maps that serve two main functions when convoluting an input feature map: normal cells that return maps of the same extent (height and width) and reduction cells in which the returned feature map height and width is reduced by a factor of two. For the reduction cell, the initial operation applied to the cell's inputs uses a stride of two (to reduce the height and width). The learned aspect of the design included elements such as which lower layer(s) each higher layer took as input, the transformations applied at that layer and to merge multiple outputs at each layer. In the studied example, the best convolutional layer (or "cell") was designed for the CIFAR-10 dataset and then applied to the ImageNet dataset by stacking copies of this cell, each with its own parameters. The approach yielded accuracy of 82.7% top-1 and 96.2% top-5. This exceeded the best human-invented architectures at a cost of 9 billion fewer FLOPS—a reduction of 28%. The system continued to exceed the manually-designed alternative at varying computation levels. The image features learned from image classification can be transferred to other computer vision problems. E.g., for object detection, the learned cells integrated with the Faster-RCNN framework improved performance by 4.0% on the COCO dataset. In the so-called Efficient Neural Architecture Search (ENAS), a controller discovers architectures by learning to search for an optimal subgraph within a large graph. The controller is trained with policy gradient to select a subgraph that maximizes the validation set's expected reward. The model corresponding to the subgraph is trained to minimize a canonical cross entropy loss. Multiple child models share parameters, ENAS requires fewer GPU-hours than other approaches and 1000-fold less than "standard" NAS. On CIFAR-10, the ENAS design achieved a test error of 2.89%, comparable to NASNet. On Penn Treebank, the ENAS design reached test perplexity of 55.8. == Evolution == An alternative approach to NAS is based on evolutionary algorithms, which has been employed by several groups. An Evolutionary Algorithm for Neural Architecture Search generally performs the following procedure. First a pool consisting of different candidate architectures along with their validation scores (fitness) is initialised. At each step the architectures in the candidate pool are mutated (e.g.: 3x3 convolution instead of a 5x5 convolution). Next the new architectures are trained from scratch for a few epochs and their validation scores are obtained. This is followed by replacing the lowest scoring architectures in the candidate pool with the better, newer architectures. This procedure is repeated multiple times and thus the candidate pool is refined over time. Mutations in the context of evolving ANNs are operations such as adding or removing a layer, which include changing the type of a layer (e.g., from convolution to pooling), changing the hyperparameters of a layer, or changing the training hyperparameters. On CIFAR-10 and ImageNet, evolution and RL performed comparably, while both slightly outperformed random search. == Bayesian optimization == Bayesian Optimization (BO), which has proven to be an efficient method for hyperparameter optimization, can also be applied to NAS. In this context, the objective function maps an architecture to its validation error after being trained for a number of epochs. At each iteration, BO uses a surrogate to model this objective function based on previously obtained architectures and their validation errors. One then chooses the next architecture to evaluate by maximizing an acquisition function, such as expected improvement, which provides a balance between exploration and exploitation. Acquisition function maximization and objective function evaluation are often computationally expensive for NAS, and make the application of BO challenging in this context. Recently, BANANAS has achieved promising results in this direction by introducing a high-performing instantiation of BO coupled to a neural predictor. == Hill-climbing == Another group used a hill climbing procedure that applies network morphisms, followed by short cosine-annealing optimization runs. The approach yielded competitive results, requiring resources on the same order of magnitude as training a single network. E.g., on CIFAR-10, the method designed and trained a network with an error rate below 5% in 12 hours on a single GPU. == Multi-objective search == While most approaches solely focus on finding architecture with maximal predictive performance, for most practical applications other objectives are relevant, such as memory consumption, model size or inference time (i.e., the time required to obtain a prediction). Because of that, researchers created a multi-objective search. LEMONADE is an evolutionary algorithm that adopted Lamarckism to efficiently optimize multiple objectives. In every generation, child networks are generated to improve the Pareto frontier with respect to the current population of ANNs. Neural Architect is claimed to be a resource-aware multi-objective RL-based NAS with network embedding and performance prediction. Network embedding encodes an existing network to a trainable embedding vector. Based on the embedding, a controller network generates transformations of the target network. A multi-objective reward function considers network accuracy, computational resource and training time. The reward is predicted by multiple performance simulation networks that are pre-trained or co-trained with the controller network. The controller network is trained via policy gradient. Following a modification, the resulting candidate network is evaluated by both an accuracy network and a training time network. The results are combined by a reward engine that passes its output back to the controller network. == One-shot models == RL or evolution-based NAS require thousands of GPU-days of searching/training to achieve state-of-the-art computer vision results as described in the NASNet, mNASNet and MobileNetV3 papers. To reduce computational cost, many recent NAS methods rely on the weight-sharing idea. In this approach, a single overparameterized supernetwork (also known as the one-shot model) is defined. A supernetwork is a very large Directed Acyclic Graph (DAG) whose subgraphs are different candidate neural networks. Thus, in a supernetwork, the weights are shared among a large number of different sub-architectures that have edges in common, each of which is considered as a path within the supernet. The essential idea is to train one supernetwork that spans many options for the final design rather than generating and training thousands of networks independently. In addition to the learned parameters, a set of architecture parameters are learnt to depict preference for one module over another. Such methods reduce the required computational resources to only a few GPU days. More recent works further combine this weight-sharing paradigm, with a continuous relaxation of the search space, which enables the use of gradient-based optimization methods. These approaches are generally referred to as differentiable NAS and have proven very efficient in exploring the search space of ne
Deepfake pornography
Deepfake pornography is a form of non-consensual AI pornography created by altering existing photographs or videos using deepfake technology to modify the appearance of the participants. The use of deepfake pornography has sparked controversy because it involves the making and sharing of realistic videos featuring non-consenting individuals and is sometimes used for revenge porn. Many countries have criminalized this "new voyeurism" through legislative measures and technological solutions. == History == The term "deepfake" was coined in 2017 on a Reddit forum where users shared altered pornographic videos created using machine learning algorithms. It is a combination of the word "deep learning", which refers to the program used to create the videos, and "fake" meaning the videos are not real. Deepfake pornography was originally created on a small individual scale using a combination of machine learning algorithms, computer vision techniques, and AI software. The process began by gathering a large amount of source material (including both images and videos) of a person's face, and then using a deep learning model to train a Generative Adversarial Network to create a fake video that convincingly swaps the face of the source material onto the body of a pornographic performer. However, the production process has significantly evolved since 2018, with the advent of several public apps that have largely automated the process. While several AI "nudification" apps emerged on mainstream platforms like Google Play and the Apple App Store around 2023, major tech storefronts have since implemented stricter policies and automated detection to ban such software. Consequently, the proliferation of non-consensual deepfake pornography has largely shifted to decentralized websites, specialized online forums, and third-party messaging bot ecosystems. Deepfake pornography is sometimes confused with fake nude photography, but the two are mostly different. Fake nude photography typically uses non-sexual images and merely makes it appear that the people in them are nude. == Notable cases == Deepfake technology has been used to create non-consensual and pornographic images and videos of famous women. One of the earliest examples occurred in 2017 when a deepfake pornographic video of Gal Gadot was created by a Reddit user and quickly spread online. Since then, there have been numerous instances of similar deepfake content targeting other female celebrities, such as Emma Watson, Natalie Portman, and Scarlett Johansson. Johansson spoke publicly on the issue in December 2018, condemning the practice but also refusing legal action because she views the harassment as inevitable. === Rana Ayyub === In 2018, Rana Ayyub, an Indian investigative journalist, was the target of an online hate campaign stemming from her condemnation of the Indian government, specifically her speaking out against the rape of an eight-year-old Kashmiri girl. Ayyub was bombarded with rape and death threats, and had a doctored pornographic video of her circulated online. In a Huffington Post article, Ayyub discussed the long-lasting psychological and social effects this experience has had on her. She explained that she continued to struggle with her mental health and how the images and videos continued to resurface whenever she took a high-profile case. === Atrioc controversy === In 2023, Twitch streamer Atrioc stirred controversy when he accidentally revealed deepfake pornographic material featuring female Twitch streamers while on live. The influencer has since admitted to paying for AI generated porn, and apologized to the women and his fans. === Taylor Swift === In January 2024, AI-generated sexually explicit images of American singer Taylor Swift were posted on X (formerly Twitter), and spread to other platforms such as Facebook, Reddit and Instagram. One tweet with the images was viewed over 45 million times before being removed. A report from 404 Media found that the images appeared to have originated from a Telegram group, whose members used tools such as Microsoft Designer to generate the images, using misspellings and keyword hacks to work around Designer's content filters. After the material was posted, Swift's fans posted concert footage and images to bury the deepfake images, and reported the accounts posting the deepfakes. Searches for Swift's name were temporarily disabled on X, returning an error message instead. Graphika, a disinformation research firm, traced the creation of the images back to a 4chan community. A source close to Swift told the Daily Mail that she would be considering legal action, saying, "Whether or not legal action will be taken is being decided, but there is one thing that is clear: These fake AI-generated images are abusive, offensive, exploitative, and done without Taylor's consent and/or knowledge." The controversy drew condemnation from White House Press Secretary Karine Jean-Pierre, Microsoft CEO Satya Nadella, the Rape, Abuse & Incest National Network, and SAG-AFTRA. Several US politicians called for federal legislation against deepfake pornography. Later in the month, US senators Dick Durbin, Lindsey Graham, Amy Klobuchar and Josh Hawley introduced a bipartisan bill that would allow victims to sue individuals who produced or possessed "digital forgeries" with intent to distribute, or those who received the material knowing it was made non-consensually. === 2024 Telegram deepfake scandal === It emerged in South Korea in August 2024, that many teachers and female students were victims of deepfake images created by users who utilized AI technology. Journalist Ko Narin of The Hankyoreh uncovered the deepfake images through Telegram chats. On Telegram, group chats were created specifically for image-based sexual abuse of women, including middle and high school students, teachers, and even family members. Women with photos on social media platforms like KakaoTalk, Instagram, and Facebook are often targeted as well. Perpetrators use AI bots to generate fake images, which are then sold or widely shared, along with the victims' social media accounts, phone numbers, and KakaoTalk usernames. One Telegram group reportedly drew around 220,000 members, according to a Guardian report. Investigations revealed numerous chat groups on Telegram where users, mainly teenagers, create and share explicit deepfake images of classmates and teachers. The issue came in the wake of a troubling history of digital sex crimes, notably the notorious Nth Room case in 2019. The Korean Teachers Union estimated that more than 200 schools had been affected by these incidents. Activists called for a "national emergency" declaration to address the problem. South Korean police reported over 800 deepfake sex crime cases by the end of September 2024, a stark rise from just 156 cases in 2021, with most victims and offenders being teenagers. On September 21, 6,000 people gathered at Marronnier Park in northeastern Seoul to demand stronger legal action against deepfake crimes targeting women. On September 26, following widespread outrage over the Telegram scandal, South Korean lawmakers passed a bill criminalizing the possession or viewing of sexually explicit deepfake images and videos, imposing penalties that include prison terms and fines. Under the new law, those caught buying, saving, or watching such material could face up to three years in prison or fines up to 30 million won ($22,600). At the time the bill was proposed, creating sexually explicit deepfakes for distribution carried a maximum penalty of five years, but the new legislation would increase this to seven years, regardless of intent. By October 2024, it was estimated that "nudify" deep fake bots on Telegram were up to four million monthly users. === 2025–2026 Grok/X chatbot deepfake scandal === In December 2025, Bloomberg reported that X users found Grok would comply with unconsensual requests to digitally undress individuals, including minors, or show them performing sexually explicit acts. The majority of these prompts were targeted at women and girls. An analysis of 20,000 images generated by Grok between December 25, 2025 and January 1, 2026 showed 2% were of people in bikinis or transparent clothes and appeared to be 18 or younger, including 30 of "young or very young" women or girls. A separate analysis conducted over 24 hours from January 5 to 6 calculated that users had Grok create 6,700 sexually suggestive or nudified images per hour. xAI responded to requests for comment from media organizations with the automated reply, "Legacy Media Lies". The bot's image generation sparked an international backlash and calls for legal or regulatory action from officials in the European Union, United Kingdom, Poland, France, India, Malaysia, and Brazil. === Fernandes–Ulmen case === German TV presenter Collien Fernandes, filed a complaint against her ex-husband, actor Christian Ulmen, for several accusation including, ident
Wonder.land
Wonder.land (stylised as wonder.land) is a musical with music by Damon Albarn and lyrics and book by Moira Buffini. Inspired by Lewis Carroll's novels Alice's Adventures in Wonderland (1865) and Through the Looking-Glass (1871), it had its world premiere at the Palace Theatre in Manchester in July 2015 as part of the Manchester International Festival. The musical moved to London's Royal National Theatre in November 2015 before opening at the Théâtre du Châtelet in Paris in 2016. Licencing for potential future smaller scale productions is held by United Agents UK. == Background == The musical is inspired by the novels Alice in Wonderland and Through the Looking-Glass, written by Lewis Carroll. It was announced on 21 January 2015 that the show would premiere in July of that year as part of the Manchester International Festival, with tickets going on sale the following day. The musical, a co-production by the Manchester International Festival, the Royal National Theatre and the Théâtre du Châtelet in Paris, marks the 150th anniversary of the publication of Alice's Adventures in Wonderland. The idea for a musical based on Alice in Wonderland came from Manchester International Festival artistic director Alex Poots. Damon Albarn had collaborated with the festival on Monkey: Journey to the West and Dr Dee. The musical has a book by Moira Buffini. It was directed by Rufus Norris, with set design by Rae Smith, costume design by Katrina Lindsay, lighting design by Paule Constable, projections by 59 Productions and choreography by Javier De Frutos. The musical's score was composed by Damon Albarn, with lyrics by Moira Buffini, sound design by Paul Arditti and musical direction by David Shrubsole. == Production history == The musical began previews at the Palace Theatre in Manchester on 29 June 2015. It opened on 2 July for a limited run until 12 July. A revised version moved to the Royal National Theatre, where it ran at the Olivier Theatre from 27 November 2015 to 30 April 2016. The production had a limited run, from 7 to 16 June 2016, at the Theatre Du Chatelet in Paris. == Synopsis == This synopsis is based on the final version, as seen at the National Theatre and the Théâtre du Châtelet. Earlier performances significantly differed in songs and plot. === Act 1 === AI, the MC, explains that virtual technology is "a portal to boundless lands" ("Prologue"). Aly's mother, Bianca, is exasperated with her for spending the weekend indoors on her phone. Aly accompanies Bianca to the supermarket, and thinks that her life is being ruined by her parents due to dysfunctional problems ("Who's Ruining Your Life?") Her alcoholic father, Matt, is also at the supermarket; he and Bianca argue about their divorce and his gambling. Aly goes home and picks up her phone. She tries to engage with schoolmates, who bully her ("Network"). Aly begins to wish that she is someone else. She finds the virtual online game Wonder.land. In its strange world, Aly creates an avatar: beautiful, kind Alice ("Wonder.land"). Wonder.land has one rule: malice causes deletion from the game. Aly and Alice become friends and encounter the Cheshire Cat, who explains that you can be anyone you want ("Fabulous"). Aly decides to go on a quest; Alice follows the white rabbit down a hole, falling past unusual objects and musical notes ("Falling"). The next morning, Aly is too distracted by Wonder.land to listen to Bianca's complaints about her baby brother Charlie. She plays the game at school before her phone is confiscated by stern headmistress Ms Manxome, who tells her students that taking pleasures from them is for their own good ("I'm Right"). Aly goes to Ms Manxome's office to retrieve her phone. Ms Manxome returns it, warning that if she catches her with it again, "it's a beheading – I mean, detention." Aly sees the girls who bullied her, and they bully her again until a teacher arrives. Aly's friend, Luke, is late and is given detention. Aly goes on her phone and takes out her frustration and sadness on Alice, whose tears form a pool until she is interrupted by the quarrelsome twins Dum and Dee ("Freaks"). Alice tries to befriend them, but they insult her and Aly makes her fight them. Dum and Dee cry, and Aly and Alice see a large mouse who is attracted by Alice's fighting. They are joined by the Dodo, the Mock Turtle and Humpty, who all have problems. The Dodo is stressed because his parents want him to save the planet; Dum and Dee are dancers who hate pressure; Humpty has problems with her parents; the Mock Turtle lacks self-esteem, and the mouse is lustful. Wonderland is a hiding place from teenage life ("Crap Life"). Aly returns to reality when asked a math question she cannot answer. Confronting the three bullies, Aly mocks the facial hair of one and hides in the bathroom. She again immerses herself in Wonder.land, where Alice meets a Caterpillar who is obsessed with identity ("Who are You?"). Aly is interrupted by the girls, who ridicule her father's gambling addiction and poverty before beating her up. Aly seeks understanding from Alice, who tries to get Aly to tell her what is wrong. Aly tells Alice about her family and how she hates her life, and is surprised that Alice has similar problems ("Secrets"). Luke comes into the girls' bathroom because Kieran has threatened him with violence, and hides in a cubicle when Kieran enters. Aly defends Luke, and makes Kieran leave. Luke reveals that the reason Kieran hates him is because, like himself, he is gay. Aly is amazed, and they skip class and play games on their phones. Luke plays Zombie Swarm, and Aly plays Wonder.land. Ms Manxome enters the bathroom; Luke hides his phone, but Aly does not. Ms Manxome confiscates the phone for three months, and Aly and Luke leave. Ms Manxome finds that Aly did not lock her phone, and Alice is calling her. Ms Manxome begins to talk to her, and Alice thinks she is talking to Aly. Aly complains to Luke about her phone being taken away. Matt then takes them out for tea to celebrate his new job at the local garden centre ("In Clover"). At the tea shop, Matt maniacally dances on the tables and plays with spoons; asked to stop, he punches a waiter. Bianca arrives, and they argue again. Aly begins to notice that Wonder.land is invading reality; the MC emerges from a gigantic teapot, and the landscape outside becomes surreal ("Chances"). === Act 2 === Ms Manxome manipulates Alice around Wonder.land on Aly's phone, buys many things, and makes Alice's hair red ("Entre Act"). She tells Alice about her plans to dominate and destroy the online world, and Alice thinks she is talking to Aly ("Me"). Aly, Matt, Bianca, and Charlie are at the police station. PC Rook unsuccessfully tries to get Matt to make a statement (since he is charged with assault and affray), but Matt and Bianca argue again. Aly laments the loss of her family's unity ("Heartless Useless"). In Wonder.land, Ms Manxome is hostile when she meets Dum and Dee, the Mock Turtle, the Dodo, Humpty and the Mouse. She makes Alice chase them away, but Alice and Ms Manxome are driven away by Alice's friends, who are worried about the change in her ("Me (Reprise)"). Bianca learns that Aly missed a detention and had her phone confiscated. Concerned that she is losing Aly to technology, she bans her from the internet ("Gadget"). Charlie vomits, and Aly is left to clean it up. She looks for an internet cafe to go to Wonder.land, the only place she is truly happy ("Everyone Loves Charlie"). At the cafe, Aly cannot log into Wonder.land and her avatar seems to be in use. She sees Alice receive a Vorpal sword, bought by Ms Manxome with the money on Aly's phone. Alice is no longer Alice but the Red Queen, and Ms Manxome tells her to kill her friends. Alice, knowing the person controlling her is not Aly, cannot rebel; she lashes out at her friends, bullying and trying to hurt them. The MC warns that Alice has a deletion warning – any more malice, and she will be deleted. Aly now knows that Ms Manxome controls her phone and avatar ("O Children"). Aly enlists Luke to help and decides to break into Ms. Manxome's office to retrieve the phone. Luke agrees to meet her at the school gates. Matt and Bianca wonder if they should reconcile ("Man of Broken Glass"). At the school, Luke is reluctant to get involved; Aly decides to break into the office anyway. Luke contacts the girls who bullied Aly and tells them about Ms Manxome playing on Aly's stolen phone. They decide to spread the word that it is not Aly ("Fabulous (Reprise)"). Bianca goes to the police because Aly is missing, and gives her phone to Matt. Aly is likely to also be in Wonder.land. The avatars prepare for war against Alice but disagree about a strategy. At the police station, Matt hacks into Wonder.land sees Alice, and realizes that she is controlled by someone other than Aly. The White Rabbit appears (delighting Alice), but Ms Manxome makes Alice push him aside. The borderline between Wonder.land and
Proximal gradient methods for learning
Proximal gradient (forward backward splitting) methods for learning is an area of research in optimization and statistical learning theory which studies algorithms for a general class of convex regularization problems where the regularization penalty may not be differentiable. One such example is ℓ 1 {\displaystyle \ell _{1}} regularization (also known as Lasso) of the form min w ∈ R d 1 n ∑ i = 1 n ( y i − ⟨ w , x i ⟩ ) 2 + λ ‖ w ‖ 1 , where x i ∈ R d and y i ∈ R . {\displaystyle \min _{w\in \mathbb {R} ^{d}}{\frac {1}{n}}\sum _{i=1}^{n}(y_{i}-\langle w,x_{i}\rangle )^{2}+\lambda \|w\|_{1},\quad {\text{ where }}x_{i}\in \mathbb {R} ^{d}{\text{ and }}y_{i}\in \mathbb {R} .} Proximal gradient methods offer a general framework for solving regularization problems from statistical learning theory with penalties that are tailored to a specific problem application. Such customized penalties can help to induce certain structure in problem solutions, such as sparsity (in the case of lasso) or group structure (in the case of group lasso). == Relevant background == Proximal gradient methods are applicable in a wide variety of scenarios for solving convex optimization problems of the form min x ∈ H F ( x ) + R ( x ) , {\displaystyle \min _{x\in {\mathcal {H}}}F(x)+R(x),} where F {\displaystyle F} is convex and differentiable with Lipschitz continuous gradient, R {\displaystyle R} is a convex, lower semicontinuous function which is possibly nondifferentiable, and H {\displaystyle {\mathcal {H}}} is some set, typically a Hilbert space. The usual criterion of x {\displaystyle x} minimizes F ( x ) + R ( x ) {\displaystyle F(x)+R(x)} if and only if ∇ ( F + R ) ( x ) = 0 {\displaystyle \nabla (F+R)(x)=0} in the convex, differentiable setting is now replaced by 0 ∈ ∂ ( F + R ) ( x ) , {\displaystyle 0\in \partial (F+R)(x),} where ∂ φ {\displaystyle \partial \varphi } denotes the subdifferential of a real-valued, convex function φ {\displaystyle \varphi } . Given a convex function φ : H → R {\displaystyle \varphi :{\mathcal {H}}\to \mathbb {R} } an important operator to consider is its proximal operator prox φ : H → H {\displaystyle \operatorname {prox} _{\varphi }:{\mathcal {H}}\to {\mathcal {H}}} defined by prox φ ( u ) = arg min x ∈ H φ ( x ) + 1 2 ‖ u − x ‖ 2 2 , {\displaystyle \operatorname {prox} _{\varphi }(u)=\operatorname {arg} \min _{x\in {\mathcal {H}}}\varphi (x)+{\frac {1}{2}}\|u-x\|_{2}^{2},} which is well-defined because of the strict convexity of the ℓ 2 {\displaystyle \ell _{2}} norm. The proximal operator can be seen as a generalization of a projection. We see that the proximity operator is important because x ∗ {\displaystyle x^{}} is a minimizer to the problem min x ∈ H F ( x ) + R ( x ) {\displaystyle \min _{x\in {\mathcal {H}}}F(x)+R(x)} if and only if x ∗ = prox γ R ( x ∗ − γ ∇ F ( x ∗ ) ) , {\displaystyle x^{}=\operatorname {prox} _{\gamma R}\left(x^{}-\gamma \nabla F(x^{})\right),} where γ > 0 {\displaystyle \gamma >0} is any positive real number. === Moreau decomposition === One important technique related to proximal gradient methods is the Moreau decomposition, which decomposes the identity operator as the sum of two proximity operators. Namely, let φ : X → R {\displaystyle \varphi :{\mathcal {X}}\to \mathbb {R} } be a lower semicontinuous, convex function on a vector space X {\displaystyle {\mathcal {X}}} . We define its Fenchel conjugate φ ∗ : X → R {\displaystyle \varphi ^{}:{\mathcal {X}}\to \mathbb {R} } to be the function φ ∗ ( u ) := sup x ∈ X ⟨ x , u ⟩ − φ ( x ) . {\displaystyle \varphi ^{}(u):=\sup _{x\in {\mathcal {X}}}\langle x,u\rangle -\varphi (x).} The general form of Moreau's decomposition states that for any x ∈ X {\displaystyle x\in {\mathcal {X}}} and any γ > 0 {\displaystyle \gamma >0} that x = prox γ φ ( x ) + γ prox φ ∗ / γ ( x / γ ) , {\displaystyle x=\operatorname {prox} _{\gamma \varphi }(x)+\gamma \operatorname {prox} _{\varphi ^{}/\gamma }(x/\gamma ),} which for γ = 1 {\displaystyle \gamma =1} implies that x = prox φ ( x ) + prox φ ∗ ( x ) {\displaystyle x=\operatorname {prox} _{\varphi }(x)+\operatorname {prox} _{\varphi ^{}}(x)} . The Moreau decomposition can be seen to be a generalization of the usual orthogonal decomposition of a vector space, analogous with the fact that proximity operators are generalizations of projections. In certain situations it may be easier to compute the proximity operator for the conjugate φ ∗ {\displaystyle \varphi ^{}} instead of the function φ {\displaystyle \varphi } , and therefore the Moreau decomposition can be applied. This is the case for group lasso. == Lasso regularization == Consider the regularized empirical risk minimization problem with square loss and with the ℓ 1 {\displaystyle \ell _{1}} norm as the regularization penalty: min w ∈ R d 1 n ∑ i = 1 n ( y i − ⟨ w , x i ⟩ ) 2 + λ ‖ w ‖ 1 , {\displaystyle \min _{w\in \mathbb {R} ^{d}}{\frac {1}{n}}\sum _{i=1}^{n}(y_{i}-\langle w,x_{i}\rangle )^{2}+\lambda \|w\|_{1},} where x i ∈ R d and y i ∈ R . {\displaystyle x_{i}\in \mathbb {R} ^{d}{\text{ and }}y_{i}\in \mathbb {R} .} The ℓ 1 {\displaystyle \ell _{1}} regularization problem is sometimes referred to as lasso (least absolute shrinkage and selection operator). Such ℓ 1 {\displaystyle \ell _{1}} regularization problems are interesting because they induce sparse solutions, that is, solutions w {\displaystyle w} to the minimization problem have relatively few nonzero components. Lasso can be seen to be a convex relaxation of the non-convex problem min w ∈ R d 1 n ∑ i = 1 n ( y i − ⟨ w , x i ⟩ ) 2 + λ ‖ w ‖ 0 , {\displaystyle \min _{w\in \mathbb {R} ^{d}}{\frac {1}{n}}\sum _{i=1}^{n}(y_{i}-\langle w,x_{i}\rangle )^{2}+\lambda \|w\|_{0},} where ‖ w ‖ 0 {\displaystyle \|w\|_{0}} denotes the ℓ 0 {\displaystyle \ell _{0}} "norm", which is the number of nonzero entries of the vector w {\displaystyle w} . Sparse solutions are of particular interest in learning theory for interpretability of results: a sparse solution can identify a small number of important factors. === Solving for L1 proximity operator === For simplicity we restrict our attention to the problem where λ = 1 {\displaystyle \lambda =1} . To solve the problem min w ∈ R d 1 n ∑ i = 1 n ( y i − ⟨ w , x i ⟩ ) 2 + ‖ w ‖ 1 , {\displaystyle \min _{w\in \mathbb {R} ^{d}}{\frac {1}{n}}\sum _{i=1}^{n}(y_{i}-\langle w,x_{i}\rangle )^{2}+\|w\|_{1},} we consider our objective function in two parts: a convex, differentiable term F ( w ) = 1 n ∑ i = 1 n ( y i − ⟨ w , x i ⟩ ) 2 {\displaystyle F(w)={\frac {1}{n}}\sum _{i=1}^{n}(y_{i}-\langle w,x_{i}\rangle )^{2}} and a convex function R ( w ) = ‖ w ‖ 1 {\displaystyle R(w)=\|w\|_{1}} . Note that R {\displaystyle R} is not strictly convex. Let us compute the proximity operator for R ( w ) {\displaystyle R(w)} . First we find an alternative characterization of the proximity operator prox R ( x ) {\displaystyle \operatorname {prox} _{R}(x)} as follows: u = prox R ( x ) ⟺ 0 ∈ ∂ ( R ( u ) + 1 2 ‖ u − x ‖ 2 2 ) ⟺ 0 ∈ ∂ R ( u ) + u − x ⟺ x − u ∈ ∂ R ( u ) . {\displaystyle {\begin{aligned}u=\operatorname {prox} _{R}(x)\iff &0\in \partial \left(R(u)+{\frac {1}{2}}\|u-x\|_{2}^{2}\right)\\\iff &0\in \partial R(u)+u-x\\\iff &x-u\in \partial R(u).\end{aligned}}} For R ( w ) = ‖ w ‖ 1 {\displaystyle R(w)=\|w\|_{1}} it is easy to compute ∂ R ( w ) {\displaystyle \partial R(w)} : the i {\displaystyle i} th entry of ∂ R ( w ) {\displaystyle \partial R(w)} is precisely ∂ | w i | = { 1 , w i > 0 − 1 , w i < 0 [ − 1 , 1 ] , w i = 0. {\displaystyle \partial |w_{i}|={\begin{cases}1,&w_{i}>0\\-1,&w_{i}<0\\\left[-1,1\right],&w_{i}=0.\end{cases}}} Using the recharacterization of the proximity operator given above, for the choice of R ( w ) = ‖ w ‖ 1 {\displaystyle R(w)=\|w\|_{1}} and γ > 0 {\displaystyle \gamma >0} we have that prox γ R ( x ) {\displaystyle \operatorname {prox} _{\gamma R}(x)} is defined entrywise by ( prox γ R ( x ) ) i = { x i − γ , x i > γ 0 , | x i | ≤ γ x i + γ , x i < − γ , {\displaystyle \left(\operatorname {prox} _{\gamma R}(x)\right)_{i}={\begin{cases}x_{i}-\gamma ,&x_{i}>\gamma \\0,&|x_{i}|\leq \gamma \\x_{i}+\gamma ,&x_{i}<-\gamma ,\end{cases}}} which is known as the soft thresholding operator S γ ( x ) = prox γ ‖ ⋅ ‖ 1 ( x ) {\displaystyle S_{\gamma }(x)=\operatorname {prox} _{\gamma \|\cdot \|_{1}}(x)} . === Fixed point iterative schemes === To finally solve the lasso problem we consider the fixed point equation shown earlier: x ∗ = prox γ R ( x ∗ − γ ∇ F ( x ∗ ) ) . {\displaystyle x^{}=\operatorname {prox} _{\gamma R}\left(x^{}-\gamma \nabla F(x^{})\right).} Given that we have computed the form of the proximity operator explicitly, then we can define a standard fixed point iteration procedure. Namely, fix some initial w 0 ∈ R d {\displaystyle w^{0}\in \mathbb {R} ^{d}} , and for k = 1 , 2 , … {\displaystyle k=1,2,\ldots } define w k + 1 = S γ ( w k − γ ∇ F ( w k ) ) . {\displaystyle w^{k+1}=S_{\gamma }\left(w^{k}-\gamma \nabla F\l
Artificial intelligence engineering
Artificial intelligence engineering (AI engineering) is a technical discipline that focuses on the design, development, and deployment of AI systems. AI engineering involves applying engineering principles and methodologies to create scalable, efficient, and reliable AI-based solutions. It merges aspects of data engineering and software engineering to create real-world applications in diverse domains such as healthcare, finance, autonomous systems, and industrial automation. == Terminology ambiguity == According to Chip Huyen's book AI Engineering: Building Applications with Foundation Models, the term AI engineering refers to the process of building applications that use foundation models, which are typically models developed by a small number of research laboratories and made available as a service. Huyen distinguishes this from machine learning (ML) engineering, which involves building and deploying models developed in-house. She notes that most practical AI systems combine both approaches. For example, a customer-support chatbot may use a generative model to produce responses while also incorporating locally built components such as request classifiers or scoring mechanisms to assess response quality. As a result, the terms AI engineering and ML engineering are often used together or interchangeably in practice. The distinction and broader usage of the term have been discussed in industry publications and interviews, where AI engineering has been described as an emerging discipline focused on productionizing applications built with foundation models. == Key components == AI engineering integrates a variety of technical domains and practices, all of which are essential to building scalable, reliable, and ethical AI systems. === Data engineering and infrastructure === Data serves as the cornerstone of AI systems, necessitating careful engineering to ensure premium quality, wide spread availability, and usability. AI engineers gather large, diverse datasets from multiple sources such as databases, APIs, and real-time streams. This data undergoes cleaning, normalization, and preprocessing, often facilitated by automated data pipelines that manage extraction, transformation, and loading (ETL) processes. Efficient storage solutions, such as SQL (or NoSQL) databases and data lakes, must be selected based on data characteristics and use cases. Security measures, including encryption and access controls, are critical for protecting sensitive information and ensuring compliance with regulations like GDPR. Scalability is essential, frequently involving cloud services and distributed computing frameworks to handle growing data volumes effectively. === Algorithm selection and optimization === Selecting the appropriate algorithm is crucial for the success of any AI system. Engineers evaluate the problem (which could be classification or regression, for example) to determine the most suitable machine learning algorithm, including deep learning paradigms. Once an algorithm is chosen, optimizing it through hyperparameter tuning is essential to enhance efficiency and accuracy. Techniques such as grid search or Bayesian optimization are employed, and engineers often utilize parallelization to expedite training processes, particularly for large models and datasets. For existing models, techniques like transfer learning can be applied to adapt pre-trained models for specific tasks, reducing the time and resources needed for training. === Deep learning engineering === Deep learning is particularly important for tasks involving large and complex datasets. Engineers design neural network architectures tailored to specific applications, such as convolutional neural networks for visual tasks or recurrent neural networks for sequence-based tasks. Transfer learning, where pre-trained models are fine-tuned for specific use cases, helps streamline development and often enhances performance. Optimization for deployment in resource-constrained environments, such as mobile devices, involves techniques like pruning and quantization to minimize model size while maintaining performance. Engineers also mitigate data imbalance through augmentation and synthetic data generation, ensuring robust model performance across various classes. === Natural language processing === Natural language processing (NLP) is a crucial component of AI engineering, focused on enabling machines to understand and generate human language. The process begins with text preprocessing to prepare data for machine learning models. Recent advancements, particularly transformer-based models like BERT and GPT, have greatly improved the ability to understand context in language. AI engineers work on various NLP tasks, including sentiment analysis, machine translation, and information extraction. These tasks require sophisticated models that utilize attention mechanisms to enhance accuracy. Applications range from virtual assistants and chatbots to more specialized tasks like named-entity recognition (NER) and Part of speech (POS) tagging. === Reasoning and decision-making systems === Developing systems capable of reasoning and decision-making is a significant aspect of AI engineering. Whether starting from scratch or building on existing frameworks, engineers create solutions that operate on data or logical rules. Symbolic AI employs formal logic and predefined rules for inference, while probabilistic reasoning techniques like Bayesian networks help address uncertainty. These models are essential for applications in dynamic environments, such as autonomous vehicles, where real-time decision-making is critical. === Security === Security is a critical consideration in AI engineering, particularly as AI systems become increasingly integrated into sensitive and mission-critical applications. AI engineers implement robust security measures to protect models from adversarial attacks, such as evasion and poisoning, which can compromise system integrity and performance. Techniques such as adversarial training, where models are exposed to malicious inputs during development, help harden systems against these attacks. Additionally, securing the data used to train AI models is of paramount importance. Encryption, secure data storage, and access control mechanisms are employed to safeguard sensitive information from unauthorized access and breaches. AI systems also require constant monitoring to detect and mitigate vulnerabilities that may arise post-deployment. In high-stakes environments like autonomous systems and healthcare, engineers incorporate redundancy and fail-safe mechanisms to ensure that AI models continue to function correctly in the presence of security threats. === Ethics and compliance === As AI systems increasingly influence societal aspects, ethics and compliance are vital components of AI engineering. Engineers design models to mitigate risks such as data poisoning and ensure that AI systems adhere to legal frameworks, such as data protection regulations like GDPR. Privacy-preserving techniques, including data anonymization and differential privacy, are employed to safeguard personal information and ensure compliance with international standards. Ethical considerations focus on reducing bias in AI systems, preventing discrimination based on race, gender, or other protected characteristics. By developing fair and accountable AI solutions, engineers contribute to the creation of technologies that are both technically sound and socially responsible. == Workload == An AI engineer's workload revolves around the AI system's life cycle, which is a complex, multi-stage process. This process may involve building models from scratch or using pre-existing models through transfer learning, depending on the project's requirements. Each approach presents unique challenges and influences the time, resources, and technical decisions involved. === Problem definition and requirements analysis === Regardless of whether a model is built from scratch or based on a pre-existing model, the work begins with a clear understanding of the problem. The engineer must define the scope, understand the business context, and identify specific AI objectives that align with strategic goals. This stage includes consulting with stakeholders to establish key performance indicators (KPIs) and operational requirements. When developing a model from scratch, the engineer must also decide which algorithms are most suitable for the task. Conversely, when using a pre-trained model, the workload shifts toward evaluating existing models and selecting the one most aligned with the task. The use of pre-trained models often allows for a more targeted focus on fine-tuning, as opposed to designing an entirely new model architecture. === Data acquisition and preparation === Data acquisition and preparation are critical stages regardless of the development method chosen, as the performance of any AI system relies heavily on high-quality, re