In scientific disciplines, a toy problem or a puzzlelike problem is a problem that is not of immediate scientific interest, yet is used as an expository device to illustrate a trait that may be shared by other, more complicated, instances of the problem, or as a way to explain a particular, more general, problem solving technique. A toy problem is useful to test and demonstrate methodologies. Researchers can use toy problems to compare the performance of different algorithms. They are also good for game designing. For instance, while engineering a large system, the large problem is often broken down into many smaller toy problems which have been well understood in detail. Often these problems distill a few important aspects of complicated problems so that they can be studied in isolation. Toy problems are thus often very useful in providing intuition about specific phenomena in more complicated problems. As an example, in the field of artificial intelligence, classical puzzles, games and problems are often used as toy problems. These include sliding-block puzzles, N-Queens problem, missionaries and cannibals problem, tic-tac-toe, chess, Tower of Hanoi and others.
Audio mining
Audio mining is a technique by which the content of an audio signal can be automatically analyzed and searched. It is most commonly used in the field of automatic speech recognition, where the analysis tries to identify any speech within the audio. The term audio mining is sometimes used interchangeably with audio indexing, phonetic searching, phonetic indexing, speech indexing, audio analytics, speech analytics, word spotting, and information retrieval. Audio indexing, however, is mostly used to describe the pre-process of audio mining, in which the audio file is broken down into a searchable index of words. == History == Academic research on audio mining began in the late 1970s in schools like Carnegie Mellon University, Columbia University, the Georgia Institute of Technology, and the University of Texas. Audio data indexing and retrieval began to receive attention and demand in the early 1990s, when multimedia content started to develop and the volume of audio content significantly increased. Before audio mining became the mainstream method, written transcripts of audio content were created and manually analyzed. == Process == Audio mining is typically split into four components: audio indexing, speech processing and recognition systems, feature extraction and audio classification. The audio will typically be processed by a speech recognition system in order to identify word or phoneme units that are likely to occur in the spoken content. This information may either be used immediately in pre-defined searches for keywords or phrases (a real-time "word spotting" system), or the output of the speech recognizer may be stored in an index file. One or more audio mining index files can then be loaded at a later date in order to run searches for keywords or phrases. The results of a search will normally be in terms of hits, which are regions within files that are good matches for the chosen keywords. The user may then be able to listen to the audio corresponding to these hits in order to verify if a correct match was found. === Audio Indexing === In audio, there is the main problem of information retrieval - there is a need to locate the text documents that contain the search key. Unlike humans, a computer is not able to distinguish between the different types of audios such as speed, mood, noise, music or human speech - an effective searching method is needed. Hence, audio indexing allows efficient search for information by analyzing an entire file using speech recognition. An index of content is then produced, bearing words and their locations done through content-based audio retrieval, focusing on extracted audio features. It is done through mainly two methods: Large Vocabulary Continuous Speech Recognition (LVCSR) and Phonetic-based Indexing. ==== Large Vocabulary Continuous Speech Recognizers (LVCSR) ==== In text-based indexing or large vocabulary continuous speech recognition (LVCSR), the audio file is first broken down into recognizable phonemes. It is then run through a dictionary that can contain several hundred thousand entries and matched with words and phrases to produce a full text transcript. A user can then simply search a desired word term and the relevant portion of the audio content will be returned. If the text or word could not be found in the dictionary, the system will choose the next most similar entry it can find. The system uses a language understanding model to create a confidence level for its matches. If the confidence level be below 100 percent, the system will provide options of all the found matches. ===== Advantages and disadvantages ===== The main draw of LVCSR is its high accuracy and high searching speed. In LVCSR, statistical methods are used to predict the likelihood of different word sequences, hence the accuracy is much higher than the single word lookup of a phonetic search. If the word can be found, the probability of the word spoken is very high. Meanwhile, while initial processing of audio takes a fair bit of time, searching is quick as just a simple test to text matching is needed. On the other hand, LVCSR is susceptible to common issues of speech recognition. The inherent random nature of audio and problems of external noise all affect the accuracies of text-based indexing. Another problem with LVCSR is its over reliance on its dictionary database. LVCSR only recognizes words that are found in their dictionary databases, and these dictionaries and databases are unable to keep up with the constant evolving of new terminology, names and words. Should the dictionary not contain a word, there is no way for the system to identify or predict it. This reduces the accuracy and reliability of the system. This is named the Out-of-vocabulary (OOV) problem. Audio mining systems try to cope with OOV by continuously updating the dictionary and language model used, but the problem still remains significant and has probed a search for alternatives. Additionally, due to the need to constantly update and maintain task-based knowledge and large training databases to cope with the OOV problem, high computational costs are incurred. This makes LVCSR an expensive approach to audio mining. ==== Phonetic-based Indexing ==== Phonetic-based indexing also breaks the audio file into recognizable phonemes, but instead of converting them to a text index, they are kept as they are and analyzed to create a phonetic-based index. The process of phonetic-based indexing can be split into two phases. The first phase is indexing. It begins by converting the input media into a standard audio representation format (PCM). Then, an acoustic model is applied to the speech. This acoustic model represents characteristics of both an acoustic channel (an environment in which the speech was uttered and a transducer through which it was recorded) and a natural language (in which human beings expressed the input speech). This produces a corresponding phonetic search track, or phonetic audio track (PAT), a highly compressed representation of the phonetic content of the input media. The second phase is searching. The user's search query term is parsed into a possible phoneme string using a phonetic dictionary. Then, multiple PAT files can be scanned at high speed during a single search for likely phonetic sequences that closely match corresponding strings of phonemes in the query term. ===== Advantages and disadvantages ===== Phonetic indexing is most attractive as it is largely unaffected by linguistic issues such as unrecognized words and spelling errors. Phonetic preprocessing maintains an open vocabulary that does not require updating. That makes it particularly useful for searching specialized terminology or words in foreign languages that do not commonly appear in dictionaries. It is also more effective for searching audio files with disruptive background noise and/or unclear utterances as it can compile results based on the sounds it can discern, and should the user wish to, they can search through the options until they find the desired item. Furthermore, in contrast to LVCSR, it can process audio files very quickly as there are very few unique phonemes between languages. However, phonemes cannot be effectively indexed like an entire word, thus searching on a phonetic-based system is slow. An issue with phonetic indexing is its low accuracy. Phoneme-based searches result in more false matches than text-based indexing. This is especially prevalent for short search terms, which have a stronger likelihood of sounding similar to other words or being part of bigger words. It could also return irrelevant results from other languages. Unless the system recognizes exactly the entire word, or understands phonetic sequences of languages, it is difficult for phonetic-based indexing to return accurate findings. === Speech processing and recognition system === Deemed as the most critical and complex component of audio mining, speech recognition requires the knowledge of human speech production system and its modeling. To correspond the Human speech production system, the electrical speech production system is developed to consist of: Speech generation Speech perception Voiced & unvoiced speech Model of human speech The electrical speech production system converts acoustic signal into corresponding representation of the spoken through the acoustic models in their software where all phonemes are represented. A statistical language model aids in the process by identifying how likely words are to follow each other in certain languages. Put together with a complex probability analysis, the speech recognition system is capable of taking an unknown speech signal and transcribing it into words based on the program's dictionary. ASR (automatic speech recognition) system includes: Acoustic analysis: input sound waveform is transformed into a feature Acoustic model: establishes relationship between speech signal and phonemes, pronunciation model and lang
AI Snake Oil
AI Snake Oil: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference is a 2024 non-fiction book written by scholars Arvind Narayanan and Sayash Kapoor. It is a critique of the tech industry's overly inflated promises and capabilities of artificial intelligence (AI) as well as a debunking of the flawed science fueling AI hype while attempting to outline both the potential positives and negatives that come with different modes of the technology. == Contents == === Publication === The book was published in September 2024 by the Princeton University Press. AI Snake Oil consists of 360 pages and features eight chapters, and sections for acknowledgements, references, and an index. An updated edition with a new preface and epilogue by the authors was published in September 2025. The authors use the term "AI snake oil" derived from the U.S. idiom for a fraudulent remedy, to describe overhyped AI systems. === Chapter one: Introduction === Narayanan and Kapoor argue that many individuals do not yet have the literacy to detect functioning aspects of AI compared to potential snake oil, which they identify as "AI that does not and cannot work as advertised". Some of the major examples utilized by the authors include Allstate's 2013 use of predictive AI, as well as the concern surrounding actors and AI attempting to replicate or use their likeness. Important discussions regarding discrimination are brought up and explored in the first chapter, including the false arrests of six Black individuals due to errors with AI facial recognition tools. The chapter concludes with a comparison to the Industrial Revolution, where Narayanan and Kapoor highlight the extensive human labour that is necessary for artificial intelligence technologies to function. === Chapter two: How Predictive AI Goes Wrong === Chapter two focuses on predictive artificial intelligence, and criticizes the overestimation of the capabilities of the technology. === Chapter three: Why Can't AI Predict the Future? === Chapter three works to inform the reader about the history of early computational prediction attempts, with examples from companies like Simulatics. === Chapter four: The Long Road to Generative AI === The fourth chapter goes in more in-depth in explorations of generative AI. Generative AI software examples include ChatGPT, Midjourney, and DALL-E. The section begins with a positive example of generative AI. As the chapter progresses, the authors begin to provide examples of harm produced by generative AI, including the suicide of a Belgian man after connecting with Chai, a generative chatbot. Issues of deepfakes and preservation of artistic property are also discussed. The use of generative AI to create non-consensual pornographic deepfake content is discussed in relation to female celebrities. === Chapter five: Is Advanced AI an Existential Threat? === The fifth chapter draws attention the AGI, or Artificial General Intelligence. The authors describe AGI as "AI that can perform most or economically relevant tasks as effectively as any human". They summarize that many contributors to the field of artificial intelligence believe AGI to be an impending threat that demands attention. However, they argue that the perceived threat of AGI would only exist if the technology continually functioned reliably. In order to better illustrate the hype surrounding AGI, Narayanan and Kapoor use the Ladder of Generality, which is described as a visual tool in which "each rung represents a way of computing that is more flexible, and more general, than the previous one". They note that we are not yet aware of the next rungs on the ladder, or if the ladder will eventually result in a dead end. The rungs that have been identified so far are as follows: (0, or floor) special purpose hardware, (1) programmable computers, (2) stored program computers, (3) machine learning, (4) deep learning, (5) pretrained models, and, finally, (6) instruction-tuned models. The potential for future rungs and what those rungs might be are currently undetermined. The chapter also discusses the ELIZA effect, which Lawrence Switzky discusses in his article "ELIZA Effects". Switzky attributes the coined term ELIZA Effect to Sherry Turke, who defined it as "our more general tendency to treat responsive computer programs as more intelligent than they really are". === Chapter six: Why Can't AI Fix Social Media? === The sixth chapter focuses on content moderation, why it is important, and how it has been and could be affected by artificial automation. The first issue raised in regard to AI-driven content moderation is the inability for computers and machines to understand context and nuance, resulting in potential for discriminatory moderation and shadow banning. While they note that there are issues with automating content moderation, Narayanan and Kapoor also highlight the psychological impact on human content moderators and their labour. They indicate the hidden labour behind moderation, which is often outsourced to less developed countries, where labourers sort through potentially traumatizing content for pay. However, the discussion focuses more heavily on why automated moderation can be problematic, including discriminatory algorithms and lack of nuance. To balance their argument, issues of discrimination and bias are also discussed in relation the human content moderators. To automate moderation, there are two types of AI used, which are fingerprint matching and machine learning. === Chapter seven: Why Do Myths about AI Persist? === The seventh chapter outlines possible factors that contribute to hype surrounding AI. Narayanan and Kapoor explain how companies often promote their new AI models without properly disclosing how the model works, and what it is learning from. They attribute hype to several different groups, including journalists, researchers, and companies. They explain the impact of companies and the misplaced hype that they spread can be attributed to greed and a desire to grow corporate funds. For journalists, one of the stated sources of hype, they argue that news media has a tendency to prioritize financial incentives over validity and quality of writing. As well, Narayanan and Kapoor point out the emergence of company statement regurgitation in news media, leading to clickbait. Hype from researchers is potentially linked to lack of reproducibility in studies as well as leakage, which occurs when AI models are tested on their training data. === Chapter eight: Where do we go from here? === The final chapter, chapter eight, turns its attention to the future. The authors express their ideas and predictions for how the technology will evolve and be utilized in the upcoming years. == Authors == Author Narayanan is a computer science professor at Princeton University. Kapoor is a doctoral candidate at the same university, and both scholars are located at the Center for Information Technology at Princeton. In 2023, Narayanan and Kapoor appeared on the TIME100 Artificial Intelligence list, which features influential figures in the field. == Reception == Nature, a science and technology peer-reviewed journal, released an article highlighting the top "10 essential reads from the past year", listing Arvind Narayanan and Sayash Kapoor's AI Snake Oil. The article states the that text is "one of the best on this controversial subject". Elizabeth Quill, in her review of the text in Science News, writes that the authors "squarely achieve their stated goal: to empower people to distinguish AI that works well from AI snake oil". Joshua Rothman of The New Yorker writes that "compared with many technologists, Narayanan, Kapoor, and Vallor [Shannon Vallor, University of Edinburgh], are deeply skeptical about today's A.I. technology and what it can achieve. Perhaps they shouldn't be". Rothman argues, following an interview with prominent computer scientist Geoffrey Hinton of University of Toronto, that the potential for AI to replicate complexity is already here and continues to be heavily funded, enhancing the prospective capabilities of the technology. However, he does praise the author's ability to address questions regarding the existential human experience. Alexya Martinez discusses the text in a book review for Journalism and Mass Communication Quarterly, critiquing AI Snake Oil for its extensive focus on the West. Martinez writes that Narayanan and Kapoor "do not fully explore how AI impacts other countries", and suggests more focus on countries outside of the United States to enhance their argument.
Willy's Chocolate Experience
Willy's Chocolate Experience was an unlicensed event based on Charlie and the Chocolate Factory that took place in Glasgow, Scotland, in February 2024. The event was promoted as an immersive and interactive family experience, illustrated on a promotional website with "dreamlike" AI-generated images. Once it was discovered that the event was held in a sparsely decorated warehouse, many customers complained, and the police were called to the venue. The event went viral on the Internet and attracted worldwide media attention. The event drew comparisons to the 2008 Lapland New Forest controversy, the 2014 Tumblr fan convention DashCon, and Billy McFarland's 2017 Fyre Festival. == Background and advertising == The event was stated to take place over the weekend of 24–25 February 2024. Promotional material advertised "stunning and intricately designed settings inspired by Roald Dahl's timeless tale" and "an array of delectable treats scattered throughout the experience". Both the website and promotional material used poor-quality AI-generated images, which included several spelling errors such as "cartchy tuns" and "a pasadise of sweet teats" and nonsensical words such as "catgacating" and "exarserdray". Tickets cost up to £35 per person. While the event was being promoted in early February, a Reddit user who saw Facebook advertisements suspected it to be a scam and was surprised that people were apparently buying tickets based solely on AI-generated images. The event was organised by House of Illuminati, a company registered to Billy Coull which claimed to offer "unparalleled immersive experiences". An investigation by Third Force News conducted after the event described Coull's previous "murky involvement in the charity sector." Coull had previously registered several other companies and claimed to work as a "consultant" for the now-defunct brand Empowerity, formerly known as the charity Gowanbank Community Hub. In 2021, Gowanbank was forced to remove claims of a £95-per-ticket fundraising "gala" at DoubleTree Glasgow which had been falsely advertised to feature TV personalities and performers including Gok Wan and Joe Black. Coull had claimed to be a doctor with a fake degree from a false university that provided "metaphysical degrees", and had attempted to use the charity to win the 2022 Glasgow City Council election in the seat of Greater Pollok, though he never registered for the election. In the summer of 2023, he independently published 17 AI-generated books on various topics, including vaccine conspiracy theories. Rolling Stone concluded that House of Illuminati's websites and event descriptions were likely written by an AI chatbot, such as ChatGPT. Three actors were hired to portray "Willy McDuff", a character based on Willy Wonka. One of them, Paul Connell, said that the cast were given one day to learn the script. Another actor playing Willy McDuff was 18-year-old Michael Archibald; the experience was his first ever acting job, and he was given the script at 6 pm on Friday before the event began on Saturday. Kirsty Paterson, an actress who played one of the Oompa-Loompas (called "Wonkidoodles" in the script), said that the job offer had been posted on Indeed.com and offered £500 for two days of work. The day before the event, the actors attended a dress rehearsal at the sparsely decorated venue. They were told that others would be working through the night on the production. When they returned on the day of the event, the venue was in the same condition. Paterson was given her costume an hour before the event opened, saying that "We were just handed an Amazon box that probably arrived that morning." == Script == The script for the event is titled Wonkidoodles at McDuff's Chocolate Factory: A Script, and describes Willy McDuff leading an audience through the Garden of Enchantment and the Twilight Tunnel. Once there, they are confronted by a character called The Unknown, described as "an evil chocolate maker who lives in the walls" who seeks to steal the magical "Anti-Graffiti Gobstopper" from McDuff's Imagination Lab. The gobstopper is "a sweet so powerful, it can make any room sparkle without lifting a finger". McDuff defeats The Unknown by amplifying the power of the gobstopper and causing his enemy to be "gently swept up by a robotic vacuum, humorously ending the confrontation". The script was unusual in that it included stage directions for the audience, and descriptions of their reactions. Connell described it as "15 pages of AI-generated gibberish of me just monologuing these mad things", and compared the vacuum cleaner plot point to that of the Nintendo video game Luigi's Mansion. Interviewed after the event, Coull claimed to have written the script himself, using AI only to "check spelling, grammar, and continuity" as he said he had dyslexia. == Event == The event was held at the Box Hub Warehouse event space in Whiteinch, an industrial area of Glasgow. Customers described the venue as "little more than an abandoned, empty warehouse", with set dressings including a small bouncy castle, AI-generated backdrop images pinned to some of the walls, and props which were "strewn about on bare concrete floors". The venue's windows were dirty and its air conditioning systems were left exposed. Paterson has stated that by the time she saw the venue, she had already signed her contract and "didn't want to disappoint the kids", and thus chose to proceed with the work. The Unknown was played by a 16-year-old actress named Felicia Dawkins, who wore a silver mask and a black cloak. Young children were frightened by the character, who appeared from behind a large rectangular mirror. Despite the script calling for The Unknown to be defeated with a vacuum cleaner, no such prop was provided, and actors were instead asked to improvise. Connell said that he and other employees were told to give each child "two jelly beans and a quarter of a cup of lemonade", although the limited supply of jelly beans quickly ran out. Paterson and another "Wonkidoodle" actress, Jenny Fogarty, said that after the first three 45-minute performances, the cast were told to abandon the script and instead let guests walk through the venue, a process that Paterson said took "about two minutes". The character of The Unknown, previously introduced as the main antagonist, was now "scaring children for no reason". One of the actors playing McDuff improvised the idea that children should pull a "silly face" at The Unknown to scare them away, but Dawkins said that, in other cases, she "just had to awkwardly walk back to my corner". Connell was told he would be given a 15-minute break every 45 minutes, but on the day of the event, he played Willy McDuff for three and a half hours without a break. After returning from a lunch break, Connell encountered a crowd of customers demanding refunds from Coull, and the other actors were unsure what to do next. After being told that the event was now cancelled halfway through its opening day, the actors left and went to a pub. Upon returning to the venue some time later, Connell said that he felt "the threat of violence had become quite high" and that there were two police vans and two squad cars at the scene. == Customer reviews and response == Willy's Chocolate Experience was widely criticised by those who attended it, many of whom demanded refunds. One customer, who had driven with his children for two hours to reach the event, described it as an "absolute con". Other visitors who arrived after the event was closed and were not informed of its cancellation requested compensation for wasted rail fares. Following the event's cancellation, Coull offered to refund 850 people, a statement repeated by the event's Facebook page. Some Facebook users stated that they had received their money back. Paterson and Fogarty stated that they only received half of their paycheque. Box Hub, the organisation that had rented the warehouse to House of Illuminati, issued an apology on House of Illuminati's behalf, stating that they "either have no regards for the families and young children they have disappointed or are too embarrassed to comment", and offered to provide a venue free of charge for those who attended the event. House of Illuminati later stated that they would not host any future events. Coull deleted his LinkedIn profile, his YouTube channel, and his personal website in response to the controversy. A few days after the event, Connell said he felt that Coull was "probably one of the most disliked people in Glasgow right now". In an interview with The Sunday Times, Coull apologised for how the event turned out, saying he would accept responsibility. == Fundraising == In an interview with Wired magazine, Connell stated that he and the other actors were working with parents to provide a free show for the children who attended. Some items from the event were later auctioned for charity. The venue auctioned the leftover hand-written "even
Asian Digital Finance Forum & Awards
Asian Digital Finance Forum & Awards (also known as Asian Digital Finance Forum and Awards) is a forum and honorary awards platform convened in Colombo, Sri Lanka. It has been hosted in a hybrid format (virtual and in-person), with editions reported in 2022, 2023 and 2025. The event is organised by the Asian FinTech Academy (AFTA) in collaboration with a number of local and international institutions. == Overview == The forum has featured international academic, industry, and policy speakers and has recognised institutions and individuals for contributions related to digital finance and fintech innovation. Media coverage has described participation and recognition at the forum as spanning multiple regions, with institutions and individuals from South Asia, Southeast Asia, East Asia, the Middle East, Europe, and North America featured across different editions. == Awards and recognition == The forum and awards were held in a hybrid format with virtual and in-person proceedings at Hilton Colombo in the 2022 and 2023 editions. The Asian Digital Finance Forum & Awards presents honorary recognitions to institutions and individuals for contributions to digital finance, financial inclusion, and related regulatory, technological, and policy developments. Media coverage has described the recognitions as non-competitive and based on demonstrated leadership and impact rather than open nominations. In 2025, the forum and awards served as an anchor initiative associated with the Asia International Digital Economy & AI in Finance Summit at Port City Colombo, with an emphasis on artificial intelligence in finance, financial inclusion, and governance-related themes. === 2022 === According to reporting by Daily FT, institutions recognised at the 2022 edition included Sri Lanka’s Bank of Ceylon, Commercial Bank of Ceylon, Hatton National Bank, and People’s Bank, alongside international organisations and fintech-sector contributors. === 2023 === Coverage of the 2023 forum described recognitions awarded to India’s International Financial Services Centres Authority (IFSCA) for regulatory innovation, as well as to digital finance and payments platforms including Dialog Genie and SLT-Mobitel mCash. IDEMIA’s Asia–Pacific operations were also recognised for contributions related to biometric and digital identity technologies in financial services. === 2025 === For the 2025 edition, institutional honourees reported in the media included Nium (Singapore), recognised for cross-border payments optimisation, and Paytm (India), recognised for AI-powered financial inclusion initiatives. A Visionary Award for Next-Generation Financial Hub Development was presented to Port City Colombo in recognition of its fintech- and AI-oriented development strategy. Individual honourees reported for 2025 included Sopnendu Mohanty (Singapore), Neil Tan (Hong Kong), Purvi Munot (United Arab Emirates), and Amira Abdelaziz (Egypt), recognised for contributions spanning fintech governance, ecosystem development, inclusive wealth technology, and AI-driven financial policy and regulation. In 2025, media reports described the awards as being subject to an independent validation framework. The process was led by Dr. Sivaguru S. Sritharan, appointed as Global Validation Chair, and involved independent research, analytical review, and benchmarking against international standards, with recognitions characterised as honorary and non-competitive.
Projection-slice theorem
In mathematics, the projection-slice theorem, central slice theorem or Fourier slice theorem in two dimensions states that the results of the following two calculations are equal: Take a two-dimensional function f(r), project (e.g. using the Radon transform) it onto a (one-dimensional) line, and do a Fourier transform of that projection. Take that same function, but do a two-dimensional Fourier transform first, and then slice the function through its origin, parallel to the projection line. In operator terms, if F1 and F2 are the 1- and 2-dimensional Fourier transform operators mentioned above, P1 is the projection operator (which projects a 2-D function onto a 1-D line), S1 is a slice operator (which extracts a 1-D central slice from a function), then F 1 P 1 = S 1 F 2 . {\displaystyle F_{1}P_{1}=S_{1}F_{2}.} This idea can be extended to higher dimensions. This theorem is used, for example, in the analysis of medical CT scans where a "projection" is an x-ray image of an internal organ. The Fourier transforms of these images are seen to be slices through the Fourier transform of the 3-dimensional density of the internal organ, and these slices can be interpolated to build up a complete Fourier transform of that density. The inverse Fourier transform is then used to recover the 3-dimensional density of the object. This technique was first derived by Ronald N. Bracewell in 1956 for a radio-astronomy problem. == The projection-slice theorem in N dimensions == In N dimensions, the projection-slice theorem states that the Fourier transform of the projection of an N-dimensional function f(r) onto an m-dimensional linear submanifold is equal to an m-dimensional slice of the N-dimensional Fourier transform of that function consisting of an m-dimensional linear submanifold through the origin in the Fourier space which is parallel to the projection submanifold. In operator terms: F m P m = S m F N . {\displaystyle F_{m}P_{m}=S_{m}F_{N}.\,} == The generalized Fourier-slice theorem == In addition to generalizing to N dimensions, the projection-slice theorem can be further generalized with an arbitrary change of basis. For convenience of notation, we consider the change of basis to be represented as B, an N-by-N invertible matrix operating on N-dimensional column vectors. Then the generalized Fourier-slice theorem can be stated as F m P m B = S m B − T | B − T | F N {\displaystyle F_{m}P_{m}B=S_{m}{\frac {B^{-T}}{|B^{-T}|}}F_{N}} where B − T = ( B − 1 ) T {\displaystyle B^{-T}=(B^{-1})^{T}} is the transpose of the inverse of the change of basis transform. == Proof in two dimensions == The projection-slice theorem is easily proven for the case of two dimensions. Without loss of generality, we can take the projection line to be the x-axis. There is no loss of generality because if we use a shifted and rotated line, the law still applies. Using a shifted line (in y) gives the same projection and therefore the same 1D Fourier transform results. The rotated function is the Fourier pair of the rotated Fourier transform, for which the theorem again holds. If f(x, y) is a two-dimensional function, then the projection of f(x, y) onto the x axis is p(x) where p ( x ) = ∫ − ∞ ∞ f ( x , y ) d y . {\displaystyle p(x)=\int _{-\infty }^{\infty }f(x,y)\,dy.} The Fourier transform of f ( x , y ) {\displaystyle f(x,y)} is F ( k x , k y ) = ∫ − ∞ ∞ ∫ − ∞ ∞ f ( x , y ) e − 2 π i ( x k x + y k y ) d x d y . {\displaystyle F(k_{x},k_{y})=\int _{-\infty }^{\infty }\int _{-\infty }^{\infty }f(x,y)\,e^{-2\pi i(xk_{x}+yk_{y})}\,dxdy.} The slice is then s ( k x ) {\displaystyle s(k_{x})} s ( k x ) = F ( k x , 0 ) = ∫ − ∞ ∞ ∫ − ∞ ∞ f ( x , y ) e − 2 π i x k x d x d y {\displaystyle s(k_{x})=F(k_{x},0)=\int _{-\infty }^{\infty }\int _{-\infty }^{\infty }f(x,y)\,e^{-2\pi ixk_{x}}\,dxdy} = ∫ − ∞ ∞ [ ∫ − ∞ ∞ f ( x , y ) d y ] e − 2 π i x k x d x {\displaystyle =\int _{-\infty }^{\infty }\left[\int _{-\infty }^{\infty }f(x,y)\,dy\right]\,e^{-2\pi ixk_{x}}dx} = ∫ − ∞ ∞ p ( x ) e − 2 π i x k x d x {\displaystyle =\int _{-\infty }^{\infty }p(x)\,e^{-2\pi ixk_{x}}dx} which is just the Fourier transform of p(x). The proof for higher dimensions is easily generalized from the above example. == The FHA cycle == If the two-dimensional function f(r) is circularly symmetric, it may be represented as f(r), where r = |r|. In this case the projection onto any projection line will be the Abel transform of f(r). The two-dimensional Fourier transform of f(r) will be a circularly symmetric function given by the zeroth-order Hankel transform of f(r), which will therefore also represent any slice through the origin. The projection-slice theorem then states that the Fourier transform of the projection equals the slice or F 1 A 1 = H , {\displaystyle F_{1}A_{1}=H,} where A1 represents the Abel-transform operator, projecting a two-dimensional circularly symmetric function onto a one-dimensional line, F1 represents the 1-D Fourier-transform operator, and H represents the zeroth-order Hankel-transform operator. == Extension to fan beam or cone-beam CT == The projection-slice theorem is suitable for CT image reconstruction with parallel beam projections. It does not directly apply to fanbeam or conebeam CT. The theorem was extended to fan-beam and conebeam CT image reconstruction by Shuang-ren Zhao in 1995.
Stable Diffusion
Stable Diffusion is a deep learning, text-to-image model released in 2022 based on diffusion techniques. The generative artificial intelligence technology is the premier product of Stability AI and is considered to be a part of the ongoing AI boom. It is primarily used to generate detailed images conditioned on text descriptions, though it can also be applied to other tasks such as inpainting, outpainting, and generating image-to-image translations guided by a text prompt. Its development involved researchers from the CompVis Group at LMU Munich and Runway with a computational donation from Stability and training data from non-profit organizations. Stable Diffusion is a latent diffusion model, a kind of deep generative artificial neural network. Its code and model weights have been released publicly, and an optimized version can run on most consumer hardware equipped with a modest GPU with as little as 2.4 GB VRAM. This marked a departure from previous proprietary text-to-image models such as DALL-E and Midjourney which were accessible only via cloud services. == Development == Stable Diffusion originated from a project called Latent Diffusion, developed in Germany by researchers at LMU Munich in Munich and Heidelberg University. Four of the original 5 authors (Robin Rombach, Andreas Blattmann, Patrick Esser and Dominik Lorenz) later joined Stability AI and released subsequent versions of Stable Diffusion. The technical license for the model was released by the CompVis group at LMU Munich. Development was led by Patrick Esser of Runway and Robin Rombach of CompVis, who were among the researchers who had earlier invented the latent diffusion model architecture used by Stable Diffusion. Stability AI also credited EleutherAI and LAION (a German nonprofit which assembled the dataset on which Stable Diffusion was trained) as supporters of the project. == Technology == === Architecture === Diffusion models, introduced in 2015, are trained with the objective of removing successive applications of Gaussian noise on training images, which can be thought of as a sequence of denoising autoencoders. The name diffusion is from the thermodynamic diffusion, since they were first developed with inspiration from thermodynamics. Models in Stable Diffusion series before SD 3 all used a variant of diffusion models, called latent diffusion model (LDM), developed in 2021 by the CompVis (Computer Vision & Learning) group at LMU Munich. Stable Diffusion consists of 3 parts: the variational autoencoder (VAE), U-Net, and an optional text encoder. The VAE encoder compresses the image from pixel space to a smaller dimensional latent space, capturing a more fundamental semantic meaning of the image. Gaussian noise is iteratively applied to the compressed latent representation during forward diffusion. The U-Net block, composed of a ResNet backbone, denoises the output from forward diffusion backwards to obtain a latent representation. Finally, the VAE decoder generates the final image by converting the representation back into pixel space. The denoising step can be flexibly conditioned on a string of text, an image, or another modality. The encoded conditioning data is exposed to denoising U-Nets via a cross-attention mechanism. For conditioning on text, the fixed, pretrained CLIP ViT-L/14 text encoder is used to transform text prompts to an embedding space. Researchers point to increased computational efficiency for training and generation as an advantage of LDMs. With 860 million parameters in the U-Net and 123 million in the text encoder, Stable Diffusion is considered relatively lightweight by 2022 standards, and unlike other diffusion models, it can run on consumer GPUs, and even CPU-only if using the OpenVINO version of Stable Diffusion. ==== SD XL ==== The XL version uses the same LDM architecture as previous versions, except larger: larger UNet backbone, larger cross-attention context, two text encoders instead of one, and trained on multiple aspect ratios (not just the square aspect ratio like previous versions). The SD XL Refiner, released at the same time, has the same architecture as SD XL, but it was trained for adding fine details to preexisting images via text-conditional img2img. ==== SD 3.0 ==== The 3.0 version completely changes the backbone. Not a UNet, but a Rectified Flow Transformer, which implements the rectified flow method with a Transformer. The Transformer architecture used for SD 3.0 has three "tracks", for original text encoding, transformed text encoding, and image encoding (in latent space). The transformed text encoding and image encoding are mixed during each transformer block. The architecture is named "multimodal diffusion transformer (MMDiT), where the "multimodal" means that it mixes text and image encodings inside its operations. This differs from previous versions of DiT, where the text encoding affects the image encoding, but not vice versa. === Training data === Stable Diffusion was trained on pairs of images and captions taken from LAION-5B, a publicly available dataset derived from Common Crawl data scraped from the web, where 5 billion image-text pairs were classified based on language and filtered into separate datasets by resolution, a predicted likelihood of containing a watermark, and predicted "aesthetic" score (e.g. subjective visual quality). The dataset was created by LAION, a German non-profit which receives funding from Stability AI. The Stable Diffusion model was trained on three subsets of LAION-5B: laion2B-en, laion-high-resolution, and laion-aesthetics v2 5+. A third-party analysis of the model's training data identified that out of a smaller subset of 12 million images taken from the original wider dataset used, approximately 47% of the sample size of images came from 100 different domains, with Pinterest taking up 8.5% of the subset, followed by websites such as WordPress, Blogspot, Flickr, DeviantArt and Wikimedia Commons. An investigation by Bayerischer Rundfunk showed that LAION's datasets, hosted on Hugging Face, contain large amounts of private and sensitive data. === Training procedures === The model was initially trained on the laion2B-en and laion-high-resolution subsets, with the last few rounds of training done on LAION-Aesthetics v2 5+, a subset of 600 million captioned images which the LAION-Aesthetics Predictor V2 predicted that humans would, on average, give a score of at least 5 out of 10 when asked to rate how much they liked them. The LAION-Aesthetics v2 5+ subset also excluded low-resolution images and images which LAION-5B-WatermarkDetection identified as carrying a watermark with greater than 80% probability. Final rounds of training additionally dropped 10% of text conditioning to improve Classifier-Free Diffusion Guidance. The model was trained using 256 Nvidia A100 GPUs on Amazon Web Services for a total of 150,000 GPU-hours, at a cost of $600,000. === Limitations === Stable Diffusion has issues with degradation and inaccuracies in certain scenarios. Initial releases of the model were trained on a dataset that consists of 512×512 resolution images, meaning that the quality of generated images noticeably degrades when user specifications deviate from its "expected" 512×512 resolution; the version 2.0 update of the Stable Diffusion model later introduced the ability to natively generate images at 768×768 resolution. Another challenge is in generating human limbs due to poor data quality of limbs in the LAION database. The model is insufficiently trained to replicate human limbs and faces due to the lack of representative features in the database, and prompting the model to generate images of such type can confound the model. In addition to human limbs, Stable Diffusion is unable to generate legible ambigrams and some other forms of text and typography. Stable Diffusion XL (SDXL) version 1.0, released in July 2023, introduced native 1024x1024 resolution and improved generation for limbs and text. Accessibility for individual developers can also be a problem. In order to customize the model for new use cases that are not included in the dataset, such as generating anime characters ("waifu diffusion"), new data and further training are required. Fine-tuned adaptations of Stable Diffusion created through additional retraining have been used for a variety of different use-cases, from medical imaging to algorithmically generated music. However, this fine-tuning process is sensitive to the quality of new data; low resolution images or different resolutions from the original data can not only fail to learn the new task but degrade the overall performance of the model. Even when the model is additionally trained on high quality images, it is difficult for individuals to run models in consumer electronics. For example, the training process for waifu-diffusion requires a minimum 30 GB of VRAM, which exceeds the usual resource provided in such consumer GPUs as Nvidia's GeForce 30 series, w