The ELVIS Act or Ensuring Likeness Voice and Image Security Act, signed into law by Tennessee Governor Bill Lee on March 21, 2024, marked a significant milestone in the area of regulation of artificial intelligence and public sector policies for artists in the era of artificial intelligence (AI) and AI alignment. It was noted as the first enacted legislation in the United States specifically designed to protect musicians from the unauthorized use of their voices through artificial intelligence technologies and against audio deepfakes and voice cloning. This legislation distinguishes itself by adding penalties for copying a performer's voice. == Origin and advocacy == The inception of the ELVIS Act has been attributed to Gebre Waddell, founder of Sound Credit, who initially conceptualized a framework in 2023 that later evolved into the legislation. Representative Justin J. Pearson acknowledged Waddell's pivotal role during the March 4 House Floor Session on the bill. Leading Tennessee musicians supported the ELVIS Act. Tennessee Governor Bill Lee endorsed it as a Governor's Bill, and it was introduced in the Tennessee Legislature as House Bill 2091 by William Lamberth (R-44) and Senate Bill 2096 by Jack Johnson (R-27). The ELVIS Act is an amendment to a 1984 law that was the result of the Elvis Presley estate litigation for controlling how his likeness could be used after death. == Lobbying from the recording industry == The legislative journey of the ELVIS Act included a broad coalition of music industry stakeholders, including: These organizations, led by the Recording Academy and the RIAA, played roles in drafting the legislation, advocating for passage, and rallying support among the industry and legislators. The act gained momentum through discussions that bridged industry concerns with legislative action. This collaborative process led to a proposal that specifically targets the use of AI to create unauthorized reproductions of artists' voices and images. == Opposition == The ELVIS Act saw industry opposition from the Motion Picture Association, including testimony in the House Banking & Consumer Affairs Subcommittee, including remarks that the law risks "interference with our members’ ability to portray real people and events." TechNet, representing companies such as OpenAI, Google and Amazon, expressed their opposition in the hearing to the bill as drafted, asserting that the language was too broadly written and could have unintended consequences. Other concerns included its potential application to cover bands, but lawmakers assured people that this was not the intention. The bill passed the Tennessee House and Senate with a unanimous, bi-partisan vote including 93 ayes and 0 Noes in the House, and 30 ayes and 0 noes in the Senate. == Passage == By explicitly addressing AI impersonation, the ELVIS Act originated a legal approach to safeguarding personal rights, in the context of digital and technological advancements. It extends protections to an artist's voice and likeness, areas vulnerable to exploitation with the proliferation of AI technologies that occurred in 2023. The legislation received widespread support from the music industry, signaling a significant step forward in the ongoing effort to balance innovation with the protection of individual rights and creative integrity. It was reported as underscoring Tennessee's commitment to its musical heritage and showed the state as a leader in adapting copyright and privacy protections to the modern technological landscape. Artists including Chris Janson and Luke Bryan appeared at the signing ceremony hosted at Robert's Western World to support the new law and commemorate its passing. == Legal precedent == The ELVIS Act was reported as representing a development in the discourse surrounding AI, intellectual property, and personal rights. It was hoped by proponents to set a precedent for future legislative efforts both within and beyond Tennessee, offering a model for how states and potentially the federal government could address similar challenges. As AI technology continues to evolve, the act represents a foundational framework for protecting the authenticity and rights of artists, ensuring contributions remain protected. The act prohibits usage of AI to clone the voice of an artist without consent and can be criminally enforced as a Class A misdemeanor. This legislation's success was hoped by its supporters to inspire similar actions in other states, contributing to a unified approach to copyright and privacy in the digital age. Such a national response would reinforce the importance of safeguarding artists' rights against unauthorized use of their voices and likenesses.
RagTime
RagTime is a frame-oriented business publishing software which combines word processing, spreadsheets, simple drawings, image processing, and charts, in a single document/program, integrated software. It is often used to create forms, reports, documentation, desktop publishing, and in office environments. Typical users are business clients, educational institutions, administrations, architects, and also private users. Ragtime includes the following modules: Page layout (forms, templates etc.) Word processing Image processing Spreadsheets, similar to Microsoft Excel Formulas and functions which can be used throughout, in text, graphics, and spreadsheets Charts in different types of diagrams Drawings in vector graphics including lines, polygons, Bézier curves and more Slide show (presentation of RagTime documents) Audio/video Buttons (pop-up menus, switches, and more) that can be used within RagTime documents Import/export of various file formats Support of the AppleScript scripting language available system-wide under macOS == Principle == RagTime differs from most other comparable programs or software packages in its strict frame-oriented design: all content is contained within frames on each page. The content can have a fixed position within its frame or, if it is text or a spreadsheet, flow into another frame that is connected to the first frame via a so-called “pipeline”. RagTime has no different document types for different types of data; all content is stored in a single compound document type. Thus, a RagTime document not only can contain multiple pages, but also multiple layouts within the same document; e.g. spreadsheets in addition to text and images. The RagTime filename extension is .rtd (RagTime document); for templates the extension is .rtt (RagTime template). The current version is RagTime 6.6.5. It is available for OS X (10.6-10.14) and Windows (XP/Vista/7/8/10). == Extensions == FileTime – allows accessing “FileMaker Pro” databases from RagTime documents under OS X RagTime Connect – ODBC database connection for RagTime 6 (Mac and Windows) Johannes – print extension for the simple creation of stapled or folded brochures, booklets etc. PowerFunctions – additional functions for a more effective creation of intelligent documents for exchanging data and for use in mixed Mac/Windows environments MetaFormula – SYLK-based extension that allows calculating text as formula == History == RagTime has been developed since 1985 for the Macintosh – originally named MacFrame – and was published in 1986. When released, it already had the present name, which was chosen following the then-available software package Lotus Jazz. In the European Macintosh market, RagTime quickly gained a prominent position that continues to this day, even though the market share has decreased. Despite repeated attempts, the program could not gain acceptance in the North American market due to its high cost ($395 in 1990). The North American sales office closed in 1991, shortly after Claris Corporation released ClarisWorks which duplicated much of the functionality of RagTime for a lower price. After the manufacturer – first Brüning & Everth, followed by B&E Software and today RagTime.de Development – had focused on the Macintosh only for a very long time, it also released a Windows version, RagTime 5.0, in 1999. However, the program could not assume great significance against established competitors, especially Microsoft Office. Until mid-2006 RagTime was, in addition to the commercial version, also available as a free version (RagTime Solo) for personal use. RagTime Solo included the same features and performance (except for spelling and Syllabification) dictionaries), but was not allowed for use in commercial environments. In other languages RagTime Solo was distributed as RagTime Privat. In a press release from July 5, 2006, RagTime announced the discontinuation of RagTime Solo: “… the RagTime Solo license conditions were often misinterpreted or deliberately flouted. Therefore we discontinued RagTime Solo, there will be no private version of RagTime 6 anymore.” After a successful start of the RagTime 6.0 software, sales edged significantly lower in the following years. Disagreements arose among the shareholders about the continuation of the company, which filed for bankruptcy in July 2007. As a result, the rights to RagTime were taken over by the newly established company RagTime.de Development GmbH, which was responsible for the development. The sales partner RagTime.de Sales GmbH distributed the RagTime products until October 2015. Today RagTime.de Development GmbH is also responsible for sales. The last level of development is the extensively revamped version RagTime 6.6 of 8 October 2015, which also includes new OS X features (e.g. high-resolution “Retina” displays) and supports Windows 10. == Programming == RagTime 1-3 were developed in Pascal, since version 4 the development is completely coded in C++. External programming and automation can be implemented via AppleScript on a Mac, and via OLE/COM-API (e.g. Visual Basic) under Windows. On a Mac, RagTime provides a comprehensive AppleScript library, for the automation of almost any task, from automatic document creation to the export of PDF documents. RagTime also supports “recordings” by use of the “AppleScript Editor”, which allows recording the interactive RagTime operation as an AppleScript program sequence. AppleScripts can be saved in the RagTime document and called via menu or shortcut keys. On Windows, RagTime (since version 6) disposes over an OLE/COM API, which allows automating many RagTime components via external programming. For that purpose there is a type library that installs the available RagTime OLE/COM object catalogue. Programming can be realized in all programming languages supported by Microsoft.
Viola–Jones object detection framework
The Viola–Jones object detection framework is a machine learning object detection framework proposed in 2001 by Paul Viola and Michael Jones. It was motivated primarily by the problem of face detection, although it can be adapted to the detection of other object classes. In short, it consists of a sequence of classifiers. Each classifier is a single perceptron with several binary masks (Haar features). To detect faces in an image, a sliding window is computed over the image. For each image, the classifiers are applied. If at any point, a classifier outputs "no face detected", then the window is considered to contain no face. Otherwise, if all classifiers output "face detected", then the window is considered to contain a face. The algorithm is efficient for its time, able to detect faces in 384 by 288 pixel images at 15 frames per second on a conventional 700 MHz Intel Pentium III. It is also robust, achieving high precision and recall. While it has lower accuracy than more modern methods such as convolutional neural network, its efficiency and compact size (only around 50k parameters, compared to millions of parameters for typical CNN like DeepFace) means it is still used in cases with limited computational power. For example, in the original paper, they reported that this face detector could run on the Compaq iPAQ at 2 fps (this device has a low power StrongARM without floating point hardware). == Problem description == Face detection is a binary classification problem combined with a localization problem: given a picture, decide whether it contains faces, and construct bounding boxes for the faces. To make the task more manageable, the Viola–Jones algorithm only detects full view (no occlusion), frontal (no head-turning), upright (no rotation), well-lit, full-sized (occupying most of the frame) faces in fixed-resolution images. The restrictions are not as severe as they appear, as one can normalize the picture to bring it closer to the requirements for Viola-Jones. any image can be scaled to a fixed resolution for a general picture with a face of unknown size and orientation, one can perform blob detection to discover potential faces, then scale and rotate them into the upright, full-sized position. the brightness of the image can be corrected by white balancing. the bounding boxes can be found by sliding a window across the entire picture, and marking down every window that contains a face. This would generally detect the same face multiple times, for which duplication removal methods, such as non-maximal suppression, can be used. The "frontal" requirement is non-negotiable, as there is no simple transformation on the image that can turn a face from a side view to a frontal view. However, one can train multiple Viola-Jones classifiers, one for each angle: one for frontal view, one for 3/4 view, one for profile view, a few more for the angles in-between them. Then one can at run time execute all these classifiers in parallel to detect faces at different view angles. The "full-view" requirement is also non-negotiable, and cannot be simply dealt with by training more Viola-Jones classifiers, since there are too many possible ways to occlude a face. == Components of the framework == A full presentation of the algorithm is in. Consider an image I ( x , y ) {\displaystyle I(x,y)} of fixed resolution ( M , N ) {\displaystyle (M,N)} . Our task is to make a binary decision: whether it is a photo of a standardized face (frontal, well-lit, etc) or not. Viola–Jones is essentially a boosted feature learning algorithm, trained by running a modified AdaBoost algorithm on Haar feature classifiers to find a sequence of classifiers f 1 , f 2 , . . . , f k {\displaystyle f_{1},f_{2},...,f_{k}} . Haar feature classifiers are crude, but allows very fast computation, and the modified AdaBoost constructs a strong classifier out of many weak ones. At run time, a given image I {\displaystyle I} is tested on f 1 ( I ) , f 2 ( I ) , . . . f k ( I ) {\displaystyle f_{1}(I),f_{2}(I),...f_{k}(I)} sequentially. If at any point, f i ( I ) = 0 {\displaystyle f_{i}(I)=0} , the algorithm immediately returns "no face detected". If all classifiers return 1, then the algorithm returns "face detected". For this reason, the Viola-Jones classifier is also called "Haar cascade classifier". === Haar feature classifiers === Consider a perceptron f w , b {\displaystyle f_{w,b}} defined by two variables w ( x , y ) , b {\displaystyle w(x,y),b} . It takes in an image I ( x , y ) {\displaystyle I(x,y)} of fixed resolution, and returns f w , b ( I ) = { 1 , if ∑ x , y w ( x , y ) I ( x , y ) + b > 0 0 , else {\displaystyle f_{w,b}(I)={\begin{cases}1,\quad {\text{if }}\sum _{x,y}w(x,y)I(x,y)+b>0\\0,\quad {\text{else}}\end{cases}}} A Haar feature classifier is a perceptron f w , b {\displaystyle f_{w,b}} with a very special kind of w {\displaystyle w} that makes it extremely cheap to calculate. Namely, if we write out the matrix w ( x , y ) {\displaystyle w(x,y)} , we find that it takes only three possible values { + 1 , − 1 , 0 } {\displaystyle \{+1,-1,0\}} , and if we color the matrix with white on + 1 {\displaystyle +1} , black on − 1 {\displaystyle -1} , and transparent on 0 {\displaystyle 0} , the matrix is in one of the 5 possible patterns shown on the right. Each pattern must also be symmetric to x-reflection and y-reflection (ignoring the color change), so for example, for the horizontal white-black feature, the two rectangles must be of the same width. For the vertical white-black-white feature, the white rectangles must be of the same height, but there is no restriction on the black rectangle's height. ==== Rationale for Haar features ==== The Haar features used in the Viola-Jones algorithm are a subset of the more general Haar basis functions, which have been used previously in the realm of image-based object detection. While crude compared to alternatives such as steerable filters, Haar features are sufficiently complex to match features of typical human faces. For example: The eye region is darker than the upper-cheeks. The nose bridge region is brighter than the eyes. Composition of properties forming matchable facial features: Location and size: eyes, mouth, bridge of nose Value: oriented gradients of pixel intensities Further, the design of Haar features allows for efficient computation of f w , b ( I ) {\displaystyle f_{w,b}(I)} using only constant number of additions and subtractions, regardless of the size of the rectangular features, using the summed-area table. === Learning and using a Viola–Jones classifier === Choose a resolution ( M , N ) {\displaystyle (M,N)} for the images to be classified. In the original paper, they recommended ( M , N ) = ( 24 , 24 ) {\displaystyle (M,N)=(24,24)} . ==== Learning ==== Collect a training set, with some containing faces, and others not containing faces. Perform a certain modified AdaBoost training on the set of all Haar feature classifiers of dimension ( M , N ) {\displaystyle (M,N)} , until a desired level of precision and recall is reached. The modified AdaBoost algorithm would output a sequence of Haar feature classifiers f 1 , f 2 , . . . , f k {\displaystyle f_{1},f_{2},...,f_{k}} . The details of the modified AdaBoost algorithm is detailed below. ==== Using ==== To use a Viola-Jones classifier with f 1 , f 2 , . . . , f k {\displaystyle f_{1},f_{2},...,f_{k}} on an image I {\displaystyle I} , compute f 1 ( I ) , f 2 ( I ) , . . . f k ( I ) {\displaystyle f_{1}(I),f_{2}(I),...f_{k}(I)} sequentially. If at any point, f i ( I ) = 0 {\displaystyle f_{i}(I)=0} , the algorithm immediately returns "no face detected". If all classifiers return 1, then the algorithm returns "face detected". === Learning algorithm === The speed with which features may be evaluated does not adequately compensate for their number, however. For example, in a standard 24x24 pixel sub-window, there are a total of M = 162336 possible features, and it would be prohibitively expensive to evaluate them all when testing an image. Thus, the object detection framework employs a variant of the learning algorithm AdaBoost to both select the best features and to train classifiers that use them. This algorithm constructs a "strong" classifier as a linear combination of weighted simple “weak” classifiers. h ( x ) = sgn ( ∑ j = 1 M α j h j ( x ) ) {\displaystyle h(\mathbf {x} )=\operatorname {sgn} \left(\sum _{j=1}^{M}\alpha _{j}h_{j}(\mathbf {x} )\right)} Each weak classifier is a threshold function based on the feature f j {\displaystyle f_{j}} . h j ( x ) = { − s j if f j < θ j s j otherwise {\displaystyle h_{j}(\mathbf {x} )={\begin{cases}-s_{j}&{\text{if }}f_{j}<\theta _{j}\\s_{j}&{\text{otherwise}}\end{cases}}} The threshold value θ j {\displaystyle \theta _{j}} and the polarity s j ∈ ± 1 {\displaystyle s_{j}\in \pm 1} are determined in the training, as well as the coefficients α j {\displaystyle \alpha _{j}} . Here a simplified version of the lea
Smart data capture
Smart data capture (SDC), also known as 'intelligent data capture' or 'automated data capture', describes the branch of technology concerned with using computer vision techniques like optical character recognition (OCR), barcode scanning, object recognition and other similar technologies to extract and process information from semi-structured and unstructured data sources. IDC characterize smart data capture as an integrated hardware, software, and connectivity strategy to help organizations enable the capture of data in an efficient, repeatable, scalable, and future-proof way. Data is captured visually from barcodes, text, IDs and other objects - often from many sources simultaneously - before being converted and prepared for digital use, typically by artificial intelligence-powered software. An important feature of SDC is that it focuses not just on capturing data more efficiently but serving up easy-to-access, actionable insights at the instant of data collection to both frontline and desk-based workers, aiding decision-making and making it a two-way process. Smart data capture automates and accelerates capture, applying insights in real time and automating processes based on extracted input. Smart data capture is designed to be repeatable and scalable to reduce low-level manual tasks and eliminate human error. To achieve this goal, smart data capture solutions are often made available using specialist software installed on commodity hardware such as smartphones. However, some solutions may rely on specialized hardware such as dedicated scanning devices, wearables or shop floor robots. == Differences from OCR == Optical character recognition applications are typically concerned with the actual data capture process; they are intended to faithfully reproduce text, words, letters and symbols from a printed document. Smart data capture is multimodal, capable of extracting data from a wider range of semi-structured and unstructured sources, going beyond basic text recognition to offer a wider scope of applications. By extending functionality to provide actionable insights at the point of capture, SDC is also a two-way process (capture-display), while OCR is more commonly one-way (capture only), primarily used for data input. Smart data capture solutions typically have two parts: Data capture (which includes OCR, barcode scanning, object recognition) Functionality that then uses this data to provide actionable insights at the point of capture. == Applications == Smart data capture can be applied to almost any industry and application that requires visual information capture and interpretation. This may include: Retail Warehouse inventory control Logistics, handling and shipping Manufacturing Field service Healthcare Transport and travel Fraud detection
Scene text
Scene text is text that appears in an image captured by a camera in an outdoor environment. The detection and recognition of scene text from camera captured images are computer vision tasks which became important after smart phones with good cameras became ubiquitous. The text in scene images varies in shape, font, colour and position. The recognition of scene text is further complicated sometimes by non-uniform illumination and focus. To improve scene text recognition, the International Conference on Document Analysis and Recognition (ICDAR) conducts a robust reading competition once in two years. The competition was held in 2003, 2005 and during every ICDAR conference. International association for pattern recognition (IAPR) has created a list of datasets as Reading systems. == Text detection == Text detection is the process of detecting the text present in the image, followed by surrounding it with a rectangular bounding box. Text detection can be carried out using image based techniques or frequency based techniques. In image based techniques, an image is segmented into multiple segments. Each segment is a connected component of pixels with similar characteristics. The statistical features of connected components are utilised to group them and form the text. Machine learning approaches such as support vector machine and convolutional neural networks are used to classify the components into text and non-text. In frequency based techniques, discrete Fourier transform (DFT) or discrete wavelet transform (DWT) are used to extract the high frequency coefficients. It is assumed that the text present in an image has high frequency components and selecting only the high frequency coefficients filters the text from the non-text regions in an image. == Word recognition == In word recognition, the text is assumed to be already detected and located and the rectangular bounding box containing the text is available. The word present in the bounding box needs to be recognized. The methods available to perform word recognition can be broadly classified into top-down and bottom-up approaches. In the top-down approaches, a set of words from a dictionary is used to identify which word suits the given image. Images are not segmented in most of these methods. Hence, the top-down approach is sometimes referred as segmentation free recognition. In the bottom-up approaches, the image is segmented into multiple components and the segmented image is passed through a recognition engine. Either an off the shelf Optical character recognition (OCR) engine or a custom-trained one is used to recognise the text.
Direct voice input
Direct voice input (DVI), sometimes called voice input control (VIC), is a style of human–machine interaction "HMI" in which the user makes voice commands to issue instructions to the machine through speech recognition. In the field of military aviation, DVI has been introduced into the cockpits of several modern military aircraft, such as the Eurofighter Typhoon, the Lockheed Martin F-35 Lightning II, the Dassault Rafale, the KF-21 Boramae and the Saab JAS 39 Gripen. Such systems have also been used for various other purposes, including industry control systems and speech recognition assistance for impaired individuals. == Overview == DVI systems can be divided into two major categories of functionality: "user-dependent" or "user-independent". A user-dependent system requires that a personal voice template to be generated for a specific person; the template for this individual has to be loaded onto their assigned machine prior to use of the DVI system for it to function properly. In contrast, a user-independent system does not require any personal voice template, being intended to respond correctly to the voice of any user. They can also be categorised between "discrete recognition" and "continuous recognition". Users of a discrete recognition system must pause between each word so that the DVI system can identify the separations between each word, while a continuous speech recognition system is capable of understanding a normal rate of speech. During the mid-2000s, researchers at the National Aerospace Laboratory in the Netherlands examined the use of DVI in the "GRACE" simulator; a total of twelve pilots participated in the ensuing experiment. The tests performed reportedly revealed that, while the hardware itself functioned well, several improvements were desirable prior to real-world deployment on aircraft since DVI operations actually consumed more time in comparison to traditional existing methods. Recommendations for improvements included the adoption of simpler syntax, the achievement of a greater recognition rate, and a decrease in response times; all of the issues encountered were determined to be of a technological nature, and were deemed feasible to resolve. The researchers concluded that in cockpits, especially during emergencies where pilots have to operate entirely on their own, a DVI system could be highly relevant, but that it was not of crucial importance during most other conceivable scenarios. Around the same time, evaluations of DVI systems for civil aviation purposes were conducted within the framework of Project SafeSound, coordinated by the European Union. It involved the observation of pilot workloads in real-world cockpits and contrasting them against pilot activity in flight simulators using both conventional systems and DVI assistance. The project aimed to enhance aviation safety and to decrease the workload in both ground and flight operations via the application of enhanced audio functions. == Applications == === Aviation === Prior to its widespread deployment, a handful of conventional military aircraft were converted to trial DVI systems; examples include the Harrier AV-8B and F-16 VISTA. In another case, a General Dynamics F-16 Fighting Falcon simulator was modified with DVI for a voice control study that was undertaken by the Royal Netherlands Air Force. DVI trials have also been conducted on helicopters, including the Boeing AH-64 Apache, showing the potential to improve flight safety and mission effectiveness. Numerous modern fighter aircraft have been outfitted with DVI systems, often in combination with various other man-machine interface schemes, such as HOTAS-compliant controls and other advanced control technologies. The combination of Voice and HOTAS control schemes has sometimes been referred to as the "V-TAS" concept. A prominent fighter aircraft to be furnished with a V-TAS cockpit is the Eurofighter Typhoon. The Lockheed Martin F-35 Lightning II also features a DVI system, which was developed by Adacel. Other examples includes the Dassault Rafale and the Saab JAS 39 Gripen. Numerous aircraft have been planned to use DVI. At one stage, the United States Air Force had sought to integrate DVI upon the Lockheed Martin F-22 Raptor; however, the technology was eventually judged to pose too many technical risks at that point in time, and thus such efforts were abandoned. === Personal === By 1990, working prototypes of speech recognition systems were being demonstrated; these were being promoted for the purpose of providing an effective man-machine interface for individuals with impaired speech. Techniques employed included time-encoded digital speech and automatic token set selection. Investigations of these early DVI systems reportedly included the use of automatic diagnostic routines and limited-scale trials using volunteers. During the 2010s, various companies were offering voice recognition systems to the general public in the form of personal digital assistants. One example is the Google Voice service, which allows users to pose questions via a DVI package installed on either a personal computer, tablet, or mobile phone. Numerous digital assistants have been developed, such as Amazon Echo, Siri, and Cortana, that use DVI to interact with users. === Commercial === DVI technology has enabled automated telephone systems to be widely deployed. Many companies commonly use centralised phone systems that route callers to the correct department via such methods. Various car manufacturers have also furnished their road vehicles with DVI systems; these typically allow drivers to control infotainment systems and interact with mobile phones with more convenience than legacy methods. During the late 1980s, investigations into the use of DVI systems for controlling CNC machines and other manufacturing apparatus were underway. During the 2010s, such systems were being used for logistics and warehouse management purposes.
Chinchilla (language model)
Chinchilla is a family of large language models (LLMs) developed by the research team at Google DeepMind, presented in March 2022. == Models == It is named "chinchilla" because it is a further development over a previous model family named Gopher. Both model families were trained in order to investigate the scaling laws of large language models. It claimed to outperform GPT-3. It considerably simplifies downstream utilization because it requires much less computer power for inference and fine-tuning. Based on the training of previously employed language models, it has been determined that if one doubles the model size, one must also have twice the number of training tokens. This hypothesis has been used to train Chinchilla by DeepMind. Similar to Gopher in terms of cost, Chinchilla has 70B parameters and four times as much data. Chinchilla has an average accuracy of 67.5% on the Measuring Massive Multitask Language Understanding (MMLU) benchmark, which is 7% higher than Gopher's performance. Chinchilla was still in the testing phase as of January 12, 2023. Chinchilla contributes to developing an effective training paradigm for large autoregressive language models with limited compute resources. The Chinchilla team recommends that the number of training tokens is twice for every model size doubling, meaning that using larger, higher-quality training datasets can lead to better results on downstream tasks. It has been used for the Flamingo vision-language model. == Architecture == Both the Gopher family and Chinchilla family are families of transformer models. In particular, they are essentially the same as GPT-2, with different sizes and minor modifications. Gopher family uses RMSNorm instead of LayerNorm; relative positional encoding rather than absolute positional encoding. The Chinchilla family is the same as the Gopher family, but trained with AdamW instead of Adam optimizer. The Gopher family contains six models of increasing size, from 44 million parameters to 280 billion parameters. They refer to the largest one as "Gopher" by default. Similar naming conventions apply for the Chinchilla family. Table 1 of shows the entire Gopher family: Table 4 of compares the 70-billion-parameter Chinchilla with Gopher 280B.