Small Data: the Tiny Clues that Uncover Huge Trends is Martin Lindstrom's seventh book. It chronicles his work as a branding expert, working with consumers across the world to better understand their behavior. The theory behind the book is that businesses can better create products and services based on observing consumer behavior in their homes, as opposed to relying solely on big data. == Content == The book is based on a several year period of consumer studies for major corporations across the globe. It features case studies of the author's work interviewing consumers in their homes and using his observations to create hypotheses as to why they use products the way that they do. == Public reception == The book was a New York Times Bestseller upon release and was positively reviewed on several websites, Including Entrepreneur and Forbes. In 2016, it was named a Best Business Book by strategy+business and one of Inc. Magazine's Best Sales and Marketing books.
Automotive security
Automotive security refers to the branch of computer security focused on the cyber risks related to the automotive context. The increasingly high number of ECUs in vehicles and, alongside, the implementation of multiple different means of communication from and towards the vehicle in a remote and wireless manner led to the necessity of a branch of cybersecurity dedicated to the threats associated with vehicles. Not to be confused with automotive safety. == Causes == The implementation of multiple ECUs (Electronic Control Units) inside vehicles began in the early '70s thanks to the development of integrated circuits and microprocessors that made it economically feasible to produce the ECUs on a large scale. Since then the number of ECUs has increased to up to 100 per vehicle. These units nowadays control almost everything in the vehicle, from simple tasks such as activating the wipers to more safety-related ones like brake-by-wire or ABS (Anti-lock Braking System). Autonomous driving is also strongly reliant on the implementation of new, complex ECUs such as the ADAS, alongside sensors (lidars and radars) and their control units. Inside the vehicle, the ECUs are connected with each other through cabled or wireless communication networks, such as CAN bus (controller area network), MOST bus (Media Oriented System Transport), FlexRay (Automotive Network Communications Protocol) or RF (radio frequency) as in many implementations of TPMSs (tire-pressure monitoring systems). Many of these ECUs require data received through these networks that arrive from various sensors to operate and use such data to modify the behavior of the vehicle (e.g., the cruise control modifies the vehicle's speed depending on signals arriving from a button usually located on the steering wheel). Since the development of cheap wireless communication technologies such as Bluetooth, LTE, Wi-Fi, RFID and similar, automotive producers and OEMs have designed ECUs that implement such technologies with the goal of improving the experience of the driver and passengers. Safety-related systems such as the OnStar from General Motors, telematic units, communication between smartphones and the vehicle's speakers through Bluetooth, Android Auto and Apple CarPlay. == Threat model == Threat models of the automotive world are based on both real-world and theoretically possible attacks. Most real-world attacks aim at the safety of the people in and around the car, by modifying the cyber-physical capabilities of the vehicle (e.g., steering, braking, accelerating without requiring actions from the driver), while theoretical attacks have been supposed to focus also on privacy-related goals, such as obtaining GPS data on the vehicle, or capturing microphone signals and similar. Regarding the attack surfaces of the vehicle, they are usually divided in long-range, short-range, and local attack surfaces: LTE and DSRC can be considered long-range ones, while Bluetooth and Wi-Fi are usually considered short-range although still wireless. Finally, USB, OBD-II and all the attack surfaces that require physical access to the car are defined as local. An attacker that is able to implement the attack through a long-range surface is considered stronger and more dangerous than the one that requires physical access to the vehicle. In 2015 the possibility of attacks on vehicles already on the market has been proven possible by Miller and Valasek, that managed to disrupt the driving of a Jeep Cherokee while remotely connecting to it through remote wireless communication. === Controller area network attacks === The most common network used in vehicles and the one that is mainly used for safety-related communication is CAN, due to its real-time properties, simplicity, and cheapness. For this reason the majority of real-world attacks have been implemented against ECUs connected through this type of network. The majority of attacks demonstrated either against actual vehicles or in testbeds fall in one or more of the following categories: ==== Sniffing ==== Sniffing in the computer security field generally refers to the possibility of intercepting and logging packets or more generally data from a network. In the case of CAN, since it is a bus network, every node listens to all communication on the network. It is useful for the attacker to read data to learn the behavior of the other nodes of the network before implementing the actual attack. Usually, the final goal of the attacker is not to simply sniff the data on CAN, since the packets passing on this type of network are not usually valuable just to read. ==== Denial of service ==== Denial of service (DoS) in information security is usually described as an attack that has the objective of making a machine or a network unavailable. DoS attacks against ECUs connected to CAN buses can be done both against the network, by abusing the arbitration protocol used by CAN to always win the arbitration, and targeting the single ECU, by abusing the error handling protocol of CAN. In this second case the attacker flags the messages of the victim as faulty to convince the victim of being broken and therefore shut itself off the network. ==== Spoofing ==== Spoofing attacks comprise all cases in which an attacker, by falsifying data, sends messages pretending to be another node of the network. In automotive security usually spoofing attacks are divided into masquerade and replay attacks. Replay attacks are defined as all those where the attacker pretends to be the victim and sends sniffed data that the victim sent in a previous iteration of authentication. Masquerade attacks are, on the contrary, spoofing attacks where the data payload has been created by the attacker. == Real life automotive threat example == Security researchers Charlie Miller and Chris Valasek have successfully demonstrated remote access to a wide variety of vehicle controls using a Jeep Cherokee as the target. They were able to control the radio, environmental controls, windshield wipers, and certain engine and brake functions. The method used to hack the system was implementation of pre-programmed chip into the controller area network (CAN) bus. By inserting this chip into the CAN bus, he was able to send arbitrary message to CAN bus. One other thing that Miller has pointed out is the danger of the CAN bus, as it broadcasts the signal which the message can be caught by the hackers throughout the network. The control of the vehicle was all done remotely, manipulating the system without any physical interaction. Miller states that he could control any of some 1.4 million vehicles in the United States regardless of the location or distance, the only thing needed is for someone to turn on the vehicle to gain access. The work by Miller and Valasek replicated earlier work completed and published by academics in 2010 and 2011 on a different vehicle. The earlier work demonstrated the ability to compromise a vehicle remotely, over multiple wireless channels (including cellular), and the ability to remotely control critical components on the vehicle post-compromise, including the telematics unit and the car's brakes. While the earlier academic work was publicly visible, both in peer-reviewed scholarly publications and in the press, the Miller and Valesek work received even greater public visibility. == Security measures == The increasing complexity of devices and networks in the automotive context requires the application of security measures to limit the capabilities of a potential attacker. Since the early 2000 many different countermeasures have been proposed and, in some cases, applied. Following, a list of the most common security measures: Sub-networks: to limit the attacker capabilities even if he/she manages to access the vehicle from remote through a remotely connected ECU, the networks of the vehicle are divided in multiple sub-networks, and the most critical ECUs are not placed in the same sub-networks of the ECUs that can be accessed from remote. Gateways: the sub-networks are divided by secure gateways or firewalls that block messages from crossing from a sub-network to the other if they were not intended to. Intrusion Detection Systems (IDS): on each critical sub-network, one of the nodes (ECUs) connected to it has the goal of reading all data passing on the sub-network and detect messages that, given some rules, are considered malicious (made by an attacker). The arbitrary messages can be caught by the passenger by using IDS which will notify the owner regarding with unexpected message. Authentication protocols: in order to implement authentication on networks where it is not already implemented (such as CAN), it is possible to design an authentication protocol that works on the higher layers of the ISO OSI model, by using part of the data payload of a message to authenticate the message itself. Hardware Security Modules: since many ECUs are not powerful enough to keep real-time delays whi
Ernie Bot
Ernie Bot (Chinese: 文心一言, Pinyin: wénxīn yīyán), full name Enhanced Representation through Knowledge Integration, is an artificial intelligence chatbot developed by the Chinese technology company Baidu. Ernie Bot rivals GPT models in Chinese NLP tasks. It is built on the company's ERNIE series of large language models, which have been in development since 2019. The service was first launched for invited testing on March 16, 2023, and was released to the general public on August 31, 2023, after receiving approval from Chinese regulators. Since its public launch, Ernie Bot has undergone several updates, with newer versions like ERNIE 4.0 and 4.5 released to improve its capabilities. The service has seen rapid user adoption, reportedly reaching over 200 million users by April 2024. It has been integrated into various products, notably powering AI features for the Chinese release of Samsung's Galaxy S24 smartphones. As a product operating in China, Ernie Bot is subject to the country's censorship regulations. It has been observed to refuse answers to politically sensitive questions, such as those regarding CCP general secretary Xi Jinping, the 1989 Tiananmen Square protests and massacre, and other topics deemed taboo by the government. == History == Ernie Bot was initially released for invited testing on March 16, 2023. The live release demo was reported to have been prerecorded, which caused Baidu's stock to drop 10 percent on the day of the launch. The company's stock gained 14 percent the following day after analysts from Citigroup and Bank of America tested Ernie Bot and gave it positive preliminary reviews. On August 31, 2023, Ernie Bot was released to the public after receiving approval from Chinese regulatory authorities. By December 2023, Baidu announced the service had surpassed 100 million users. In January 2024, Hong Kong newspaper South China Morning Post reported that a university research lab linked to the People's Liberation Army (PLA) had tested Ernie Bot for military response scenarios. Baidu denied the allegations, stating it had no connection with the academic paper. That same month, Ernie was integrated into Samsung's Galaxy S24 lineup for its launch in China. The user base reportedly grew to 200 million by April 2024 and 300 million by June 2024. In September 2024, Baidu changed the chatbot's Chinese name from "Wenxin Yiyan" (文心一言) to "Wenxiaoyan" (文小言) to position it as a search assistant. On March 16, 2025, Baidu announced version 4.5 and the reasoning model ERNIE X1. The following month, at the Create2025 Baidu AI Developer Conference, the company released the Wenxin 4.5 Turbo and Wenxin X1 Turbo models, designed to be faster and less expensive to operate. == Development == Ernie Bot is based on Baidu's ERNIE (Enhanced Representation through Knowledge Integration) series of foundation models. The general training process begins with pre-training on large datasets, followed by refinement using techniques like supervised fine-tuning, reinforcement learning with human feedback, and prompt engineering. === Foundation models === ==== Ernie 3.0 ==== The model powering the initial launch of Ernie Bot. It was trained with 10 billion parameters on a 4-terabyte corpus consisting of plain text and a large-scale knowledge graph. ==== Ernie 3.5 ==== Released in June 2023. At the time of release, its performance was reported as "slightly inferior" to OpenAI's GPT-4. ==== Ernie 4.0 ==== Unveiled in October 2023 and released to paying subscribers in November. According to Baidu, this version featured improved performance over its predecessor, with information updated to April 2023. ==== Ernie X1 ==== Announced in March 2025, with Ernie X1 positioned as a specialized reasoning model. Baidu stated that performance improvements were achieved through new technologies such as "FlashMask" dynamic attention masking and a heterogeneous multimodal mixture-of-experts architecture. === Turbo Models === In June 2024, Baidu announced Ernie 4.0 Turbo. In April 2025, Ernie 4.5 Turbo and X1 Turbo were released. These models are optimized for faster response times and lower operational costs. == Service == In its subscription options, the professional plan gives users access to Ernie 4.0 with a payment either for a month or with reduced payment for auto-renewal per month. Meanwhile, Ernie 3.5 is free of charge. Ernie 4.0, the language model for Ernie bot, has information updated to April 2023. == Censorship == Ernie Bot is subject to the Chinese government's censorship regime. In public tests with journalists, Ernie Bot refused to answer questions about CCP general secretary Xi Jinping, the 1989 Tiananmen Square protests and massacre, the persecution of Uyghurs in China in Xinjiang, and the 2019–2020 Hong Kong protests. When queried about the origin of SARS-CoV-2, Ernie Bot stated that it originated among American vape users.
Image analysis
Image analysis or imagery analysis is the extraction of meaningful information from images; mainly from digital images by means of digital image processing techniques. Image analysis tasks can be as simple as reading bar coded tags or as sophisticated as identifying a person from their face. Computers are indispensable for the analysis of large amounts of data, for tasks that require complex computation, or for the extraction of quantitative information. On the other hand, the human visual cortex is an excellent image analysis apparatus, especially for extracting higher-level information, and for many applications — including medicine, security, and remote sensing — human analysts still cannot be replaced by computers. For this reason, many important image analysis tools such as edge detectors and neural networks are inspired by human visual perception models. == Digital == Digital Image Analysis or Computer Image Analysis is when a computer or electrical device automatically studies an image to obtain useful information from it. Note that the device is often a computer but may also be an electrical circuit, a digital camera or a mobile phone. It involves the fields of computer or machine vision, and medical imaging, and makes heavy use of pattern recognition, digital geometry, and signal processing. This field of computer science developed in the 1950s at academic institutions such as the MIT A.I. Lab, originally as a branch of artificial intelligence and robotics. It is the quantitative or qualitative characterization of two-dimensional (2D) or three-dimensional (3D) digital images. 2D images are, for example, to be analyzed in computer vision, and 3D images in medical imaging. The field was established in the 1950s—1970s, for example with pioneering contributions by Azriel Rosenfeld, Herbert Freeman, Jack E. Bresenham, or King-Sun Fu. == Techniques == There are many different techniques used in automatically analysing images. Each technique may be useful for a small range of tasks, however there still aren't any known methods of image analysis that are generic enough for wide ranges of tasks, compared to the abilities of a human's image analysing capabilities. Examples of image analysis techniques in different fields include: 2D and 3D object recognition, image segmentation, motion detection e.g. Single particle tracking, video tracking, optical flow, medical scan analysis, 3D Pose Estimation. == Deep learning == Since the early 2010s, deep learning methods have substantially advanced the field of image analysis. In 2012, a deep convolutional neural network (CNN) known as AlexNet achieved a significant reduction in error rates on the ImageNet large-scale image classification benchmark, demonstrating the effectiveness of deep learning for visual recognition tasks. Subsequent architectures such as ResNet introduced residual connections that enabled training of much deeper networks, further improving accuracy across image analysis tasks. Real-time object detection became practical with frameworks such as YOLO (You Only Look Once), which unified detection and classification into a single network pass. In 2020, the Vision Transformer (ViT) demonstrated that transformer architectures, originally developed for natural language processing, could achieve competitive results on image classification when applied directly to sequences of image patches. More recently, foundation models trained on large-scale datasets have enabled zero-shot generalisation across image analysis tasks. The Segment Anything Model (SAM), trained on over one billion masks, can segment arbitrary objects in images without task-specific fine-tuning. These advances have made image analysis techniques increasingly accessible through browser-based tools and open-source implementations. == Applications == The applications of digital image analysis are continuously expanding through all areas of science and industry, including: anatomy, allows for precise measurements, visualization, and statistical analysis of anatomical structures. assay micro plate reading, such as detecting where a chemical was manufactured. astronomy, such as calculating the size of a planet. automated species identification (e.g. plant and animal species) defense error level analysis filtering machine vision, such as to automatically count items in a factory conveyor belt. materials science, such as determining if a metal weld has cracks. medicine, such as detecting cancer in a mammography scan. metallography, such as determining the mineral content of a rock sample. microscopy, such as counting the germs in a swab. automatic number plate recognition; optical character recognition, such as automatic license plate detection. remote sensing, such as detecting intruders in a house, and producing land cover/land use maps. robotics, such as to avoid steering into an obstacle. security, such as detecting a person's eye color or hair color. == Object-based == Object-based image analysis (OBIA) involves two typical processes, segmentation and classification. Segmentation helps to group pixels into homogeneous objects. The objects typically correspond to individual features of interest, although over-segmentation or under-segmentation is very likely. Classification then can be performed at object levels, using various statistics of the objects as features in the classifier. Statistics can include geometry, context and texture of image objects. Over-segmentation is often preferred over under-segmentation when classifying high-resolution images. Object-based image analysis has been applied in many fields, such as cell biology, medicine, earth sciences, and remote sensing. For example, it can detect changes of cellular shapes in the process of cell differentiation.; it has also been widely used in the mapping community to generate land cover. When applied to earth images, OBIA is known as geographic object-based image analysis (GEOBIA), defined as "a sub-discipline of geoinformation science devoted to (...) partitioning remote sensing (RS) imagery into meaningful image-objects, and assessing their characteristics through spatial, spectral and temporal scale". The international GEOBIA conference has been held biannually since 2006. OBIA techniques are implemented in software such as eCognition or the Orfeo toolbox.
Glossary of machine vision
The following are common definitions related to the machine vision field. General related fields Machine vision Computer vision Image processing Signal processing == 0-9 == 1394. FireWire is Apple Inc.'s brand name for the IEEE 1394 interface. It is also known as i.Link (Sony's name) or IEEE 1394 (although the 1394 standard also defines a backplane interface). It is a personal computer (and digital audio/digital video) serial bus interface standard, offering high-speed communications and isochronous real-time data services. 1D. One-dimensional. 2D computer graphics. The computer-based generation of digital images—mostly from two-dimensional models (such as 2D geometric models, text, and digital images) and by techniques specific to them. 3D computer graphics. 3D computer graphics are different from 2D computer graphics in that a three-dimensional representation of geometric data is stored in the computer for the purposes of performing calculations and rendering 2D images. Such images may be for later display or for real-time viewing. Despite these differences, 3D computer graphics rely on many of the same algorithms as 2D computer vector graphics in the wire frame model and 2D computer raster graphics in the final rendered display. In computer graphics software, the distinction between 2D and 3D is occasionally blurred; 2D applications may use 3D techniques to achieve effects such as lighting, and primarily 3D may use 2D rendering techniques. 3D scanner. This is a device that analyzes a real-world object or environment to collect data on its shape and possibly color. The collected data can then be used to construct digital, three dimensional models useful for a wide variety of applications. == A == Aberration. Optically, defocus refers to a translation along the optical axis away from the plane or surface of best focus. In general, defocus reduces the sharpness and contrast of the image. What should be sharp, high-contrast edges in a scene become gradual transitions. Algebraic distance or algebraic error. The algebraic distance from a point xi to a curve or surface defined by f ( x , β ) = 0 {\displaystyle f(x,\beta )=0} is the value of f ( x i , β ) {\displaystyle f(x_{i},\beta )} , i.e. the residual in the least squares problem with data point (xi, 0) and model function f. This term is mainly used in computer vision.[1][2] Aperture. In context of photography or machine vision, aperture refers to the diameter of the aperture stop of a photographic lens. The aperture stop can be adjusted to control the amount of light reaching the film or image sensor. aspect ratio (image). The aspect ratio of an image is its displayed width divided by its height (usually expressed as "x:y"). Angular resolution. Describes the resolving power of any image forming device such as an optical or radio telescope, a microscope, a camera, or an eye. Automated optical inspection. == B == Barcode. A barcode (also bar code) is a machine-readable representation of information in a visual format on a surface. Blob discovery. Inspecting an image for discrete blobs of connected pixels (e.g. a black hole in a grey object) as image landmarks. These blobs frequently represent optical targets for machining, robotic capture, or manufacturing failure. Bitmap. A raster graphics image, digital image, or bitmap, is a data file or structure representing a generally rectangular grid of pixels, or points of color, on a computer monitor, paper, or other display device. == C == Camera. A camera is a device used to take pictures, either singly or in sequence. A camera that takes pictures singly is sometimes called a photo camera to distinguish it from a video camera. Camera Link. Camera Link is a serial communication protocol designed for computer vision applications based on the National Semiconductor interface Channel-link. It was designed for the purpose of standardizing scientific and industrial video products including cameras, cables and frame grabbers. The standard is maintained and administered by the Automated Imaging Association, or AIA, the global machine vision industry's trade group. Charge-coupled device. A charge-coupled device (CCD) is a sensor for recording images, consisting of an integrated circuit containing an array of linked, or coupled, capacitors. CCD sensors and cameras tend to be more sensitive, less noisy, and more expensive than CMOS sensors and cameras. CIE 1931 Color Space. In the study of the perception of color, one of the first mathematically defined color spaces was the CIE XYZ color space (also known as CIE 1931 color space), created by the International Commission on Illumination (CIE) in 1931. CMOS. CMOS ("see-moss")stands for complementary metal-oxide semiconductor, is a major class of integrated circuits. CMOS imaging sensors for machine vision are cheaper than CCD sensors but more noisy. CoaXPress. CoaXPress (CXP) is an asymmetric high speed serial communication standard over coaxial cable. CoaXPress combines high speed image data, low speed camera control and power over a single coaxial cable. The standard is maintained by JIIA, the Japan Industrial Imaging Association. Color. The perception of the frequency (or wavelength) of light, and can be compared to how pitch (or a musical note) is the perception of the frequency or wavelength of sound. Color blindness. Also known as color vision deficiency, in humans is the inability to perceive differences between some or all colors that other people can distinguish Color temperature. "White light" is commonly described by its color temperature. A traditional incandescent light source's color temperature is determined by comparing its hue with a theoretical, heated black-body radiator. The lamp's color temperature is the temperature in kelvins at which the heated black-body radiator matches the hue of the lamp. Color vision. CV is the capacity of an organism or machine to distinguish objects based on the wavelengths (or frequencies) of the light they reflect or emit. computer vision. The study and application of methods which allow computers to "understand" image content. Contrast. In visual perception, contrast is the difference in visual properties that makes an object (or its representation in an image) distinguishable from other objects and the background. C-Mount. Standardized adapter for optical lenses on CCD - cameras. C-Mount lenses have a back focal distance 17.5 mm vs. 12.5 mm for "CS-mount" lenses. A C-Mount lens can be used on a CS-Mount camera through the use of a 5 mm extension adapter. C-mount is a 1" diameter, 32 threads per inch mounting thread (1"-32UN-2A.) CS-Mount. Same as C-Mount but the focal point is 5 mm shorter. A CS-Mount lens will not work on a C-Mount camera. CS-mount is a 1" diameter, 32 threads per inch mounting thread. == D == Data matrix. A two dimensional Barcode. Depth of field. In optics, particularly photography and machine vision, the depth of field (DOF) is the distance in front of and behind the subject which appears to be in focus. Depth perception. DP is the visual ability to perceive the world in three dimensions. It is a trait common to many higher animals. Depth perception allows the beholder to accurately gauge the distance to an object. Diaphragm. In optics, a diaphragm is a thin opaque structure with an opening (aperture) at its centre. The role of the diaphragm is to stop the passage of light, except for the light passing through the aperture. == E == Edge detection. ED marks the points in a digital image at which the luminous intensity changes sharply. It also marks the points of luminous intensity changes of an object or spatial-taxon silhouette. Electromagnetic interference. Radio Frequency Interference (RFI) is electromagnetic radiation which is emitted by electrical circuits carrying rapidly changing signals, as a by-product of their normal operation, and which causes unwanted signals (interference or noise) to be induced in other circuits. == F == FireWire. FireWire (also known as i. Link or IEEE 1394) is a personal computer (and digital audio/video) serial bus interface standard, offering high-speed communications. It is often used as an interface for industrial cameras. Fixed-pattern noise. Flat-field correction. Frame grabber. An electronic device that captures individual, digital still frames from an analog video signal or a digital video stream. Fringe Projection Technique. 3D data acquisition technique employing projector displaying fringe pattern on a surface of measured piece, and one or more cameras recording image(s). Field of view. The field of view (FOV) is the part which can be seen by the machine vision system at one moment. The field of view depends from the lens of the system and from the working distance between object and camera. Focus. An image, or image point or region, is said to be in focus if light from object points is converged about as well as possible in the image; conversely, it is out of focus if light is not w
Language model benchmark
A language model benchmark is a standardized test designed to evaluate the performance of language models on various natural language processing tasks. These tests are intended for comparing different models' capabilities in areas such as language understanding, generation, and reasoning. Benchmarks generally consist of a dataset and corresponding evaluation metrics. The dataset provides text samples and annotations, while the metrics measure a model's performance on tasks like answering questions, text classification, and machine translation. These benchmarks are developed and maintained by academic institutions, research organizations, and industry players to track progress in the field. In addition to accuracy, the metrics can include throughput, energy efficiency, bias, trust, and sustainability. == Overview == === Types === Benchmarks may be described by the following adjectives, not mutually exclusive: Classical: These tasks are studied in natural language processing, even before the advent of deep learning. Examples include the Penn Treebank for testing syntactic and semantic parsing, as well as bilingual translation benchmarked by BLEU scores. Question answering: These tasks have a text question and a text answer, often multiple-choice. They can be open-book or closed-book. Open-book QA resembles reading comprehension questions, with relevant passages included as annotation in the question, in which the answer appears. Closed-book QA includes no relevant passages. Closed-book QA is also called open-domain question-answering. Before the era of large language models, open-book QA was more common, and understood as testing information retrieval methods. Closed-book QA became common since GPT-2 as a method to measure knowledge stored within model parameters. Omnibus: An omnibus benchmark combines many benchmarks, often previously published. It is intended as an all-in-one benchmarking solution. Reasoning: These tasks are usually in the question-answering format, but are intended to be more difficult than standard question answering. Multimodal: These tasks require processing not only text, but also other modalities, such as images and sound. Examples include OCR and transcription. Agency: These tasks are for a language-model–based software agent that operates a computer for a user, such as editing images, browsing the web, etc. Adversarial: A benchmark is "adversarial" if the items in the benchmark are picked specifically so that certain models do badly on them. Adversarial benchmarks are often constructed after state of the art (SOTA) models have saturated (achieved 100% performance) a benchmark, to renew the benchmark. A benchmark is "adversarial" only at a certain moment in time, since what is adversarial may cease to be adversarial as newer SOTA models appear. Public/Private: A benchmark might be partly or entirely private, meaning that some or all of the questions are not publicly available. The idea is that if a question is publicly available, then it might be used for training, which would be "training on the test set" and invalidate the result of the benchmark. Usually, only the guardians of the benchmark have access to the private subsets, and to score a model on such a benchmark, one must send the model weights, or provide API access, to the guardians. The boundary between a benchmark and a dataset is not sharp. Generally, a dataset contains three "splits": training, test, and validation. Both the test and validation splits are essentially benchmarks. In general, a benchmark is distinguished from a test/validation dataset in that a benchmark is typically intended to be used to measure the performance of many different models that are not trained specifically for doing well on the benchmark, while a test/validation set is intended to be used to measure the performance of models trained specifically on the corresponding training set. In other words, a benchmark may be thought of as a test/validation set without a corresponding training set. Conversely, certain benchmarks may be used as a training set, such as the English Gigaword or the One Billion Word Benchmark, which in modern language is just the negative log-likelihood loss on a pretraining set with 1 billion words. Indeed, the distinction between benchmark and dataset in language models became sharper after the rise of the pretraining paradigm, whereby a model is first trained on massive, unlabeled datasets to learn general language patterns, syntax, and knowledge (pretraining), and the base model is then adapted to specific, downstream tasks using smaller, labeled datasets (fine-tuning). === Lifecycle === Generally, the life cycle of a benchmark consists of the following steps: Inception: A benchmark is published. It can be simply given as a demonstration of the power of a new model (implicitly) that others then picked up as a benchmark, or as a benchmark that others are encouraged to use (explicitly). Growth: More papers and models use the benchmark, and the performance on the benchmark grows. Maturity, degeneration or deprecation: A benchmark may be saturated, after which researchers move on to other benchmarks. Progress on the benchmark may also be neglected as the field moves to focus on other benchmarks. Renewal: A saturated benchmark can be upgraded to make it no longer saturated, allowing further progress. === Construction === Like datasets, benchmarks are typically constructed by several methods, individually or in combination: Web scraping: Ready-made question-answer pairs may be scraped online, such as from websites that teach mathematics and programming. Conversion: Items may be constructed programmatically from scraped web content, such as by blanking out named entities from sentences, and asking the model to fill in the blank. This was used for making the CNN/Daily Mail Reading Comprehension Task. Crowd sourcing: Items may be constructed by paying people to write them, such as on Amazon Mechanical Turk. This was used for making the MCTest. === Evaluation === Generally, benchmarks are fully automated. This limits the questions that can be asked. For example, with mathematical questions, "proving a claim" would be difficult to automatically check, while "calculate an answer with a unique integer answer" would be automatically checkable. With programming tasks, the answer can generally be checked by running unit tests, with an upper limit on runtime. The benchmark scores are of the following kinds: For multiple choice or cloze questions, common scores are accuracy (frequency of correct answer), precision, recall, F1 score, etc. pass@n: The model is given n {\displaystyle n} attempts to solve each problem. If any attempt is correct, the model earns a point. The pass@n score is the model's average score over all problems. k@n: The model makes n {\displaystyle n} attempts to solve each problem, but only k {\displaystyle k} attempts out of them are selected for submission. If any submission is correct, the model earns a point. The k@n score is the model's average score over all problems. cons@n: The model is given n {\displaystyle n} attempts to solve each problem. If the most common answer is correct, the model earns a point. The cons@n score is the model's average score over all problems. Here "cons" stands for "consensus" or "majority voting". The pass@n score can be estimated more accurately by making N > n {\displaystyle N>n} attempts, and use the unbiased estimator 1 − ( N − c n ) ( N n ) {\displaystyle 1-{\frac {\binom {N-c}{n}}{\binom {N}{n}}}} , where c {\displaystyle c} is the number of correct attempts. For less well-formed tasks, where the output can be any sentence, there are the following commonly used scores including BLEU ROUGE, METEOR, NIST, word error rate, LEPOR, CIDEr, and SPICE. === Issues === error: Some benchmark answers may be wrong. ambiguity: Some benchmark questions may be ambiguously worded. subjective: Some benchmark questions may not have an objective answer at all. This problem generally prevents creative writing benchmarks. Similarly, this prevents benchmarking writing proofs in natural language, though benchmarking proofs in a formal language is possible. open-ended: Some benchmark questions may not have a single answer of a fixed size. This problem generally prevents programming benchmarks from using more natural tasks such as "write a program for X", and instead uses tasks such as "write a function that implements specification X". inter-annotator agreement: Some benchmark questions may be not fully objective, such that even people would not agree with 100% on what the answer should be. This is common in natural language processing tasks, such as syntactic annotation. shortcut: Some benchmark questions may be easily solved by an "unintended" shortcut. For example, in the SNLI benchmark, having a negative word like "not" in the second sentence is a strong signal for the "Contradiction" category, regardless of what the se
Aggregation (linguistics)
In linguistics, aggregation is a subtask of natural language generation, which involves merging syntactic constituents (such as sentences and phrases) together. Sometimes aggregation can be done at a conceptual level. == Examples == A simple example of syntactic aggregation is merging the two sentences John went to the shop and John bought an apple into the single sentence John went to the shop and bought an apple. Syntactic aggregation can be much more complex than this. For example, aggregation can embed one of the constituents in the other; e.g., we can aggregate John went to the shop and The shop was closed into the sentence John went to the shop, which was closed. From a pragmatic perspective, aggregating sentences together often suggests to the reader that these sentences are related to each other. If this is not the case, the reader may be confused. For example, someone who reads John went to the shop and bought an apple may infer that the apple was bought in the shop; if this is not the case, then these sentences should not be aggregated. == Algorithms and issues == Aggregation algorithms must do two things: Decide when two constituents should be aggregated Decide how two constituents should be aggregated, and create the aggregated structure The first issue, deciding when to aggregate, is poorly understood. Aggegration decisions certainly depend on the semantic relations between the constituents, as mentioned above; they also depend on the genre (e.g., bureaucratic texts tend to be more aggregated than instruction manuals). They probably should depend on rhetorical and discourse structure. The literacy level of the reader is also probably important (poor readers need shorter sentences). But we have no integrated model which brings all these factors together into a single algorithm. With regard to the second issue, there have been some studies of different types of aggregation, and how they should be carried out. Harbusch and Kempen describe several syntactic aggregation strategies. In their terminology, John went to the shop and bought an apple is an example of forward conjunction Reduction Much less is known about conceptual aggregation. Di Eugenio et al. show how conceptual aggregation can be done in an intelligent tutoring system, and demonstrate that performing such aggregation makes the system more effective (and that conceptual aggregation make a bigger impact than syntactic aggregation). == Software == Unfortunately there is not much software available for performing aggregation. However the SimpleNLG system does include limited support for basic aggregation. For example, the following code causes SimpleNLG to print out The man is hungry and buys an apple.