Automatic scorer

An automatic scorer is the computerized scoring system to keep track of scoring in ten-pin bowling. It was introduced en masse in bowling alleys in the 1970s and combined with mechanical pinsetters to detect overturned pins. By eliminating the need for manual score-keeping, these systems have introduced new bowlers into the game who otherwise would not participate because they had to count the score themselves, as many do not understand the mathematical formula involved in bowler scoring. At first, people were skeptical about whether a computer could keep an accurate score. In the twenty-first century, automatic scorers are used in most bowling centers around the world. The three manufacturers of these specialty computers have been Brunswick Bowling, AMF Bowling (later QubicaAMF), and RCA. == History == Automatic equipment is considered a cornerstone of the modern bowling center. The traditional bowling center of the early 20th century was advanced in automation when the pinsetter person ("pin boy"), who set back up by hand the bowled down pins, was replaced by a machine that automatically replaced the pins in their proper play positions. This machine came out in the 1950s. A detection system was developed from the pinsetter mechanism in the 1960s that could tell which pins had been knocked down, and that information could be transferred to a digital computer. Automatic electronic scoring was first conceived by Robert Reynolds, who was described by a newspaper story at the time as "a West Coast electronics calculator expert." He worked with the technical staff of Brunswick Bowling to develop it. The goal was realized in the late 1960s when a specialized computer was designed for the purpose of automatic scorekeeping for bowling. The field test for the automatic scorer took place at Village Lanes bowling center, Chicago in 1967. The scoring machine received approval for official use by the American Bowling Congress in August of that year. They were first used in national official league gaming on October 10, 1967. In November, Brunswick announced that they were accepting orders for the new digital computer, which cost around $3,000 per bowling lane. Bowling centers that installed these new automatic scoring devices in the 1970s charged a ten cents extra per line of scoring for the convenience. == Description == Each Automatic Scorer computer unit kept score for four lanes. It had two bowler identification panels serving two lanes each. The bowler pushed it into his named position when his turn came up so the computer knew who was bowling and score accordingly. After the bowler rolled the bowling ball down the lane and knocked down pins, the pinsetter detected which pins were down and relayed this information back to the computer for scoring. The result was then printed on a scoresheet and projected overhead onto a large screen for all to see. The Automatic Scorer digital computer was mathematically accurate, however the detection system at the pinsetter mechanism sometimes reported the wrong number of pins knocked down. The computer could be corrected manually for any errors in the system; similarly, human errors, such as neglecting to move the bowler identification mechanism, could be corrected for by manual action. The scorer could take into account bowlers' handicaps and could adjust for late-arriving bowlers. The automatic scorer is directly connected to the foul detection unit. As a result, foul line violations are automatically scored. Brunswick had put ten years of research and development into the Automatic Scorer, and by 1972 there were over 500 of these computers installed in bowling centers around the world. AMF Bowling, competitor to Brunswick, entered into the automatic scorer computer field during the 1970s and their systems were installed into their brand of bowling centers. By 1974, RCA was also making these computers for automatic scoring. == Reception and further developments == The purposes of the computerized scoring were to avoid errors by human scorers and to prevent cheating. It had the side benefit of speeding up the progress of the game and introducing new bowlers to the game. Score-keeping for bowling is based on a formula that many new to bowling were not familiar with and thought difficult to learn. These casual bowlers unfamiliar with the formula thought the scores given by the computers were confusing. Some bowlers were not comfortable with automatic scorers when they were introduced in the 1970s, so kept score using the traditional method on paper score sheets. The introduction of this device increased the popularity of the sport. Automatic scorers came to be considered a normal part of modern bowling installations worldwide, with owners and managers saying that bowlers expect such equipment to be present in bowling establishments and that business increased following their introduction. Brunswick introduced a color television style automatic scorer in 1983. Bowling center owners could use these style automatic scorers for advertising, management, videos, and live television. By the 2010s, these types of electronic visual displays could show bowler avatars and social media connections to publicize the bowlers' scores. Some are capable of being extended entertainment systems of games for children and adults. Some scoring systems support variations on traditional bowling, such as different kinds of bingo games where certain pins have to be knocked down at certain times or practice regimes where certain spares have to be accomplished. By this point, QubicaAMF Worldwide, an outgrowth of AMF, was one of the leading providers of bowling scoring equipment.

Kuaishou

Kuaishou Technology is a Chinese publicly traded partly state-owned holding company based in Haidian District, Beijing, that was founded in 2011 by Hua Su (Chinese: 宿华) and Cheng Yixiao (Chinese: 程一笑). The company, listed on the Hong Kong Stock Exchange, is known for developing a mobile app for sharing users' short videos, a social network, and video special effects editor. The app is known as Kwai in many countries outside of China. It is also known as Snack Video in India, Pakistan and Indonesia. == Ownership and governance == Kuaishou's overseas team is led by the former CEO of the application 99, and staff from Google, Facebook, Netflix, and TikTok were recruited to lead the company's international expansion. The China Internet Investment Fund, a state-owned enterprise controlled by the Cyberspace Administration of China, holds a golden share ownership stake in Kuaishou. == History == Kuaishou is China's first short video platform that was developed in 2011 by engineer Hua Su and Cheng Yixiao. Prior to co-founding Kuaishou, Su Hua had worked for both Google and Baidu as a software engineer. The company is headquartered in Haidian District, Beijing. Kuaishou's predecessor "GIF Kuaishou" was founded in March 2011. GIF Kuaishou was a mobile app with which users could make and share GIF pictures. In 2013, Kuaishou became a short-video social platform. By 2013, the app had reached 100 million daily users. By 2019, it had exceeded 200 million active daily users. In March 2017, Kuaishou closed a US$350 million investment round that was led by Tencent. In January 2018, Forbes estimated the company's valuation to be US$18 billion. In April 2018, Kuaishou's app was briefly banned from Chinese app stores after China Central Television (CCTV) reported on the platform popularizing videos of teenage mothers. In 2019, the company announced a partnership with the People's Daily, an official newspaper of the Central Committee of the Chinese Communist Party, to help it experiment with the use of artificial intelligence in news. In June 2020, following the start of the 2020–2021 China–India skirmishes, the Government of India banned Kwai along with 58 other apps, citing "data and privacy issues". In January 2021, Kuaishou announced it was planning an initial public offering (IPO) to raise approximately US$5 billion. Kuaishou's stock completed its first day of trading at $300 Hong Kong dollars (HKD) (US$38.70), more than doubling its initial offer price, and causing its market value to rise to over $1 trillion HKD (US$159 billion). In February 2021, Kuaishou made a debut on the Hong Kong Stock Exchange, with its shares soaring by 194% at the opening. The company subsequently encountered major setbacks as a result of heightened regulatory restrictions on Chinese internet firms, which contributed to its share price falling by nearly 80% from its post-IPO peak. By December 2021, Kuaishou announced a major reorganization, including the layoff of 30% of its staff, primarily targeting mid-level employees earning an annual salary of $157,000 or more. This restructuring aimed to cut costs and mitigate financial losses. In October 2022, state-owned Beijing Radio and Television Station took a minority ownership stake in Kuaishou. In April 2024, a Financial Times article citing current and former Kuaishou employees stated that the company has been running an ageist redundancy programme known internally as "Limestone", culling workers in their mid-30s. In June 2024, Kuaishou and the Sichuan international communication center launched a branch center in São Paulo, Brazil. In June 2024, Kuaishou released its diffusion transformer text-to-video model, Kling, which they claimed could generate two minutes of video at 30 frames per second and in 1080p resolution. The model has been compared to that of OpenAI's Sora text-to-video model. It is accessible to the public on Kuaishou's video editing app KwaiCut via signing up for a waitlist with a Chinese phone number. In December 2025, Kuaishou came under a cyberattack which led to a temporary influx of violent and pornographic content. == Popularity == As of 2019, it had a worldwide user base of over 200 million, leading the "Most Downloaded" lists of the Google Play and Apple App Store in eight countries, such as Brazil, where it was introduced in 2019. Its main short-video platform competitor was Douyin, which is known as TikTok outside China. Compared to Douyin, Kuaishou is more popular with older users living outside China's Tier 1 cities. Its initial popularity came from videos of Chinese rural life. The app is particularly well known for its "rustic" aesthetic and is popular among rural people. Kuaishou also relied more on e-commerce revenue than on advertising revenue compared to its main competitor. == Reception == Kwai (as the app is called outside of China) was banned in India in 2020 along with other short video apps like TikTok. Kuaishou then released the clone SnackVideo, which was subsequently also banned. The app is one of the most popular social media platforms in Brazil, where Kuaishou partnered with creators to make telenovela style content, and appeals to football fans by working with football teams CR Flamengo and Santos FC and sponsoring the tournament Copa América. Kwai was notable in Brazil for spreading information (and misinformation) about the COVID-19 vaccine and political misinformation. === Manjiao Wenhua === "Manjiao wenhua" (慢脚文化) is a sarcasm term on Chinese internet on the unethical or illegal contents on Kuaishou. State broadcaster China Central Television (CCTV) reported that many contents are about child pregnancy. "Dating, pregnancy, bearing a child...these are strictly prohibited in the real time by a minor, but these contents can easily shown to audiences here." In addition, many students from primary or secondary schools make a pose of smoking. Wang Zhenhui (王贞会) from CUPSL stated that these kinds of bad values will give negative effects to the minors.

Multi expression programming

Multi Expression Programming (MEP) is an evolutionary algorithm for generating mathematical functions describing a given set of data. MEP is a Genetic Programming variant encoding multiple solutions in the same chromosome. MEP representation is not specific (multiple representations have been tested). In the simplest variant, MEP chromosomes are linear strings of instructions. This representation was inspired by Three-address code. MEP strength consists in the ability to encode multiple solutions, of a problem, in the same chromosome. In this way, one can explore larger zones of the search space. For most of the problems this advantage comes with no running-time penalty compared with genetic programming variants encoding a single solution in a chromosome. == Representation == MEP chromosomes are arrays of instructions represented in Three-address code format. Each instruction contains a variable, a constant, or a function. If the instruction is a function, then the arguments (given as instruction's addresses) are also present. === Example of MEP program === Here is a simple MEP chromosome (labels on the left side are not a part of the chromosome): 1: a 2: b 3: + 1, 2 4: c 5: d 6: + 4, 5 7: 3, 5 == Fitness computation == When the chromosome is evaluated it is unclear which instruction will provide the output of the program. In many cases, a set of programs is obtained, some of them being completely unrelated (they do not have common instructions). For the above chromosome, here is the list of possible programs obtained during decoding: E1 = a, E2 = b, E4 = c, E5 = d, E3 = a + b. E6 = c + d. E7 = (a + b) d. Each instruction is evaluated as a possible output of the program. The fitness (or error) is computed in a standard manner. For instance, in the case of symbolic regression, the fitness is the sum of differences (in absolute value) between the expected output (called target) and the actual output. == Fitness assignment process == Which expression will represent the chromosome? Which one will give the fitness of the chromosome? In MEP, the best of them (which has the lowest error) will represent the chromosome. This is different from other GP techniques: In Linear genetic programming the last instruction will give the output. In Cartesian Genetic Programming the gene providing the output is evolved like all other genes. Note that, for many problems, this evaluation has the same complexity as in the case of encoding a single solution in each chromosome. Thus, there is no penalty in running time compared to other techniques. == Software == === MEPX === MEPX is a cross-platform (Windows, macOS, and Linux Ubuntu) free software for the automatic generation of computer programs. It can be used for data analysis, particularly for solving symbolic regression, statistical classification and time-series problems. === libmep === Libmep is a free and open source library implementing Multi Expression Programming technique. It is written in C++. === hmep === hmep is a new open source library implementing Multi Expression Programming technique in Haskell programming language.

Dynamic time warping

In time series analysis, dynamic time warping (DTW) is an algorithm for measuring similarity between two temporal sequences, which may vary in speed. For instance, similarities in walking could be detected using DTW, even if one person was walking faster than the other, or if there were accelerations and decelerations during the course of an observation. DTW has been applied to temporal sequences of video, audio, and graphics data — indeed, any data that can be turned into a one-dimensional sequence can be analyzed with DTW. A well-known application has been automatic speech recognition, to cope with different speaking speeds. Other applications include speaker recognition and online signature recognition. It can also be used in partial shape matching applications. In general, DTW is a method that calculates an optimal match between two given sequences (e.g. time series) with certain restriction and rules: Every index from the first sequence must be matched with one or more indices from the other sequence, and vice versa The first index from the first sequence must be matched with the first index from the other sequence (but it does not have to be its only match) The last index from the first sequence must be matched with the last index from the other sequence (but it does not have to be its only match) The mapping of the indices from the first sequence to indices from the other sequence must be monotonically increasing, and vice versa, i.e. if j > i {\displaystyle j>i} are indices from the first sequence, then there must not be two indices l > k {\displaystyle l>k} in the other sequence, such that index i {\displaystyle i} is matched with index l {\displaystyle l} and index j {\displaystyle j} is matched with index k {\displaystyle k} , and vice versa We can plot each match between the sequences 1 : M {\displaystyle 1:M} and 1 : N {\displaystyle 1:N} as a path in a M × N {\displaystyle M\times N} matrix from ( 1 , 1 ) {\displaystyle (1,1)} to ( M , N ) {\displaystyle (M,N)} , such that each step is one of ( 0 , 1 ) , ( 1 , 0 ) , ( 1 , 1 ) {\displaystyle (0,1),(1,0),(1,1)} . In this formulation, we see that the number of possible matches is the Delannoy number. The optimal match is denoted by the match that satisfies all the restrictions and the rules and that has the minimal cost, where the cost is computed as the sum of absolute differences, for each matched pair of indices, between their values. The sequences are "warped" non-linearly in the time dimension to determine a measure of their similarity independent of certain non-linear variations in the time dimension. This sequence alignment method is often used in time series classification. Although DTW measures a distance-like quantity between two given sequences, it doesn't guarantee the triangle inequality to hold. In addition to a similarity measure between the two sequences (a so called "warping path" is produced), by warping according to this path the two signals may be aligned in time. The signal with an original set of points X(original), Y(original) is transformed to X(warped), Y(warped). This finds applications in genetic sequence and audio synchronisation. In a related technique sequences of varying speed may be averaged using this technique see the average sequence section. This is conceptually very similar to the Needleman–Wunsch algorithm. == Implementation == This example illustrates the implementation of the dynamic time warping algorithm when the two sequences s and t are strings of discrete symbols. For two symbols x and y, d ( x , y ) {\displaystyle d(x,y)} is a distance between the symbols, e.g., d ( x , y ) = | x − y | {\displaystyle d(x,y)=|x-y|} . int DTWDistance(s: array [1..n], t: array [1..m]) { DTW := array [0..n, 0..m] for i := 0 to n for j := 0 to m DTW[i, j] := infinity DTW[0, 0] := 0 for i := 1 to n for j := 1 to m cost := d(s[i], t[j]) DTW[i, j] := cost + minimum(DTW[i-1, j ], // insertion DTW[i , j-1], // deletion DTW[i-1, j-1]) // match return DTW[n, m] } where DTW[i, j] is the distance between s[1:i] and t[1:j] with the best alignment. We sometimes want to add a locality constraint. That is, we require that if s[i] is matched with t[j], then | i − j | {\displaystyle |i-j|} is no larger than w, a window parameter. We can easily modify the above algorithm to add a locality constraint (differences marked). However, the above given modification works only if | n − m | {\displaystyle |n-m|} is no larger than w, i.e. the end point is within the window length from diagonal. In order to make the algorithm work, the window parameter w must be adapted so that | n − m | ≤ w {\displaystyle |n-m|\leq w} (see the line marked with () in the code). int DTWDistance(s: array [1..n], t: array [1..m], w: int) { DTW := array [0..n, 0..m] w := max(w, abs(n-m)) // adapt window size () for i := 0 to n for j:= 0 to m DTW[i, j] := infinity DTW[0, 0] := 0 for i := 1 to n for j := max(1, i-w) to min(m, i+w) DTW[i, j] := 0 for i := 1 to n for j := max(1, i-w) to min(m, i+w) cost := d(s[i], t[j]) DTW[i, j] := cost + minimum(DTW[i-1, j ], // insertion DTW[i , j-1], // deletion DTW[i-1, j-1]) // match return DTW[n, m] } == Warping properties == The DTW algorithm produces a discrete matching between existing elements of one series to another. In other words, it does not allow time-scaling of segments within the sequence. Other methods allow continuous warping. For example, Correlation Optimized Warping (COW) divides the sequence into uniform segments that are scaled in time using linear interpolation, to produce the best matching warping. The segment scaling causes potential creation of new elements, by time-scaling segments either down or up, and thus produces a more sensitive warping than DTW's discrete matching of raw elements. == Complexity == The time complexity of the DTW algorithm is O ( N M ) {\displaystyle O(NM)} , where N {\displaystyle N} and M {\displaystyle M} are the lengths of the two input sequences. The 50 years old quadratic time bound was broken in 2016: an algorithm due to Gold and Sharir enables computing DTW in O ( N 2 / log ⁡ log ⁡ N ) {\displaystyle O({N^{2}}/\log \log N)} time and space for two input sequences of length N {\displaystyle N} . This algorithm can also be adapted to sequences of different lengths. Despite this improvement, it was shown that a strongly subquadratic running time of the form O ( N 2 − ϵ ) {\displaystyle O(N^{2-\epsilon })} for some ϵ > 0 {\displaystyle \epsilon >0} cannot exist unless the Strong exponential time hypothesis fails. While the dynamic programming algorithm for DTW requires O ( N M ) {\displaystyle O(NM)} space in a naive implementation, the space consumption can be reduced to O ( min ( N , M ) ) {\displaystyle O(\min(N,M))} using Hirschberg's algorithm. == Fast computation == Fast techniques for computing DTW include PrunedDTW, SparseDTW, FastDTW, and the MultiscaleDTW. A common task, retrieval of similar time series, can be accelerated by using lower bounds such as LB_Keogh, LB_Improved, or LB_Petitjean. However, the Early Abandon and Pruned DTW algorithm reduces the degree of acceleration that lower bounding provides and sometimes renders it ineffective. In a survey, Wang et al. reported slightly better results with the LB_Improved lower bound than the LB_Keogh bound, and found that other techniques were inefficient. Subsequent to this survey, the LB_Enhanced bound was developed that is always tighter than LB_Keogh while also being more efficient to compute. LB_Petitjean is the tightest known lower bound that can be computed in linear time. == Average sequence == Averaging for dynamic time warping is the problem of finding an average sequence for a set of sequences. NLAAF is an exact method to average two sequences using DTW. For more than two sequences, the problem is related to that of multiple alignment and requires heuristics. DBA is currently a reference method to average a set of sequences consistently with DTW. COMASA efficiently randomizes the search for the average sequence, using DBA as a local optimization process. == Supervised learning == A nearest-neighbour classifier can achieve state-of-the-art performance when using dynamic time warping as a distance measure. == Amerced Dynamic Time Warping == Amerced Dynamic Time Warping (ADTW) is a variant of DTW designed to better control DTW's permissiveness in the alignments that it allows. The windows that classical DTW uses to constrain alignments introduce a step function. Any warping of the path is allowed within the window and none beyond it. In contrast, ADTW employs an additive penalty that is incurred each time that the path is warped. Any amount of warping is allowed, but each warping action incurs a direct penalty. ADTW significantly outperforms DTW with windowing when applied as a nearest neighbor classifier on a set of benchmark time series classification tasks. == Alternative approaches == In functional data analysis, time series are regarde

Kernel principal component analysis

In the field of multivariate statistics, kernel principal component analysis (kernel PCA) is an extension of principal component analysis (PCA) using techniques of kernel methods. Using a kernel, the originally linear operations of PCA are performed in a reproducing kernel Hilbert space. == Background: Linear PCA == Recall that conventional PCA operates on zero-centered data; that is, 1 N ∑ i = 1 N x i = 0 {\displaystyle {\frac {1}{N}}\sum _{i=1}^{N}\mathbf {x} _{i}=\mathbf {0} } , where x i {\displaystyle \mathbf {x} _{i}} is one of the N {\displaystyle N} multivariate observations. It operates by diagonalizing the covariance matrix, C = 1 N ∑ i = 1 N x i x i ⊤ {\displaystyle C={\frac {1}{N}}\sum _{i=1}^{N}\mathbf {x} _{i}\mathbf {x} _{i}^{\top }} in other words, it gives an eigendecomposition of the covariance matrix: λ v = C v {\displaystyle \lambda \mathbf {v} =C\mathbf {v} } which can be rewritten as λ x i ⊤ v = x i ⊤ C v for i = 1 , … , N {\displaystyle \lambda \mathbf {x} _{i}^{\top }\mathbf {v} =\mathbf {x} _{i}^{\top }C\mathbf {v} \quad {\textrm {for}}~i=1,\ldots ,N} . (See also: Covariance matrix as a linear operator) == Introduction of the Kernel to PCA == To understand the utility of kernel PCA, particularly for clustering, observe that, while N points cannot, in general, be linearly separated in d < N {\displaystyle d

Lossless join decomposition

In database design, a lossless join decomposition is a decomposition of a relation r {\displaystyle r} into relations r 1 , r 2 {\displaystyle r_{1},r_{2}} such that a natural join of the two smaller relations yields back the original relation. This is central in removing redundancy safely from databases while preserving the original data. Lossless join can also be called non-additive. == Definition == A relation r {\displaystyle r} on schema R {\displaystyle R} decomposes losslessly onto schemas R 1 {\displaystyle R_{1}} and R 2 {\displaystyle R_{2}} if π R 1 ( r ) ⋈ π R 2 ( r ) = r {\displaystyle \pi _{R_{1}}(r)\bowtie \pi _{R_{2}}(r)=r} , that is r {\displaystyle r} is the natural join of its projections onto the smaller schemas. A pair ( R 1 , R 2 ) {\displaystyle (R_{1},R_{2})} is a lossless-join decomposition of R {\displaystyle R} or said to have a lossless join with respect to a set of functional dependencies F {\displaystyle F} if any relation r ( R ) {\displaystyle r(R)} that satisfies F {\displaystyle F} decomposes losslessly onto R 1 {\displaystyle R_{1}} and R 2 {\displaystyle R_{2}} . Decompositions into more than two schemas can be defined in the same way. == Criteria == A decomposition R = R 1 ∪ R 2 {\displaystyle R=R_{1}\cup R_{2}} has a lossless join with respect to F {\displaystyle F} if and only if the closure of R 1 ∩ R 2 {\displaystyle R_{1}\cap R_{2}} includes R 1 ∖ R 2 {\displaystyle R_{1}\setminus R_{2}} or R 2 ∖ R 1 {\displaystyle R_{2}\setminus R_{1}} . In other words, one of the following must hold: ( R 1 ∩ R 2 ) → ( R 1 ∖ R 2 ) ∈ F + {\displaystyle (R_{1}\cap R_{2})\to (R_{1}\setminus R_{2})\in F^{+}} ( R 1 ∩ R 2 ) → ( R 2 ∖ R 1 ) ∈ F + {\displaystyle (R_{1}\cap R_{2})\to (R_{2}\setminus R_{1})\in F^{+}} === Criteria for multiple sub-schemas === Multiple sub-schemas R 1 , R 2 , . . . , R n {\displaystyle R_{1},R_{2},...,R_{n}} have a lossless join if there is some way in which we can repeatedly perform lossless joins until all the schemas have been joined into a single schema. Once we have a new sub-schema made from a lossless join, we are not allowed to use any of its isolated sub-schema to join with any of the other schemas. For example, if we can do a lossless join on a pair of schemas R i , R j {\displaystyle R_{i},R_{j}} to form a new schema R i , j {\displaystyle R_{i,j}} , we use this new schema (rather than R i {\displaystyle R_{i}} or R j {\displaystyle R_{j}} ) to form a lossless join with another schema R k {\displaystyle R_{k}} (which may already be joined (e.g., R k , l {\displaystyle R_{k,l}} )). == Example == Let R = { A , B , C , D } {\displaystyle R=\{A,B,C,D\}} be the relation schema, with attributes A, B, C and D. Let F = { A → B C } {\displaystyle F=\{A\rightarrow BC\}} be the set of functional dependencies. Decomposition into R 1 = { A , B , C } {\displaystyle R_{1}=\{A,B,C\}} and R 2 = { A , D } {\displaystyle R_{2}=\{A,D\}} is lossless under F because R 1 ∩ R 2 = A {\displaystyle R_{1}\cap R_{2}=A} and we have a functional dependency A → B C {\displaystyle A\rightarrow BC} . In other words, we have proven that ( R 1 ∩ R 2 → R 1 ∖ R 2 ) ∈ F + {\displaystyle (R_{1}\cap R_{2}\rightarrow R_{1}\setminus R_{2})\in F^{+}} .

Triplet loss

Triplet loss is a machine learning loss function widely used in one-shot learning, a setting where models are trained to generalize effectively from limited examples. It was conceived by Google researchers for their prominent FaceNet algorithm for face detection. Triplet loss is designed to support metric learning. Namely, to assist training models to learn an embedding (mapping to a feature space) where similar data points are closer together and dissimilar ones are farther apart, enabling robust discrimination across varied conditions. In the context of face detection, data points correspond to images. == Definition == The loss function is defined using triplets of training points of the form ( A , P , N ) {\displaystyle (A,P,N)} . In each triplet, A {\displaystyle A} (called an "anchor point") denotes a reference point of a particular identity, P {\displaystyle P} (called a "positive point") denotes another point of the same identity in point A {\displaystyle A} , and N {\displaystyle N} (called a "negative point") denotes a point of an identity different from the identity in point A {\displaystyle A} and P {\displaystyle P} . Let x {\displaystyle x} be some point and let f ( x ) {\displaystyle f(x)} be the embedding of x {\displaystyle x} in the finite-dimensional Euclidean space. It shall be assumed that the L2-norm of f ( x ) {\displaystyle f(x)} is unity (the L2 norm of a vector X {\displaystyle X} in a finite dimensional Euclidean space is denoted by ‖ X ‖ {\displaystyle \Vert X\Vert } .) We assemble m {\displaystyle m} triplets of points from the training dataset. The goal of training here is to ensure that, after learning, the following condition (called the "triplet constraint") is satisfied by all triplets ( A ( i ) , P ( i ) , N ( i ) ) {\displaystyle (A^{(i)},P^{(i)},N^{(i)})} in the training data set: ‖ f ( A ( i ) ) − f ( P ( i ) ) ‖ 2 2 + α < ‖ f ( A ( i ) ) − f ( N ( i ) ) ‖ 2 2 {\displaystyle \Vert f(A^{(i)})-f(P^{(i)})\Vert _{2}^{2}+\alpha <\Vert f(A^{(i)})-f(N^{(i)})\Vert _{2}^{2}} The variable α {\displaystyle \alpha } is a hyperparameter called the margin, and its value must be set manually. In the FaceNet system, its value was set as 0.2. Thus, the full form of the function to be minimized is the following: L = ∑ i = 1 m max ( ‖ f ( A ( i ) ) − f ( P ( i ) ) ‖ 2 2 − ‖ f ( A ( i ) ) − f ( N ( i ) ) ‖ 2 2 + α , 0 ) {\displaystyle L=\sum _{i=1}^{m}\max {\Big (}\Vert f(A^{(i)})-f(P^{(i)})\Vert _{2}^{2}-\Vert f(A^{(i)})-f(N^{(i)})\Vert _{2}^{2}+\alpha ,0{\Big )}} == Intuition == A baseline for understanding the effectiveness of triplet loss is the contrastive loss, which operates on pairs of samples (rather than triplets). Training with the contrastive loss pulls embeddings of similar pairs closer together, and pushes dissimilar pairs apart. Its pairwise approach is greedy, as it considers each pair in isolation. Triplet loss innovates by considering relative distances. Its goal is that the embedding of an anchor (query) point be closer to positive points than to negative points (also accounting for the margin). It does not try to further optimize the distances once this requirement is met. This is approximated by simultaneously considering two pairs (anchor-positive and anchor-negative), rather than each pair in isolation. == Triplet "mining" == One crucial implementation detail when training with triplet loss is triplet "mining", which focuses on the smart selection of triplets for optimization. This process adds an additional layer of complexity compared to contrastive loss. A naive approach to preparing training data for the triplet loss involves randomly selecting triplets from the dataset. In general, the set of valid triplets of the form ( A ( i ) , P ( i ) , N ( i ) ) {\displaystyle (A^{(i)},P^{(i)},N^{(i)})} is very large. To speed-up training convergence, it is essential to focus on challenging triplets. In the FaceNet paper, several options were explored, eventually arriving at the following. For each anchor-positive pair, the algorithm considers only semi-hard negatives. These are negatives that violate the triplet requirement (i.e, are "hard"), but lie farther from the anchor than the positive (not too hard). Restated, for each A ( i ) {\displaystyle A^{(i)}} and P ( i ) {\displaystyle P^{(i)}} , they seek N ( i ) {\displaystyle N^{(i)}} such that: The rationale for this design choice is heuristic. It may appear puzzling that the mining process neglects "very hard" negatives (i.e., closer to the anchor than the positive). Experiments conducted by the FaceNet designers found that this often leads to a convergence to degenerate local minima. Triplet mining is performed at each training step, from within the sample points contained in the training batch (this is known as online mining), after embeddings were computed for all points in the batch. While ideally the entire dataset could be used, this is impractical in general. To support a large search space for triplets, the FaceNet authors used very large batches (1800 samples). Batches are constructed by selecting a large number of same-category sample points (40), and randomly selected negatives for them. == Extensions == Triplet loss has been extended to simultaneously maintain a series of distance orders by optimizing a continuous relevance degree with a chain (i.e., ladder) of distance inequalities. This leads to the Ladder Loss, which has been demonstrated to offer performance enhancements of visual-semantic embedding in learning to rank tasks. In Natural Language Processing, triplet loss is one of the loss functions considered for BERT fine-tuning in the SBERT architecture. Other extensions involve specifying multiple negatives (multiple negatives ranking loss).