The point distribution model is a model for representing the mean geometry of a shape and some statistical modes of geometric variation inferred from a training set of shapes. == Background == The point distribution model concept has been developed by Cootes, Taylor et al. and became a standard in computer vision for the statistical study of shape and for segmentation of medical images where shape priors really help interpretation of noisy and low-contrasted pixels/voxels. The latter point leads to active shape models (ASM) and active appearance models (AAM). Point distribution models rely on landmark points. A landmark is an annotating point posed by an anatomist onto a given locus for every shape instance across the training set population. For instance, the same landmark will designate the tip of the index finger in a training set of 2D hands outlines. Principal component analysis (PCA), for instance, is a relevant tool for studying correlations of movement between groups of landmarks among the training set population. Typically, it might detect that all the landmarks located along the same finger move exactly together across the training set examples showing different finger spacing for a flat-posed hands collection. == Details == First, a set of training images are manually landmarked with enough corresponding landmarks to sufficiently approximate the geometry of the original shapes. These landmarks are aligned using the generalized procrustes analysis, which minimizes the least squared error between the points. k {\displaystyle k} aligned landmarks in two dimensions are given as X = ( x 1 , y 1 , … , x k , y k ) {\displaystyle \mathbf {X} =(x_{1},y_{1},\ldots ,x_{k},y_{k})} . It's important to note that each landmark i ∈ { 1 , … k } {\displaystyle i\in \lbrace 1,\ldots k\rbrace } should represent the same anatomical location. For example, landmark #3, ( x 3 , y 3 ) {\displaystyle (x_{3},y_{3})} might represent the tip of the ring finger across all training images. Now the shape outlines are reduced to sequences of k {\displaystyle k} landmarks, so that a given training shape is defined as the vector X ∈ R 2 k {\displaystyle \mathbf {X} \in \mathbb {R} ^{2k}} . Assuming the scattering is gaussian in this space, PCA is used to compute normalized eigenvectors and eigenvalues of the covariance matrix across all training shapes. The matrix of the top d {\displaystyle d} eigenvectors is given as P ∈ R 2 k × d {\displaystyle \mathbf {P} \in \mathbb {R} ^{2k\times d}} , and each eigenvector describes a principal mode of variation along the set. Finally, a linear combination of the eigenvectors is used to define a new shape X ′ {\displaystyle \mathbf {X} '} , mathematically defined as: X ′ = X ¯ + P b {\displaystyle \mathbf {X} '={\overline {\mathbf {X} }}+\mathbf {P} \mathbf {b} } where X ¯ {\displaystyle {\overline {\mathbf {X} }}} is defined as the mean shape across all training images, and b {\displaystyle \mathbf {b} } is a vector of scaling values for each principal component. Therefore, by modifying the variable b {\displaystyle \mathbf {b} } an infinite number of shapes can be defined. To ensure that the new shapes are all within the variation seen in the training set, it is common to only allow each element of b {\displaystyle \mathbf {b} } to be within ± {\displaystyle \pm } 3 standard deviations, where the standard deviation of a given principal component is defined as the square root of its corresponding eigenvalue. PDM's can be extended to any arbitrary number of dimensions, but are typically used in 2D image and 3D volume applications (where each landmark point is R 2 {\displaystyle \mathbb {R} ^{2}} or R 3 {\displaystyle \mathbb {R} ^{3}} ). == Discussion == An eigenvector, interpreted in euclidean space, can be seen as a sequence of k {\displaystyle k} euclidean vectors associated to corresponding landmark and designating a compound move for the whole shape. Global nonlinear variation is usually well handled provided nonlinear variation is kept to a reasonable level. Typically, a twisting nematode worm is used as an example in the teaching of kernel PCA-based methods. Due to the PCA properties: eigenvectors are mutually orthogonal, form a basis of the training set cloud in the shape space, and cross at the 0 in this space, which represents the mean shape. Also, PCA is a traditional way of fitting a closed ellipsoid to a Gaussian cloud of points (whatever their dimension): this suggests the concept of bounded variation. The idea behind PDMs is that eigenvectors can be linearly combined to create an infinity of new shape instances that will 'look like' the one in the training set. The coefficients are bounded alike the values of the corresponding eigenvalues, so as to ensure the generated 2n/3n-dimensional dot will remain into the hyper-ellipsoidal allowed domain—allowable shape domain (ASD).
Foreign key
A foreign key is a set of attributes in a table that refers to the primary key of another table, linking these two tables. In the context of relational databases, a foreign key is subject to an inclusion dependency constraint that the tuples consisting of the foreign key attributes in one relation, R, must also exist in some other (not necessarily distinct) relation, S; furthermore that those attributes must also be a candidate key in S. In other words, a foreign key is a set of attributes that references a candidate key. For example, a table called TEAM may have an attribute, MEMBER_NAME, which is a foreign key referencing a candidate key, PERSON_NAME, in the PERSON table. Since MEMBER_NAME is a foreign key, any value existing as the name of a member in TEAM must also exist as a person's name in the PERSON table; in other words, every member of a TEAM is also a PERSON. == Summary == The table containing the foreign key is called the child table, and the table containing the candidate key is called the referenced or parent table. In database relational modeling and implementation, a candidate key is a set of zero or more attributes, the values of which are guaranteed to be unique for each tuple (row) in a relation. The value or combination of values of candidate key attributes for any tuple cannot be duplicated for any other tuple in that relation. Since the purpose of the foreign key is to identify a particular row of referenced table, it is generally required that the foreign key is equal to the candidate key in some row of the primary table, or else have no value (the NULL value.). This rule is called a referential integrity constraint between the two tables. Because violations of these constraints can be the source of many database problems, most database management systems provide mechanisms to ensure that every non-null foreign key corresponds to a row of the referenced table. For example, consider a database with two tables: a CUSTOMER table that includes all customer data and an ORDER table that includes all customer orders. Suppose the business requires that each order must refer to a single customer. To reflect this in the database, a foreign key column is added to the ORDER table (e.g., CUSTOMERID), which references the primary key of CUSTOMER (e.g. ID). Because the primary key of a table must be unique, and because CUSTOMERID only contains values from that primary key field, we may assume that, when it has a value, CUSTOMERID will identify the particular customer which placed the order. However, this can no longer be assumed if the ORDER table is not kept up to date when rows of the CUSTOMER table are deleted or the ID column altered, and working with these tables may become more difficult. Many real world databases work around this problem by 'inactivating' rather than physically deleting master table foreign keys, or by complex update programs that modify all references to a foreign key when a change is needed. Foreign keys play an essential role in database design. One important part of database design is making sure that relationships between real-world entities are reflected in the database by references, using foreign keys to refer from one table to another. Another important part of database design is database normalization, in which tables are broken apart and foreign keys make it possible for them to be reconstructed. Multiple rows in the referencing (or child) table may refer to the same row in the referenced (or parent) table. In this case, the relationship between the two tables is called a one to many relationship between the referencing table and the referenced table. In addition, the child and parent table may, in fact, be the same table, i.e. the foreign key refers back to the same table. Such a foreign key is known in SQL:2003 as a self-referencing or recursive foreign key. In database management systems, this is often accomplished by linking a first and second reference to the same table. A table may have multiple foreign keys, and each foreign key can have a different parent table. Each foreign key is enforced independently by the database system. Therefore, cascading relationships between tables can be established using foreign keys. A foreign key is defined as an attribute or set of attributes in a relation whose values match a primary key in another relation. The syntax to add such a constraint to an existing table is defined in SQL:2003 as shown below. Omitting the column list in the REFERENCES clause implies that the foreign key shall reference the primary key of the referenced table. Likewise, foreign keys can be defined as part of the CREATE TABLE SQL statement. If the foreign key is a single column only, the column can be marked as such using the following syntax: Foreign keys can be defined with a stored procedure statement. child_table: the name of the table or view that contains the foreign key to be defined. parent_table: the name of the table or view that has the primary key to which the foreign key applies. The primary key must already be defined. col3 and col4: the name of the columns that make up the foreign key. The foreign key must have at least one column and at most eight columns. == Referential actions == Because the database management system enforces referential constraints, it must ensure data integrity if rows in a referenced table are to be deleted (or updated). If dependent rows in referencing tables still exist, those references have to be considered. SQL:2003 specifies 5 different referential actions that shall take place in such occurrences: CASCADE RESTRICT NO ACTION SET NULL SET DEFAULT === CASCADE === Whenever rows in the parent (referenced) table are deleted (or updated), the respective rows of the child (referencing) table with a matching foreign key column will be deleted (or updated) as well. This is called a cascade delete (or update). === RESTRICT === A value cannot be updated or deleted when a row exists in a referencing or child table that references the value in the referenced table. Similarly, a row cannot be deleted as long as there is a reference to it from a referencing or child table. To understand RESTRICT (and CASCADE) better, it may be helpful to notice the following difference, which might not be immediately clear. The referential action CASCADE modifies the "behavior" of the (child) table itself where the word CASCADE is used. For example, ON DELETE CASCADE effectively says "When the referenced row is deleted from the other table (master table), then delete also from me". However, the referential action RESTRICT modifies the "behavior" of the master table, not the child table, although the word RESTRICT appears in the child table and not in the master table! So, ON DELETE RESTRICT effectively says: "When someone tries to delete the row from the other table (master table), prevent deletion from that other table (and of course, also don't delete from me, but that's not the main point here)." RESTRICT is not supported by Microsoft SQL 2012 and earlier. === NO ACTION === NO ACTION and RESTRICT are very much alike. The main difference between NO ACTION and RESTRICT is that with NO ACTION the referential integrity check is done after trying to alter the table. RESTRICT does the check before trying to execute the UPDATE or DELETE statement. Both referential actions act the same if the referential integrity check fails: the UPDATE or DELETE statement will result in an error. In other words, when an UPDATE or DELETE statement is executed on the referenced table using the referential action NO ACTION, the DBMS verifies at the end of the statement execution that none of the referential relationships are violated. This is different from RESTRICT, which assumes at the outset that the operation will violate the constraint. Using NO ACTION, the triggers or the semantics of the statement itself may yield an end state in which no foreign key relationships are violated by the time the constraint is finally checked, thus allowing the statement to complete successfully. === SET NULL, SET DEFAULT === In general, the action taken by the DBMS for SET NULL or SET DEFAULT is the same for both ON DELETE or ON UPDATE: the value of the affected referencing attributes is changed to NULL for SET NULL, and to the specified default value for SET DEFAULT. === Triggers === Referential actions are generally implemented as implied triggers (i.e. triggers with system-generated names, often hidden.) As such, they are subject to the same limitations as user-defined triggers, and their order of execution relative to other triggers may need to be considered; in some cases it may become necessary to replace the referential action with its equivalent user-defined trigger to ensure proper execution order, or to work around mutating-table limitations. Another important limitation appears with transaction isolation: your changes to a row may not be able to fully cascade because the row is ref
Brain Imaging Data Structure
The Brain Imaging Data Structure (BIDS) is a standard for organizing, annotating, and describing data collected during neuroimaging experiments. It is based on a formalized file and directory structure and metadata files (based on JSON and TSV) with controlled vocabulary. This standard has been adopted by a multitude of labs around the world as well as databases such as OpenNeuro, SchizConnect, Developing Human Connectome Project, and FCP-INDI, and is seeing uptake in an increasing number of studies. While originally specified for MRI data, BIDS has been extended to several other imaging modalities such as MEG, EEG, and intracranial EEG (see also BIDS Extension Proposals). == History == The project is a community-driven effort. BIDS, originally OBIDS (Open Brain Imaging Data Structure), was initiated during an INCF sponsored data sharing working group meeting (January 2015) at Stanford University. It was subsequently spearheaded and maintained by Chris Gorgolewski. Since October 2019, the project is headed by a Steering Group and maintained by a separate team of maintainers, the Maintainers Group, according to a governance document that was approved of by the BIDS community in a vote. BIDS has advanced under the direction and effort of contributors, the community of researchers that appreciate the value of standardizing neuroimaging data to facilitate sharing and analysis. == BIDS Extension Proposals == BIDS can be extended in a backwards compatible way and is evolving over time. This is accomplished through BIDS Extension Proposals (BEPs), which are community-driven processes following agreed-upon guidelines. A full list of finalized BEPs and BEPs in progress can be found on the BIDS website
Social Media (Age-Restricted Users) Bill
The Social Media (Age-Restricted Users) Bill is a member's bill by National Party Member of Parliament Catherine Wedd that seeks to ban children under the age of 16 years from accessing social media by forcing social media companies to implement age verification measures. It is modelled after the Australian government's Online Safety Amendment. In mid October 2025, the New Zealand Parliament confirmed plans to introduce the social media age restriction bill. == Background == In late November 2024, the Albanese government of Australia, with support from the opposition Coalition parties, passed the Online Safety Amendment creating a world-first age verification regime targeting social media platforms operating in the country. The ban targets several social media platforms including Facebook, Instagram, Kick, Reddit, Snapchat, Threads, TikTok, Twitch, X (formerly Twitter) and YouTube. These platforms were required to implement age verification systems and to remove under-age users by 10 December 2025, when the law change came into effect. == Draft provisions == The draft Social Media (Age-Restricted Users) Bill defines social media platforms as electronic platforms that enable social media interactions between two or more end-users, facilitates communication between multiple end-users and allows users to post content on the platform. The proposed bill requires social media companies to take action to prevent users under the age of 16 from creating accounts on their platforms. It also creates a framework for courts to impose fines on platforms that fail to take reasonable steps to prevent underaged users from accessing the platform. == Legislative history == === Draft legislation === On 6 May 2025, Wedd announced a private member's bill called the "Social Media (Age-Restricted Users) Bill" that would bar access to social media platforms for people under the age of 16 years. She said that she was motivated as the mother of four children to support families, parents and teachers' efforts to manage their children's online exposure and the passage of the Australian Online Safety Amendment legislation in December 2024. Since National's coalition partner ACT New Zealand had refused to support the bill, the Sixth National Government announce it as a member's bill rather than a government bill. Prime Minister Christopher Luxon has confirmed that National would seek cross-party support for the legislation. ACT MP and the Minister of Internal Affairs Brooke van Velden said that the Government would watch the implementation of the Australian social media age restriction policy. In October 2025, Wedd's bill was drawn from the parliamentary ballot. In addition, Labour Reuben Davidson drafted a similar member's bill that would hold social media providers responsible for restricting "harmful content" and imposed NZ$50,000 fines for non-compliance. In November 2025, Luxon reiterated his support for social media age restriction legislation and said the New Zealand government would introduce a bill in 2026 before the 2026 New Zealand general election. He also confirmed that Education Minister Erica Stanford was leading an investigation into what lessons could be learnt from the Australian legislation. At the request of ACT MP Parmjeet Parmar, Parliament's Education and Workforce Committee held an inquiry into a proposed social media ban in early October 2025. The committee was led by National MP Carl Bates and received 430 submissions from 400 groups and individuals. The committee also heard from 87 in-person submissions. On 10 December 2025, the committee made 12 recommendations including restricting social media access to persons under the age of 16, re-evaluating existing legislation such as the Films, Videos, and Publications Classification Act and the Harmful Digital Communications Act 2015, and regulating online platforms and Internet service providers. The ACT party released a dissenting view disagreeing with the need for a law restricting social media access to under-16 year olds. In mid-May 2026, the Government confirmed that work on the proposed bill to ban under-16 year olds from social media had been paused. The New Zealand Parliament held a debate on the proposed bill on 13 May following a select committee inquiry into the harms caused by social media platforms. While the opposition Labour Party has agreed to support the member's bill, the ACT and Green parties opposed the proposed bill on the grounds that the rules were easy to circumvent, that at-risk groups could become more isolated, and that social media also harmed other age groups. == Responses == === Academia and civil society === In late July 2025, the New Zealand Council for Civil Liberties (NZCCL) expressed concern that the proposed social media age restriction could infringe upon the New Zealand Bill of Rights Act 1990, the Privacy Act 2020 and the United Nations' Convention on the Rights of the Child. The NZCCL also questioned the practicality of age verification software, a social media age limit and whether it would fulfil its stated goal of combating online harm. In August 2025, University of Auckland criminologist and senior lecturer Claire Meehan expressed concern that the social media age restriction legislation would cut children from their friendship and support networks. She also said that children and young people were digital natives who could use VPNs to circumvent the ban. Similar sentiments were echoed by Victoria University of Wellington media and communications lecturer Alex Beattie and "Ocean Today" Instagram social media influencer "Charlie." In October 2025, New Zealand Initiative representative Dr Eric Crampton expressed concern that a social media age restriction would involve the introduction of digital IDs. He argued that a new law was unnecessary and said that parents could limit their children's exposure to social media via Google's Family Link and Apple's equivalent. Similarly, Institute of Economic Affairs public policy fellow Matthew Lesh and the British Free Speech Union expressed concerns that young people could use VPNs to circumvent a social media ban, citing the spike in VPN usage in the United Kingdom following the passage of the Online Safety Act 2023. The advocacy group B416's co-chair Anna Curzon advocated for a social media ban on underage users, stating that social media apps "are made to be addictive" and made it difficult for parents to relate with their children. In late November 2025, B416's co-founder Anna Mowbray expressed support for the Government's social media age restriction bill but expressed disappointment that Luxon had not timed his announcement with the launch of the group's campaign. Generation-Z Aotearoa co-founder Lola Fisher has called on the New Zealand Government to consult with young people on the development of the legislation. === Government agencies and departments === In early October 2025, Privacy Commissioner Michael Webster expressed concern that social media platforms requiring users to prove their age via digital IDs could raise privacy concerns. Webster suggested that age verification systems could relay on various documents including passports. He said that age estimation technologies had high error rates and that age inference technologies relied on data mining. === Political parties === In early May 2025, the National Party government expressed support for a social media age restriction legislation. By contrast, its coalition partner ACT has opposed such legislation. ACT leader David Seymour described the ban as hasty and unworkable since it did not involve parents. Meanwhile, New Zealand First leader Winston Peters expressed support for a social media age restriction but said the bill should be subject to a select committee inquiry. The opposition Labour Party leader Chris Hipkins has expressed interest in a social media age restriction legislation but emphasised the need for consensus. Meanwhile, Green Party co-leader Chlöe Swarbrick said she wanted to learn more about the bill but described it as simplistic. Fellow Greens co-leader Marama Davidson said that the proposed bill would punish children and young people for the harm caused by big tech platforms. === Tech companies === In early October 2025, representatives of TikTok and Meta Platforms cautioned against proposed social media ban on under-16 years olds. During a one-day parliamentary inquiry, Ella Woods-Joyce, TikTok's public policy lead for Australia and New Zealand, and Mia Garlick, Meta's regional director of policy, expressed concern that the social media age restriction could send children and young people to less regulated online spaces. Woods-Joyce highlighted TikTok's policy of closing down accounts belonging to users under the age of 13 years while Garlick highlighted Meta's policy of placing users under the age of 16 in private accounts by default. In early February 2026 Meta's vice president and global head of safety, Antigone Da
Chaffing and winnowing
Chaffing and winnowing is a cryptographic technique to achieve confidentiality without using encryption when sending data over an insecure channel. The name is derived from agriculture: after grain has been harvested and threshed, it remains mixed together with inedible fibrous chaff. The chaff and grain are then separated by winnowing, and the chaff is discarded. The cryptographic technique was conceived by Ron Rivest and published in an on-line article on 18 March 1998. Although it bears similarities to both traditional encryption and steganography, it cannot be classified under either category. This technique allows the sender to deny responsibility for encrypting their message. When using chaffing and winnowing, the sender transmits the message unencrypted, in clear text. Although the sender and the receiver share a secret key, they use it only for authentication. However, a third party can make their communication confidential by simultaneously sending specially crafted messages through the same channel. == How it works == The sender (Alice) wants to send a message to the receiver (Bob). In the simplest setup, Alice enumerates the symbols in her message and sends out each in a separate packet. If the symbols are complex enough, such as natural-language text, an attacker may be able to distinguish the real symbols from poorly faked chaff symbols, posing a similar problem as steganography in needing to generate highly realistic fakes; to avoid this, the symbols can be reduced to just single 0/1 bits, and realistic fakes can then be simply randomly generated 50:50 and are indistinguishable from real symbols. In general, the method requires each symbol to arrive in-order and to be authenticated by the receiver. When implemented over networks that may change the order of packets, the sender places the symbol's serial number in the packet, the symbol itself (both unencrypted), and a message authentication code (MAC). Many MACs use a secret key Alice shares with Bob, but it is sufficient that the receiver has a method to authenticate the packets. Rivest notes an interesting property of chaffing-and-winnowing is that third parties (such as an ISP) can opportunistically add it to communications without needing permission or coordination with the sender/recipient. A third-party (Charles) who transmits Alice's packets to Bob, interleaves the packets with corresponding bogus packets (called "chaff") with corresponding serial numbers, arbitrary symbols, and a random number in place of the MAC. Charles does not need to know the key to do that (real MACs are large enough that it is extremely unlikely to generate a valid one by chance, unlike in the example). Bob uses the MAC to find the authentic messages and drops the "chaff" messages. This process is called "winnowing". An eavesdropper located between Alice and Charles can easily read Alice's message. But an eavesdropper between Charles and Bob would have to tell which packets are bogus and which are real (i.e. to winnow, or "separate the wheat from the chaff"). That is infeasible if the MAC used is secure and Charles does not leak any information on packet authenticity (e.g. via timing). If a fourth party joins the example (named Darth) who wants to send counterfeit messages to impersonate Alice, it would require Alice to disclose her secret key. If Darth cannot force Alice to disclose an authentication key (the knowledge of which would enable him to forge messages from Alice), then her messages will remain confidential. Charles, on the other hand, is no target of Darth's at all, since Charles does not even possess any secret keys that could be disclosed. == Variations == The simple variant of the chaffing and winnowing technique described above adds many bits of overhead per bit of original message. To make the transmission more efficient, Alice can process her message with an all-or-nothing transform and then send it out in much larger chunks. The chaff packets will have to be modified accordingly. Because the original message can be reconstructed only by knowing all of its chunks, Charles needs to send only enough chaff packets to make finding the correct combination of packets computationally infeasible. Chaffing and winnowing lends itself especially well to use in packet-switched network environments such as the Internet, where each message (whose payload is typically small) is sent in a separate network packet. In another variant of the technique, Charles carefully interleaves packets coming from multiple senders. That eliminates the need for Charles to generate and inject bogus packets in the communication. However, the text of Alice's message cannot be well protected from other parties who are communicating via Charles at the same time. This variant also helps protect against information leakage and traffic analysis. == Implications for law enforcement == Ron Rivest suggests that laws related to cryptography, including export controls, would not apply to chaffing and winnowing because it does not employ any encryption at all. The power to authenticate is in many cases the power to control, and handing all authentication power to the government is beyond all reason The author of the paper proposes that the security implications of handing everyone's authentication keys to the government for law-enforcement purposes would be far too risky, since possession of the key would enable someone to masquerade and communicate as another entity, such as an airline controller. Furthermore, Ron Rivest contemplates the possibility of rogue law enforcement officials framing up innocent parties by introducing the chaff into their communications, concluding that drafting a law restricting chaffing and winnowing would be far too difficult. == Trivia == The term winnowing was suggested by Ronald Rivest's father. Before the publication of Rivest's paper in 1998 other people brought to his attention a 1965 novel, Rex Stout's The Doorbell Rang, which describes the same concept and was thus included in the paper's references.
Self-supervised learning
Self-supervised learning (SSL) is a paradigm in machine learning where a model is trained on a task using the data itself to generate supervisory signals, rather than relying on externally-provided labels. In the context of neural networks, self-supervised learning aims to leverage inherent structures or relationships within the input data to create meaningful training signals. SSL tasks are designed so that solving them requires capturing essential features or relationships in the data. The input data is typically augmented or transformed in a way that creates pairs of related samples, where one sample serves as the input, and the other is used to formulate the supervisory signal. This augmentation can involve introducing noise, cropping, rotation, or other transformations. Self-supervised learning more closely imitates the way humans learn to classify objects. During SSL, the model learns in two steps. First, the task is solved based on an auxiliary or pretext classification task using pseudo-labels, which help to initialize the model parameters. Next, the actual task is performed with supervised or unsupervised learning. Self-supervised learning has produced promising results in recent years, and has found practical application in fields such as audio processing, and is being used by Facebook and others for speech recognition. == Pseudo-labels == Pseudo-labels are automatically generated labels that a model assigns to unlabeled data based on its own predictions. They are widely used in self-supervised and semi-supervised learning, where ground-truth annotations are limited or unavailable. By treating predicted labels as surrogate ground truth, learning algorithms can make use of large quantities of unlabeled data in the training process. Pseudo-labeling also plays an important role in systems that must adapt to concept drift, where the statistical properties of the data change over time. In these scenarios, the model may detect that an incoming instance deviates from previously learned behavior. The system then generates a classification result for that instance, and this predicted class is used as a pseudo-label for updating or retraining model components that are becoming outdated. This approach enables continuous adaptation in dynamic environments without requiring manual annotation. In many adaptive learning pipelines, pseudo-labels are chosen when the classifier produces sufficiently confident predictions, reducing the risk of propagating errors. These pseudo-labeled instances are then incorporated into training to refresh or evolve the model's understanding of emerging data patterns, particularly when existing components show signs of “aging” due to drift or distributional shifts. This strategy reduces reliance on manual labeling while helping maintain long-term model performance. == Types == === Autoassociative self-supervised learning === Autoassociative self-supervised learning is a specific category of self-supervised learning where a neural network is trained to reproduce or reconstruct its own input data. In other words, the model is tasked with learning a representation of the data that captures its essential features or structure, allowing it to regenerate the original input. The term "autoassociative" comes from the fact that the model is essentially associating the input data with itself. This is often achieved using autoencoders, which are a type of neural network architecture used for representation learning. Autoencoders consist of an encoder network that maps the input data to a lower-dimensional representation (latent space), and a decoder network that reconstructs the input from this representation. The training process involves presenting the model with input data and requiring it to reconstruct the same data as closely as possible. The loss function used during training typically penalizes the difference between the original input and the reconstructed output (e.g. mean squared error). By minimizing this reconstruction error, the autoencoder learns a meaningful representation of the data in its latent space. === Contrastive self-supervised learning === For a binary classification task, training data can be divided into positive examples and negative examples. Positive examples are those that match the target. For example, if training a classifier to identify birds, the positive training data would include images that contain birds. Negative examples would be images that do not. Contrastive self-supervised learning uses both positive and negative examples. The loss function in contrastive learning is used to minimize the distance between positive sample pairs, while maximizing the distance between negative sample pairs. An early example uses a pair of 1-dimensional convolutional neural networks to process a pair of images and maximize their agreement. Contrastive Language-Image Pre-training (CLIP) allows joint pretraining of a text encoder and an image encoder, such that a matching image-text pair have image encoding vector and text encoding vector that span a small angle (having a large cosine similarity). InfoNCE (Noise-Contrastive Estimation) is a method to optimize two models jointly, based on Noise Contrastive Estimation (NCE). Given a set X = { x 1 , … x N } {\displaystyle X=\left\{x_{1},\ldots x_{N}\right\}} of N {\displaystyle N} random samples containing one positive sample from p ( x t + k ∣ c t ) {\displaystyle p\left(x_{t+k}\mid c_{t}\right)} and N − 1 {\displaystyle N-1} negative samples from the 'proposal' distribution p ( x t + k ) {\displaystyle p\left(x_{t+k}\right)} , it minimizes the following loss function: L N = − E X [ log f k ( x t + k , c t ) ∑ x j ∈ X f k ( x j , c t ) ] {\displaystyle {\mathcal {L}}_{\mathrm {N} }=-\mathbb {E} _{X}\left[\log {\frac {f_{k}\left(x_{t+k},c_{t}\right)}{\sum _{x_{j}\in X}f_{k}\left(x_{j},c_{t}\right)}}\right]} === Non-contrastive self-supervised learning === Non-contrastive self-supervised learning (NCSSL) uses only positive examples. Counterintuitively, NCSSL converges on a useful local minimum rather than reaching a trivial solution, with zero loss. For the example of binary classification, it would trivially learn to classify each example as positive. Effective NCSSL requires an extra predictor on the online side that does not back-propagate on the target side. === Joint-Embedding and Predictive Architectures === A major class of self-supervised learning moves beyond contrastive pairs, instead maximizing the agreement between views while preventing collapse through statistical constraints. Rooted in Deep Canonical Correlation Analysis (Deep CCA), this approach includes Joint-Embedding Architectures (JEA) like Barlow Twins and VICReg, which enforce covariance constraints to learn invariant representations without negative sampling. Deep Latent Variable Path Modelling (DLVPM) generalizes this to multimodal systems, using path models to enforce correlation and orthogonality across diverse data types. In 2022 Yann LeCun introduced Joint-Embedding Predictive Architectures (JEPA) as a step towards decision making, reasoning, and autonomous human intelligence in machines, including self-improvement through autonomous learning. Founded in representation learning, LeCun included the concept of a “world model” in JEPA which aims to enable machines to replicate human intellect by providing machines with a concept for the world in which they exist. Unlike autoencoders, JEPAs operate entirely in latent space, avoiding pixel-level noise to focus on semantic structure. Rather than just learning invariance, JEPAs learn by predicting masked latent representations from visible context. JEPA has been applied to domains such as image analysis, audio processing, and motion in images and video. == Comparison with other forms of machine learning == SSL belongs to supervised learning methods insofar as the goal is to generate a classified output from the input. At the same time, however, it does not require the explicit use of labeled input-output pairs. Instead, correlations, metadata embedded in the data, or domain knowledge present in the input are implicitly and autonomously extracted from the data. These supervisory signals, extracted from the data, can then be used for training. SSL is similar to unsupervised learning in that it does not require labels in the sample data. Unlike unsupervised learning, however, learning is not done using inherent data structures. Semi-supervised learning combines supervised and unsupervised learning, requiring only a small portion of the learning data be labeled. In transfer learning, a model designed for one task is reused on a different task. Training an autoencoder intrinsically constitutes a self-supervised process, because the output pattern needs to become an optimal reconstruction of the input pattern itself. However, in current jargon, the term 'self-supervised' often refers to tasks based on a pretext-task training setup
Consumer relationship system
Consumer relationship systems (CRS) are specialized customer relationship management (CRM) software applications that are used to handle a company's dealings with its customers. Current consumer relationship systems integrate the software with telephone and call recording systems as well as with corporate systems for input and reporting. Customers can provide input from the company's website directly into the CRS. These systems are popular because they can deliver the 'voice of the consumer' that contributes to product quality improvement and that ultimately increases corporate profits. Consumer relationship systems that provide automated support as well as advanced systems may have artificial intelligence (AI) interfaces that can extract and analyse data collected, or handle basic questions and complaints. == History == The first CRS was developed in the 1980s. In 1981 Michael Wilke and Robert Thornton founded Wilke/Thornton, Inc in Columbus, Ohio, to develop new CRS software.