Why We Post is a research project funded by the European Research Council and launched in 2012 by Daniel Miller with the objective of examining the global impact of new social media. The study is based on ethnographic data collected through the course of 15 months in China, India, Turkey, Italy, United Kingdom, Trinidad, Chile and Brazil. The results of this project were released on 29 February 2016. This included the first three of eleven Open Access books (available via UCL Press), a five-week e-course (MOOC) on FutureLearn in English, also available in Chinese, Portuguese, Hindi, Tamil, Italian, Turkish, and Spanish on UCLeXtend. In addition a website containing key discoveries, stories and over 100 films is available in the same 8 languages.
Machine unlearning
Machine unlearning is a branch of machine learning focused on removing specific undesired element, such as private data, wrong or manipulated training data, outdated information, copyrighted material, harmful content, dangerous abilities, or misinformation, without needing to rebuild models from the ground up. Large language models, like the ones powering ChatGPT, may be asked not just to remove specific elements but also to unlearn a "concept," "fact," or "knowledge," which aren't easily linked to specific examples. New terms such as "model editing," "concept editing," and "knowledge unlearning" have emerged to describe this process. == History == Early research efforts were largely motivated by Article 17 of the GDPR, the European Union's privacy regulation commonly known as the "right to be forgotten" (RTBF), introduced in 2014. The GDPR did not anticipate that the development of large language models would make data erasure a complex task. This issue has since led to research on "machine unlearning," with a growing focus on removing copyrighted material, harmful content, dangerous capabilities, and misinformation. Just as early experiences in humans shape later ones, some concepts are more fundamental and harder to unlearn. A piece of knowledge may be so deeply embedded in the model's knowledge graph that unlearning it could cause internal contradictions, requiring adjustments to other parts of the graph to resolve them. Researchers have now also started studying unlearning in the context of removing incorrect or adversarially manipulated training data such as systematically biased labels or poisoning attacks. == Motivations == At present, machine unlearning is motivated by a growing range of concerns that extend well beyond the field's original focus on data privacy. A widely used taxonomy in the literature distinguishes two high-level categories of motivation. Access revocation covers cases where a data subject or rights holder requests the removal of data they own or control. This is most commonly associated with RTBF established by the European Union's General Data Protection Regulation (GDPR) and analogous legislation such as the California Consumer Privacy Act (CCPA). These regulations grant individuals the legal right to request erasure of their personal data from any system that has processed it, including models that were trained on it. Access revocation also encompasses the removal of copyrighted or pay-walled content that was incorporated into training corpora without the necessary licenses, a concern that has become prominent with the widespread use of largely web-scraped pre-training datasets. Model correction covers cases where the model exhibits undesirable behavior arising from the training data, regardless of any individual's request. This includes: Removal of toxic, biased, or unsafe outputs introduced by harmful content in the training set Correction of stale or factually incorrect associations, such as outdated knowledge encoded in a deployed model Removal of dangerous capabilities, such as detailed knowledge of the synthesis of chemical or biological agents Correction of the influence of data poisoning or adversarial attacks that have corrupted model behavior This second category has been formalized as corrective machine unlearning, which frames unlearning as a post-training mechanism for repairing the effects of bad or harmful training data. It is closely related to the AI safety literature, where data filtering alone has been found insufficient to prevent hazardous knowledge from being encoded in model weights, motivating unlearning as a complementary risk mitigation strategy. A further distinction has been drawn in the literature between removal {eliminating the influence of specific training data on model parameters) and suppression (preventing the model from generating specific outputs regardless of how that knowledge is encoded). These two goals are not equivalent: removing training data does not guarantee meaningful output suppression, and suppressing outputs does not constitute removal of the underlying training data's influence. == SISA Training == SISA is a training strategy consisting of four mechanisms designed to make machine unlearning more efficient by structuring how models are trained and updated. Its goal is to allow a system to remove the influence of specific data points without retraining an entire model from scratch. By reorganizing training data and workflows, SISA reduces the computational burden of unlearning requests. Sharding divides the training dataset into multiple disjoint subsets, or shards. Each shard is used to train a separate model instance. This ensures that a single data point affects only one shard, so unlearning it requires updating only the corresponding shard rather than the full model. Isolation refers to training each shard independently, with nothing shared across shards during the training process. This separation prevents cross-contamination between shards, ensuring that forgetting data in one shard does not require adjustments to any others. Slicing breaks the data within each shard into sequential slices and stores model states after each slice is trained on. When an unlearning request targets a piece of data, the system can roll back to the checkpoint before the point was seen and retrain only from that slice forward. This reduces retraining time even within a shard. Aggregation occurs at inference, when the model is queried. It combines the outputs of each shard to determine the output of the overall model. This is often through majority voting or averaging. This allows SISA-trained systems to behave like a single model despite being composed of multiple shard-level models. Together, these mechanisms enable machine learning systems to forget specific data points with far lower computational cost than full retraining. The trade-off is that sharding and slicing can lead to reduced model accuracy, worse generalization, and increased storage requirements for the intermediate checkpoints. This can be tolerable based on the needs of the individual or organization to comply with "right to be forgotten" or efficiently recover from backdoor attacks. == Algorithms == Machine unlearning algorithms are broadly categorized into exact and approximate methods, reflecting a fundamental trade-off between formal guarantees and computational tractability. === Exact Unlearning === Exact unlearning methods produce a model that is statistically indistinguishable from one retrained from scratch on the dataset with the forget data removed. The canonical framework for exact unlearning is SISA Training (Sharded, Isolated, Sliced, and Aggregated), introduced by Bourtoule et al. (2021). SISA partitions the training dataset into disjoint shards and trains a separate sub-model on each. At inference time, predictions are aggregated across sub-models. When an unlearning request is received, only the sub-model corresponding to the shard containing the target data requires retraining, reducing computational overhead proportionally to the number of shards. Exact methods provide the strongest guarantees but become prohibitively expensive for large pre-trained neural networks and are generally limited to settings where training can be structured in advance. === Approximate Unlearning === Approximate unlearning methods seek to produce a model whose behavior is sufficiently close to an exactly unlearned model without the cost of full retraining. These methods dominate practical applications. Common approaches include: Gradient Ascent: The model is fine-tuned by maximizing the loss on the forget set, directly degrading its performance on targeted data. This is the most direct approach but risks destabilizing performance on retained data. Random Labelling: The model is fine-tuned on the forget set using randomly shuffled labels, confusing its associations with the targeted data while producing a less aggressive weight shift than pure gradient ascent. Gradient Difference: Combines gradient ascent on the forget set with simultaneous gradient descent on the retain set, using the retain objective as a regularizer to preserve general model utility. KL Divergence Regularization: Minimizes the KL divergence between the outputs of the unlearned model and the original model on the retain set, anchoring behavior on data the model should remember. Weight Pruning and Fine-tuning: Parameters with the smallest L1-norm are pruned — targeting weights most weakly associated with general knowledge and potentially most associated with the forget set — followed by fine-tuning on the retain set to restore utility. Layer Reset and Fine-tuning: The first or last k layers are re-initialized to random weights and the model is subsequently fine-tuned on the retain set. This is a coarse but computationally simple approach. Selective Synaptic Dampening: Uses influence functions to estimate the effect of individual trainin
Data transformation (computing)
In computing, data transformation is the process of converting data from one format or structure into another format or structure. It is a fundamental aspect of most data integration and data management tasks such as data wrangling, data warehousing, data integration and application integration. Data transformation can be simple or complex based on the required changes to the data between the source (initial) data and the target (final) data. Data transformation is typically performed via a mixture of manual and automated steps. Tools and technologies used for data transformation can vary widely based on the format, structure, complexity, and volume of the data being transformed. A master data recast is another form of data transformation where the entire database of data values is transformed or recast without extracting the data from the database. All data in a well-designed database is directly or indirectly related to a limited set of master database tables by a network of foreign key constraints. Each foreign key constraint is dependent upon a unique database index from the parent database table. Therefore, when the proper master database table is recast with a different unique index, the directly and indirectly related data are also recast or restated. The directly and indirectly related data may also still be viewed in the original form since the original unique index still exists with the master data. Also, the database recast must be done in such a way as to not impact the applications architecture software. When the data mapping is indirect via a mediating data model, the process is also called data mediation. == Data transformation process == Data transformation can be divided into the following steps, each applicable as needed based on the complexity of the transformation required. Data discovery Data mapping Code generation Code execution Data review These steps are often the focus of developers or technical data analysts who may use multiple specialized tools to perform their tasks. The steps can be described as follows: Data discovery is the first step in the data transformation process. Typically the data is profiled using profiling tools or sometimes using manually written profiling scripts to better understand the structure and characteristics of the data and decide how it needs to be transformed. Data mapping is the process of defining how individual fields are mapped, modified, joined, filtered, aggregated etc. to produce the final desired output. Developers or technical data analysts traditionally perform data mapping since they work in the specific technologies to define the transformation rules (e.g. visual ETL tools, transformation languages). Code generation is the process of generating executable code (e.g. SQL, Python, R, or other executable instructions) that will transform the data based on the desired and defined data mapping rules. Typically, the data transformation technologies generate this code based on the definitions or metadata defined by the developers. Code execution is the step whereby the generated code is executed against the data to create the desired output. The executed code may be tightly integrated into the transformation tool, or it may require separate steps by the developer to manually execute the generated code. Data review is the final step in the process, which focuses on ensuring the output data meets the transformation requirements. It is typically the business user or final end-user of the data that performs this step. Any anomalies or errors in the data that are found and communicated back to the developer or data analyst as new requirements to be implemented in the transformation process. == Types of data transformation == === Batch data transformation === Traditionally, data transformation has been a bulk or batch process, whereby developers write code or implement transformation rules in a data integration tool, and then execute that code or those rules on large volumes of data. This process can follow the linear set of steps as described in the data transformation process above. Batch data transformation is the cornerstone of virtually all data integration technologies such as data warehousing, data migration and application integration. When data must be transformed and delivered with low latency, the term "microbatch" is often used. This refers to small batches of data (e.g. a small number of rows or a small set of data objects) that can be processed very quickly and delivered to the target system when needed. === Benefits of batch data transformation === Traditional data transformation processes have served companies well for decades. The various tools and technologies (data profiling, data visualization, data cleansing, data integration etc.) have matured and most (if not all) enterprises transform enormous volumes of data that feed internal and external applications, data warehouses and other data stores. === Limitations of traditional data transformation === This traditional process also has limitations that hamper its overall efficiency and effectiveness. The people who need to use the data (e.g. business users) do not play a direct role in the data transformation process. Typically, users hand over the data transformation task to developers who have the necessary coding or technical skills to define the transformations and execute them on the data. This process leaves the bulk of the work of defining the required transformations to the developer, which often in turn do not have the same domain knowledge as the business user. The developer interprets the business user requirements and implements the related code/logic. This has the potential of introducing errors into the process (through misinterpreted requirements), and also increases the time to arrive at a solution. This problem has given rise to the need for agility and self-service in data integration (i.e. empowering the user of the data and enabling them to transform the data themselves interactively). There are companies that provide self-service data transformation tools. They are aiming to efficiently analyze, map and transform large volumes of data without the technical knowledge and process complexity that currently exists. While these companies use traditional batch transformation, their tools enable more interactivity for users through visual platforms and easily repeated scripts. Still, there might be some compatibility issues (e.g. new data sources like IoT may not work correctly with older tools) and compliance limitations due to the difference in data governance, preparation and audit practices. === Interactive data transformation === Interactive data transformation (IDT) is an emerging capability that allows business analysts and business users the ability to directly interact with large datasets through a visual interface, understand the characteristics of the data (via automated data profiling or visualization), and change or correct the data through simple interactions such as clicking or selecting certain elements of the data. Although interactive data transformation follows the same data integration process steps as batch data integration, the key difference is that the steps are not necessarily followed in a linear fashion and typically don't require significant technical skills for completion. There are a number of companies that provide interactive data transformation tools, including Trifacta, Alteryx and Paxata. They are aiming to efficiently analyze, map and transform large volumes of data while at the same time abstracting away some of the technical complexity and processes which take place under the hood. Interactive data transformation solutions provide an integrated visual interface that combines the previously disparate steps of data analysis, data mapping and code generation/execution and data inspection. That is, if changes are made at one step (like for example renaming), the software automatically updates the preceding or following steps accordingly. Interfaces for interactive data transformation incorporate visualizations to show the user patterns and anomalies in the data so they can identify erroneous or outlying values. Once they've finished transforming the data, the system can generate executable code/logic, which can be executed or applied to subsequent similar data sets. By removing the developer from the process, interactive data transformation systems shorten the time needed to prepare and transform the data, eliminate costly errors in the interpretation of user requirements and empower business users and analysts to control their data and interact with it as needed. == Transformational languages == There are numerous languages available for performing data transformation. Many transformation languages require a grammar to be provided. In many cases, the grammar is structured using something closely resembling Backus–Naur form (BNF). There are numerous languages
Instapoetry
Instapoetry is a style of poetry that emerged after the advent of social media, especially on Instagram. The term has been used to describe poems written specifically for being shared online, most commonly on Instagram, but also other platforms including Twitter, Tumblr, and TikTok. The style usually consists of short, direct lines in aesthetically pleasing fonts that are sometimes accompanied by an image or drawing, often without rhyme schemes or meter, and dealing with commonplace themes. Literary critics, poets, and writers have contended with Instapoetry's focus on brevity and plainness compared to traditional poetry, criticizing it for reproducing rather than subverting normative ideas on social media platforms that favor popularity and accessibility over craft and depth. == History == Instapoetry developed as a result of young, predominantly women, amateur poets sharing their output to expand their readership, who began using social media as their preferred method of distribution rather than traditional publishing methods. The term "Instapoetry" is a portmanteau of the words "Instagram" and "poetry," and was created by other writers trying to define and understand the new extension of "instant poetry" shared via social media, most prominently Instagram. In its most basic form, Instapoetry usually consists of bite-sized verses that consider political and social subjects such as immigration, domestic violence, sexual assault, love, culture, feminism, gun violence, war, racism, LGBTQ rights, and other social justice topics. All of these elements are usually made to fit social media feeds that are easily accessible through applications on smartphones. == Scholarship == Despite the diversity of poetry on Instagram, the Brazilian linguist Bruna Osaki Fazano found that shared "aspects of the compositional form, theme and style" mean that it can be understood as a specific genre. Camilla Holm Soelseth argues that taking on the platform-specific tasks of a social media creator is a prerequisite for being an Instapoet. Writing in Poetics Today, JuEunhae Knox combined quantitative and qualitative analysis to show that Instapoetry is a cohesive genre, in part because "the sheer volume and rapidity of content production in turn encourages posts that are not only visually appealing but also immediately recognizable as Instapoems". Instapoetry has been seen as a practice that serves as a form of self-staging for poets and "[crafts] authenticity". Eirik Vassenden describes the work of Norwegian poet Trygve Skaug as appearing to offer a "simple, almost direct access to the inner self". Vassenden writes that poems such as Rupi Kaur's "if you are not enough for yourself / you will never be enough / for someone else" are "authentic" to such an extent that they are not literary. Kiera Obbard describes how Rupi Kaur uses humour as a rhetorical device in her poetry performances to tell personal stories of trauma and challenge social inequalities. Scholars have also studied the work of specific Instapoets, such as Rupi Kaur, R.M. Drake, Aja Monet, Yrsa Daley-Ward, Nayyirah Waheed, Atticus, Nikita Gill and Trygve Skaug. == Overview == Academics have shown appreciation for the way in which Instapoetry has stimulated interest in poetry in general. Meanwhile, it has been argued that since Instapoets avoid critical evaluations, academics, and the publishing industry, Instapoets qualify more as online celebrities than literary figures. Additionally, although Instapoetry has been characterized as anti-establishment, Alyson Miller noted traditional or even conservative views in the online posts of Instapoets in contrast with the activist views the style is associated with, and that there is a contradiction between "the extra-textual commentary surrounding Instapoetry, particularly by way of interviews and artistic statements, and the content of works which repeatedly reinscribe conservative, patriarchal, and heteronormative worldviews". Thom Young, a poet and high school English teacher, created a parody Instagram page as a way to mock Instapoets and their work, describing it as "fidget-spinner poetry. Like they're just scrolling on their devices, to read something instantly, while the libraries are empty. I think people today don't want to read anything that causes a whole lot of critical thinking." According to Johnathan Ford's piece in the Financial Times, as Instagram's algorithms have limited prospective Instapoets' reach-per-post, it has pushed them to pay to promote their material. Popular Instagram accounts will be promoted to the front of users' feeds, with the app's algorithm, in the view of critics, favoring the spread of bland, inauthentic, or clichéd content while preventing disciplined poetry from reaching new audiences. == Writers described as Instapoets == Rupi Kaur Atticus Amanda Lovelace Tyler Knott Gregson Najwa Zebian Lang Leav Nikita Gill Upile Chisala Tendai M. Shaba Donna Ashworth Trista Mateer
Data Management Association
The Data Management Association (DAMA), formerly known as the Data Administration Management Association, is a global not-for-profit organization which aims to advance concepts and practices about information management and data management. It describes itself as vendor-independent, all-volunteer organization, and has a membership consisting of technical and business professionals. Its international branch is called DAMA International (or DAMA-I), and DAMA also has various continental and national branches around the world. == History == The Data Management Association International was founded in 1980 in Los Angeles. Other early chapters were: San Francisco, Portland, Seattle, Minneapolis, New York, and Washington D.C. == Data Management Body of Knowledge == DAMA has published the Data Management Body of Knowledge (DMBOK), which contains suggestions on best practices and suggestions of a common vernacular for enterprise data management. The first edition (DAMA-DMBOK) was published on 2009 November 1, the second edition (DAMA-DMBOK2) was published on 2017 July 1., and the Revised second edition (DAMA-DMBOK2 rev.2) was published on 2019 March 19. DMBOK has been described by the authors as being an "equivalent" to the Project Management Body of Knowledge (PMBOK) and Business Analysis Body of Knowledge (BABOK). It encompasses topics such as data architecture, security, quality, modelling, governance, big data, data science, and more. DMBOK also includes the DAMA Data Wheel, an infographic which represents core data management practices. The center of the infographic is data governance, and the surrounding segments each represent a different aspect of data management: Data architecture Data modeling and design Data storage and operations Data security Data integration and interoperability Document management Content management Master data management Reference data and master data Data warehousing Metadata management Data quality Business intelligence Data science == Professional Accreditation == DAMA also provides a professional data management certification for individuals known as a Certified Data Management Professional (CDMP), which is based on the DMBOK as a study reference. There are four levels of certification based on career experience and exam results. The highest level, Fellow, requires 25 years of experience and nomination by DAMA members. It is an example of one of many competing certifications for data management professionals.
Class activation mapping
Class activation mapping methods are explainable AI (XAI) techniques used to visualize the regions of an input image that are the most relevant for a particular task, especially image classification, in convolutional neural networks (CNNs). These methods generate heatmaps by weighting the feature maps from a convolutional layer according to their relevance to the target class. In the field of artificial intelligence, generically defined as "the effort to automate intellectual tasks normally performed by humans", machine learning and deep learning were created. They both use statistical and computational methods to learn patterns from data, reducing the need for manually coded rules. Machine learning models are trained on input data and the known respective answers, learning the underlying patterns or structures present in the data. Traditional Machine learning algorithms employ manually designed feature sets, posing a direct link between machine learning designers and employed features. Deep learning is a subfield of machine learning, based on the concept of successive layers of representation, in which the data is progressively unfolded in different ways, to extract relevant and informative patterns in data analysis. Deep learning algorithms are defined as feature learning algorithms automatically learning hierarchical feature representations from raw data, extracting increasingly abstract features through multiple layers. CNNs are a specific architecture of deep learning models, designed to process spatially structured data, such as images, exploiting a series of convolution, non-linear activation and pooling operations to extract relevant features, contained in the so-called feature maps from input data. CNNs have demonstrated to be highly effective in a variety of computer vision and image processing tasks. CNNs (and deep learning models more broadly) are described as black boxes due to their complex and non-transparent internal layers of representation. The need for clearer indications on its internal working and decision-making process gave birth to XAI techniques. Among the proposed XAI techniques for computer vision tasks, Class activation mapping methods can show which pixels in an input image are important to the predicted logit for a class of interest, in a classification task. Class activation mapping methods were originally developed for class-discriminative scenarios to visualize which parts of the input image influenced the classification decision, namely to visually highlight the regions of those feature maps that contribute most strongly to the prediction of a given class. More advanced versions of these methods are not limited to image classification tasks, but have been extended also to several vision-related tasks, such as object detection, image captioning, visual question answering and image segmentation. == Background == The following methods laid the groundwork for the class activation maps approaches, forming the conceptual basis of using gradients to highlight class-discriminative regions. === Class model visualization and saliency maps for convolutional neural networks === The class model visualization and image-specific saliency maps approaches have been presented in the foundational work "Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps" by Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman and it generalizes the deconvnet method by Zeiler and Fergus. Class model visualization synthesizes an artificial input image that strongly activates the output neurons associated with a target class. Given a trained, fixed model, this method starts with a zero-initialized image, backpropagates the gradients from the class score to the image pixels, updates the image pixels increasing the specific class scores and it repeats the pixel updating process, showing an encoded (idealized version) prototype of the class of interest. Image-specific class saliency visualization method provides a visual explanation by highlighting the most relevant pixels in an image for predicting a certain class C of interest. This is done by computing the gradient of the class score with respect to the input image, I 0 , {\displaystyle I_{0},} w = ∂ S C ∂ I | I 0 {\displaystyle w=\left.{\frac {\partial S_{C}}{\partial I}}\right|_{I_{0}}} approximating the model locally (around I 0 {\displaystyle I_{0}} ) as linear, using a first-order Taylor expansion: S C ( I ) ≈ w C T I + b {\displaystyle S_{C}(I)\approx w_{C}^{T}I+b} . The magnitude of w C {\displaystyle w_{C}} , the gradient, indicates the importancy of the pixels: larger gradients suggest greater influence on the prediction. Once the gradient is known, the saliency map is defined as the maximum absolute gradient across the color channels: M i j = m a x C | ∂ S C ∂ I i j C | {\displaystyle M_{ij}=max_{C}\left|{\frac {\partial S_{C}}{\partial I_{ij}^{C}}}\right|} resulting in an saliency map (i.e. heatmap). === Guided backpropagation === The concept of guided backpropagation can be traced for the first time in the paper by Springenberg et al. "Striving For Simplicity: The All Convolutional Net" and also this method builds upon the work by Zeiler and Fergus "Visualizing and Understanding Convolutional Networks". Guided backpropagation core is to understand what a CNN is learning, by visualizing the patterns that activate more strongly individual neurons (or filters), in architectures which do not rely on max-pooling layer. When propagating gradients back through a rectified linear unit (ReLU), guided backpropagation passes the gradient if and only if the input to the ReLU was positive (forward pass) and the output gradient is positive (backward signal), tackling both inactive neurons, negative gradients and suppressing the noise. The result displays sharper, high-resolution visualizations of what each neuron is responding to. Guided backpropagation represents a simple and practical method for model interpretability, helping understand how and where neural networks detect semantic concepts across layers. Moreover, it can be applied to any network architecture, due to its working principle. == Base versions == Class activation mapping and gradient-weighted class activation mapping are the original and most widely used methods for visual explanations in convolutional neural networks. These methods serve as the foundation for many later developments in explainable AI. Notation: In this article, the symbols i and j represent integer indices that disappear inside sums or averages, while x and y are the continuous (or up-sampled integer) coordinates of the final heat-map that is plotted. === Class activation mapping (CAM) === Class activation mapping (CAM) was the first, and the original, version of CAM methods, and it gave the name to the whole category. The approach was firstly introduced by Zhou et al. in their seminal work "Learning Deep Features for Discriminative Localization". This approach achieves class-specific heatmaps by modifying image classification CNN architectures, replacing fully-connected layers with convolutional layers and a final global average pooling layer. Its main scope is to localize and highlight discriminative regions of an input image that a CNN uses to identify a particular class, without needing explicit bounding box annotations. ==== Global average pooling (GAP) ==== Global average pooling (GAP) represents the key element in the original CAM approach. It is a dimensionality reduction technique and, similarly to other pooling layers, it allows the downsampling of the feature maps, calculating representative values for a specific region of the feature map. The particularity of GAP is that it calculates a single value for an entire feature map, significantly reducing the model dimensions. ==== Mathematical description ==== The mathematical description considers as its key the combination of convolutional and GAP layers. In CAM, it is mandatory to have the GAP layer after the last convolutional layer and before the final linear classifier layer. This last element of the architecture connects the output logits (the network predictions) y C {\displaystyle y^{C}} , to the GAP values, with its respective fine-tuned weights, w k C {\displaystyle w_{k}^{C}} . Considering A k {\displaystyle A^{k}} as the last feature maps of the last convolutional layer, GAP produces one value for each feature map, by averaging all the matrix elements (i, j) of the feature map: F k = 1 m n ∑ i = 1 m ∑ j = 1 n A i j k {\displaystyle F^{k}={\frac {1}{mn}}\sum _{i=1}^{m}\sum _{j=1}^{n}A_{ij}^{k}} with A k = [ A 11 k A 12 k ⋯ A 1 n k A 21 k A 22 k ⋯ A 2 n k ⋮ ⋮ ⋱ ⋮ A m 1 k A m 2 k ⋯ A m n k ] = { A i j k ∣ 1 ≤ i ≤ m , 1 ≤ j ≤ n } {\displaystyle A^{k}={\begin{bmatrix}A_{11}^{k}&A_{12}^{k}&\cdots &A_{1n}^{k}\\A_{21}^{k}&A_{22}^{k}&\cdots &A_{2n}^{k}\\\vdots &\vdots &\ddots &\vdots \\A_{m1}^{k}&A_{m2}^{k}&\cdots &A_{mn}^{k}\end{bmatrix}}=\left\{A_{
Communications security
Communications security is the discipline of preventing unauthorized interceptors from accessing telecommunications in an intelligible form, while still delivering content to the intended recipients. In the North Atlantic Treaty Organization culture, including United States Department of Defense culture, it is often referred to by the abbreviation COMSEC. The field includes cryptographic security, transmission security, emissions security and physical security of COMSEC equipment and associated keying material. COMSEC is used to protect both classified and unclassified traffic on military communications networks, including voice, video, and data. It is used for both analog and digital applications, and both wired and wireless links. Voice over secure internet protocol VOSIP has become the de facto standard for securing voice communication, replacing the need for Secure Terminal Equipment (STE) in much of NATO, including the U.S.A. USCENTCOM moved entirely to VOSIP in 2008. == Specialties == Cryptographic security: The component of communications security that results from the provision of technically sound cryptosystems and their proper use. This includes ensuring message confidentiality and authenticity. Emission security (EMSEC): The protection resulting from all measures taken to deny unauthorized persons information of value that might be derived from communications systems and cryptographic equipment intercepts and the interception and analysis of compromising emanations from cryptographic equipment, information systems, and telecommunications systems. Transmission security (TRANSEC): The component of communications security that results from the application of measures designed to protect transmissions from interception and exploitation by means other than cryptanalysis (e.g. frequency hopping and spread spectrum). Physical security: The component of communications security that results from all physical measures necessary to safeguard classified equipment, material, and documents from access thereto or observation thereof by unauthorized persons. == Related terms == ACES – Automated Communications Engineering Software AEK – Algorithmic Encryption Key AKMS – the Army Key Management System CCI – Controlled Cryptographic Item - equipment which contains COMSEC embedded devices CT3 – Common Tier 3 DTD – Data Transfer Device ICOM – Integrated COMSEC, e.g. a radio with built in encryption KEK – Key Encryption Key KG-30 – family of COMSEC equipment KOI-18 – Tape Reader General Purpose KPK – Key production key KYK-13 – Electronic Transfer Device KYX-15 – Electronic Transfer Device LCMS – Local COMSEC Management Software OTAR – Over the Air Rekeying OWK – Over the Wire Key SKL – Simple Key Loader SOI – Signal operating instructions STE – Secure Terminal Equipment (secure phone) STU-III – (obsolete secure phone, replaced by STE) TED – Trunk Encryption Device such as the WALBURN/KG family TEK – Traffic Encryption Key TPI – Two person integrity TSEC – Telecommunications Security (sometimes referred to in error transmission security or TRANSEC) Types of COMSEC equipment: Authentication equipment Crypto equipment: Any equipment that embodies cryptographic logic or performs one or more cryptographic functions (key generation, encryption, and authentication). Crypto-ancillary equipment: Equipment designed specifically to facilitate efficient or reliable operation of crypto-equipment, without performing cryptographic functions itself. Crypto-production equipment: Equipment used to produce or load keying material == DoD Electronic Key Management System == The Electronic Key Management System (EKMS) is a United States Department of Defense (DoD) key management, COMSEC material distribution, and logistics support system. The National Security Agency (NSA) established the EKMS program to supply electronic key to COMSEC devices in securely and timely manner, and to provide COMSEC managers with an automated system capable of ordering, generation, production, distribution, storage, security accounting, and access control. The Army's platform in the four-tiered EKMS, AKMS, automates frequency management and COMSEC management operations. It eliminates paper keying material, hardcopy Signal operating instructions (SOI) and saves the time and resources required for courier distribution. It has 4 components: LCMS provides automation for the detailed accounting required for every COMSEC account, and electronic key generation and distribution capability. ACES is the frequency management portion of AKMS. ACES has been designated by the Military Communications Electronics Board as the joint standard for use by all services in development of frequency management and crypto-net planning. CT3 with DTD software is in a fielded, ruggedized hand-held device that handles, views, stores, and loads SOI, Key, and electronic protection data. DTD provides an improved net-control device to automate crypto-net control operations for communications networks employing electronically keyed COMSEC equipment. SKL is a hand-held PDA that handles, views, stores, and loads SOI, Key, and electronic protection data. == Key Management Infrastructure (KMI) Program == KMI is intended to replace the legacy Electronic Key Management System to provide a means for securely ordering, generating, producing, distributing, managing, and auditing cryptographic products (e.g., asymmetric keys, symmetric keys, manual cryptographic systems, and cryptographic applications). This system is currently being fielded by Major Commands and variants will be required for non-DoD Agencies with a COMSEC Mission.