AI for Business

Explore the best AI for Business — independent reviews, comparisons, pricing and step-by-step how-to guides, curated by Aizhi.

• 

  Purged cross-validation

  Purged cross-validation is a variant of k-fold cross-validation designed to prevent look-ahead bias in time series and other structured data, developed in 2017 by Marcos López de Prado at Guggenheim Partners and Cornell University. It is primarily used in financial machine learning to ensure the independence of training and testing samples when labels depend on future events. It provides an alternative to conventional cross-validation and walk-forward backtesting methods, which often yield overly optimistic performance estimates due to information leakage and overfitting. == Motivation == Standard cross-validation assumes that observations are independently and identically distributed (IID), which often does not hold in time series or financial datasets. If the label of a test sample overlaps in time with the features or labels in the training set, the result may be data leakage and overfitting. Purged cross-validation addresses this issue by removing overlapping observations and, optionally, adding a temporal buffer ("embargo") around the test set to further reduce the risk of leakage. The figure below illustrates standard 5 Fold Cross-Validation == Purging == Purging removes from the training set any observation whose timestamp falls within the time range of formation of a label in the test set. This can be the case for train set observations before and after the test set. Their removal ensures that the algorithm cannot learn during train time information that will be used to assess the performance of the algorithm. See the figure below for an illustration of purging. == Embargoing == Embargoing addresses a more subtle form of leakage: even if an observation does not directly overlap the test set, it may still be affected by test events due to market reaction lag or downstream dependencies. To guard against this, a percentage-based embargo is imposed after each test fold. For example, with a 5% embargo and 1000 observations, the 50 observations following each test fold are excluded from training. Unlike purging, embargoing can only occur after the test set. The figure below illustrates the application of embargo: == Applications == Purged and embargoed cross-validation has been useful in: Backtesting of trading strategies Validation of classifiers on labeled event-driven returns Any machine learning task with overlapping label horizons == Example == To illustrate the effect of purging and embargoing, consider the figures below. Both diagrams show the structure of 5-fold cross-validation over a 20-day period. In each row, blue squares indicate training samples and red squares denote test samples. Each label is defined based on the value of the next two observations, hence creating an overlap. If this overlap is left untreated, test set information leaks into the train set. The second figure applies the Purged CV procedure. Notice how purging removes overlapping observations from the training set and the embargo widens the gap between test and training data. This approach ensures that the evaluation more closely resembles a true out-of-sample test and reduces the risk of backtest overfitting. == Combinatorial Purged Cross-Validation == Walk-forward backtesting analysis, another common cross-validation technique in finance, preserves temporal order but evaluates the model on a single sequence of test sets. This leads to high variance in performance estimation, as results are contingent on a specific historical path. Combinatorial Purged Cross-Validation (CPCV) addresses this limitation by systematically constructing multiple train-test splits, purging overlapping samples, and enforcing an embargo period to prevent information leakage. The result is a distribution of out-of-sample performance estimates, enabling robust statistical inference and more realistic assessment of a model's predictive power. === Methodology === CPCV divides a time-series dataset into N sequential, non-overlapping groups. These groups preserve the temporal order of observations. Then, all combinations of k groups (where k < N) are selected as test sets, with the remaining N − k groups used for training. For each combination, the model is trained and evaluated under strict controls to prevent leakage. To eliminate potential contamination between training and test sets, CPCV introduces two additional mechanisms: Purging: Any training observations whose label horizon overlaps with the test period are excluded. This ensures that future information does not influence model training. Embargoing: After the end of each test period, a fixed number of observations (typically a small percentage) are removed from the training set. This prevents leakage due to delayed market reactions or auto-correlated features. Each data point appears in multiple test sets across different combinations. Because test groups are drawn combinatorially, this process produces multiple backtest "paths," each of which simulates a plausible market scenario. From these paths, practitioners can compute a distribution of performance statistics such as the Sharpe ratio, drawdown, or classification accuracy. === Formal definition === Let N be the number of sequential groups into which the dataset is divided, and let k be the number of groups selected as the test set for each split. Then: The number of unique train-test combinations is given by the binomial coefficient: ( N k ) {\displaystyle {\binom {N}{k}}} Each observation is used in k {\displaystyle k} test sets and contributes to φ [ N , k ] {\displaystyle \varphi [N,k]} unique backtest paths: φ [ N , k ] = k N ( N k ) {\displaystyle \varphi [N,k]={\frac {k}{N}}{\binom {N}{k}}} This yields a distribution of performance metrics rather than a single point estimate, making it possible to apply Monte Carlo-based or probabilistic techniques to assess model robustness. === Illustrative example === Consider the case where N = 6 and k = 2. The number of possible test set combinations is ( 6 2 ) = 15 {\displaystyle {\binom {6}{2}}=15} . Each of the six groups appears in five test splits. Consequently, five distinct backtest paths can be constructed, each incorporating one appearance from every group. ==== Test group assignment matrix ==== This table shows the 15 test combinations. An "x" indicates that the corresponding group is included in the test set for that split. ==== Backtest path assignment ==== Each group contributes to five different backtest paths. The number in each cell indicates the path to which the group's result is assigned for that split. === Advantages === Combinatorial Purged Cross-Validation offers several key benefits over conventional methods: It produces a distribution of performance metrics, enabling more rigorous statistical inference. The method systematically eliminates lookahead bias through purging and embargoing. By simulating multiple historical scenarios, it reduces the dependence on any single market regime or realization. It supports high-confidence comparisons between competing models or strategies. CPCV is commonly used in quantitative strategy research, especially for evaluating predictive models such as classifiers, regressors, and portfolio optimizers. It has been applied to estimate realistic Sharpe ratios, assess the risk of overfitting, and support the use of statistical tools such as the Deflated Sharpe Ratio (DSR). === Limitations === The main limitation of CPCV stems from its high computational cost. However, this cost can be managed by sampling a finite number of splits from the space of all possible combinations.

  Read more →
• 

  Brill tagger

  The Brill tagger is an inductive method for part-of-speech tagging. It was described and invented by Eric Brill in his 1993 PhD thesis. It can be summarized as an "error-driven transformation-based tagger". It is: a form of supervised learning, which aims to minimize error; and, a transformation-based process, in the sense that a tag is assigned to each word and changed using a set of predefined rules. In the transformation process, if the word is known, it first assigns the most frequent tag, or if the word is unknown, it naively assigns the tag "noun" to it. High accuracy is eventually achieved by applying these rules iteratively and changing the incorrect tags. This approach ensures that valuable information such as the morphosyntactic construction of words is employed in an automatic tagging process. == Algorithm == The algorithm starts with initialization, which is the assignment of tags based on their probability for each word (for example, "dog" is more often a noun than a verb). Then "patches" are determined via rules that correct (probable) tagging errors made in the initialization phase: Initialization: Known words (in vocabulary): assigning the most frequent tag associated to a form of the word Unknown word == Rules and processing == The input text is first tokenized, or broken into words. Typically in natural language processing, contractions such as "'s", "n't", and the like are considered separate word tokens, as are punctuation marks. A dictionary and some morphological rules then provide an initial tag for each word token. For example, a simple lookup would reveal that "dog" may be a noun or a verb (the most frequent tag is simply chosen), while an unknown word will be assigned some tag(s) based on capitalization, various prefix or suffix strings, etc. (such morphological analyses, which Brill calls Lexical Rules, may vary between implementations). After all word tokens have (provisional) tags, contextual rules apply iteratively, to correct the tags by examining small amounts of context. This is where the Brill method differs from other part of speech tagging methods such as those using Hidden Markov Models. Rules are reapplied repeatedly, until a threshold is reached, or no more rules can apply. Brill rules are of the general form: tag1 → tag2 IF Condition where the Condition tests the preceding and/or following word tokens, or their tags (the notation for such rules differs between implementations). For example, in Brill's notation: IN NN WDPREVTAG DT while would change the tag of a word from IN (preposition) to NN (common noun), if the preceding word's tag is DT (determiner) and the word itself is "while". This covers cases like "all the while" or "in a while", where "while" should be tagged as a noun rather than its more common use as a conjunction (many rules are more general). Rules should only operate if the tag being changed is also known to be permissible, for the word in question or in principle (for example, most adjectives in English can also be used as nouns). Rules of this kind can be implemented by simple Finite-state machines. See Part of speech tagging for more general information including descriptions of the Penn Treebank and other sets of tags. Typical Brill taggers use a few hundred rules, which may be developed by linguistic intuition or by machine learning on a pre-tagged corpus. == Code == Brill's code pages at Johns Hopkins University are no longer on the web. An archived version of a mirror of the Brill tagger at its latest version as it was available at Plymouth Tech can be found on Archive.org. The software uses the MIT License.

  Read more →
• 

  Mean shift

  Mean shift is a non-parametric feature-space mathematical analysis technique for locating the maxima of a density function, a so-called mode-seeking algorithm. Application domains include cluster analysis in computer vision and image processing. == History == The mean shift procedure is usually credited to work by Fukunaga and Hostetler in 1975. It is, however, reminiscent of earlier work by Schnell in 1964. == Overview == Mean shift is a procedure for locating the maxima—the modes—of a density function given discrete data sampled from that function. This is an iterative method, and we start with an initial estimate x {\displaystyle x} . Let a kernel function K ( x i − x ) {\displaystyle K(x_{i}-x)} be given. This function determines the weight of nearby points for re-estimation of the mean. Typically a Gaussian kernel on the distance to the current estimate is used, K ( x i − x ) = e − c | | x i − x | | 2 {\displaystyle K(x_{i}-x)=e^{-c||x_{i}-x||^{2}}} . The weighted mean of the density in the window determined by K {\displaystyle K} is m ( x ) = ∑ x i ∈ N ( x ) K ( x i − x ) x i ∑ x i ∈ N ( x ) K ( x i − x ) {\displaystyle m(x)={\frac {\sum _{x_{i}\in N(x)}K(x_{i}-x)x_{i}}{\sum _{x_{i}\in N(x)}K(x_{i}-x)}}} where N ( x ) {\displaystyle N(x)} is the neighborhood of x {\displaystyle x} , a set of points for which K ( x i − x ) ≠ 0 {\displaystyle K(x_{i}-x)\neq 0} . The difference m ( x ) − x {\displaystyle m(x)-x} is called mean shift in Fukunaga and Hostetler. The mean-shift algorithm now sets x ← m ( x ) {\displaystyle x\leftarrow m(x)} , and repeats the estimation until m ( x ) {\displaystyle m(x)} converges. Although the mean shift algorithm has been widely used in many applications, a rigid proof for the convergence of the algorithm using a general kernel in a high dimensional space is still not known. Aliyari Ghassabeh showed the convergence of the mean shift algorithm in one dimension with a differentiable, convex, and strictly decreasing profile function. However, the one-dimensional case has limited real world applications. Also, the convergence of the algorithm in higher dimensions with a finite number of the stationary (or isolated) points has been proved. However, sufficient conditions for a general kernel function to have finite stationary (or isolated) points have not been provided. Gaussian Mean-Shift is an Expectation–maximization algorithm. == Details == Let data be a finite set S {\displaystyle S} embedded in the n {\displaystyle n} -dimensional Euclidean space, X {\displaystyle X} . Let K {\displaystyle K} be a flat kernel that is the characteristic function of the λ {\displaystyle \lambda } -ball in X {\displaystyle X} , In each iteration of the algorithm, s ← m ( s ) {\displaystyle s\leftarrow m(s)} is performed for all s ∈ S {\displaystyle s\in S} simultaneously. The first question, then, is how to estimate the density function given a sparse set of samples. One of the simplest approaches is to just smooth the data, e.g., by convolving it with a fixed kernel of width h {\displaystyle h} , where x i {\displaystyle x_{i}} are the input samples and k ( r ) {\displaystyle k(r)} is the kernel function (or Parzen window). h {\displaystyle h} is the only parameter in the algorithm and is called the bandwidth. This approach is known as kernel density estimation or the Parzen window technique. Once we have computed f ( x ) {\displaystyle f(x)} from the equation above, we can find its local maxima using gradient ascent or some other optimization technique. The problem with this "brute force" approach is that, for higher dimensions, it becomes computationally prohibitive to evaluate f ( x ) {\displaystyle f(x)} over the complete search space. Instead, mean shift uses a variant of what is known in the optimization literature as multiple restart gradient descent. Starting at some guess for a local maximum, y k {\displaystyle y_{k}} , which can be a random input data point x 1 {\displaystyle x_{1}} , mean shift computes the gradient of the density estimate f ( x ) {\displaystyle f(x)} at y k {\displaystyle y_{k}} and takes an uphill step in that direction. == Types of kernels == Kernel definition: Let X {\displaystyle X} be the n {\displaystyle n} -dimensional Euclidean space, R n {\displaystyle \mathbb {R} ^{n}} . The norm of x {\displaystyle x} is a non-negative number, ‖ x ‖ 2 = x ⊤ x ≥ 0 {\displaystyle \|x\|^{2}=x^{\top }x\geq 0} . A function K : X → R {\displaystyle K:X\rightarrow \mathbb {R} } is said to be a kernel if there exists a profile, k : [ 0 , ∞ ] → R {\displaystyle k:[0,\infty ]\rightarrow \mathbb {R} } , such that K ( x ) = k ( ‖ x ‖ 2 ) {\displaystyle K(x)=k(\|x\|^{2})} and k is non-negative. k is non-increasing: k ( a ) ≥ k ( b ) {\displaystyle k(a)\geq k(b)} if a < b {\displaystyle a Read more →

• 

  Microapp

  A microapp is a super-specialized application designed to perform one task or use case with the only objective of doing it well. They follow the single responsibility principle, which states that "a class should have one and only one reason to change." Micro applications help developers create less complex applications while reducing costs by breaking down monolithic systems into groups of independent services acting as one system. A good example of Microapps would be https://docs.citrix.com/en-us/legacy-archive/downloads/microapps.pdfthat provide single purpose action from Salesforce and over 40 applications on its workspace. == Requirements and characteristics == Microapps usually are accessible on any device, display, or operating system without installation on the viewer's device. To qualify as a microapp, the entity must: be built and deployed as an independent software module bring together various media types into a single experience have advanced security and compliance features be functionally-extensible comply with granular data demands be agnostic single use case oriented Microapps differentiate from traditional web or mobile applications by how the end-user interacts with them. Consequently, they can be embedded in websites or viewed online to bypass app stores and are typically built to provide a focused experience to the user. == Usage == Microapps are typically used for commercial purposes to reduce development costs for projects not requiring the large scope of a traditional web or mobile application. In addition, they are often used to showcase in-depth information or enrich marketing material with interactivity. Lately, micro apps are being used to boost productivity by providing quick tools to people to reuse best practices. Users have been interacting with microapps for a while with suites like Microsoft 365 and Google Workspace, where each one of their end-user services could be considered as a microapp. All these microapps share a unique identity manager to provide a unified user experience. == Benefits == Replacing monolith systems with microapps provide several advantages like: Reduce complexity for developers and users. Smaller, more cohesive, and maintainable codebases Scalable organizations with decoupled, autonomous teams Allows for hyper-specialization Independent deployment Multi-stack == Cloud-native microapps == Technologies like Kubernetes, or OpenShift, allow companies to replace their monolith and legacy systems with modular software taking advantage of microapps on reducing costs and improve reliability and security. == Microapps vs. microservices == There is a widespread misunderstanding between these two concepts, which is the key difference. Microservices is an architectural style that is systems-centric, meaning it decouples the presentation and data layer using web services APIs. On the other side, micro apps behave more as a super-architecture style (that embraces microservices among other types), and it is user-centric, meaning they decouple the whole monolith system onto modules that are designed to interact with final users. Both architectural styles rely on modularity to provide high performance, scalability, and resilience. == Considerations == Developing Micro apps requires a different approach than traditional software, and user experience is crucial. The following considerations are essential for switching to microapps. To run multiple microapps is required a single identity management system. Microservices are well suited to make microapps more powerful Apps with different levels of maturity might create a non-unified user experience. Duplication of dependencies can create security issues and inefficiencies. Suitable for well-organized teams

  Read more →
• 

  MLOps

  MLOps or ML Ops is a paradigm that aims to deploy and maintain machine learning models in production reliably and efficiently. It bridges the gap between machine learning development and production operations, ensuring that models are robust, scalable, and aligned with business goals. The word is a compound of "machine learning" and the continuous delivery practice (CI/CD) of DevOps in the software field. Machine learning models are tested and developed in isolated experimental systems. When an algorithm is ready to be launched, MLOps is practiced between data scientists, DevOps, and machine learning engineers to transition the algorithm to production systems. Similar to DevOps or DataOps approaches, MLOps seeks to increase automation and improve the quality of production models, while also focusing on business and regulatory requirements. While MLOps started as a set of best practices, it is slowly evolving into an independent approach to ML lifecycle management. MLOps applies to the entire lifecycle - from integrating with model generation (software development lifecycle, continuous integration/continuous delivery), orchestration, and deployment, to health, diagnostics, governance, and business metrics. == Definition == MLOps is a paradigm, including aspects like best practices, sets of concepts, as well as a development culture when it comes to the end-to-end conceptualization, implementation, monitoring, deployment, and scalability of machine learning products. Most of all, it is an engineering practice that leverages three contributing disciplines: machine learning, software engineering (especially DevOps), and data engineering. MLOps is aimed at productionizing machine learning systems by bridging the gap between development (Dev) and operations (Ops). Essentially, MLOps aims to facilitate the creation of machine learning products by leveraging these principles: CI/CD automation, workflow orchestration, reproducibility; versioning of data, model, and code; collaboration; continuous ML training and evaluation; ML metadata tracking and logging; continuous monitoring; and feedback loops. == History == Interest in operationalizing machine learning systems began to grow in the mid-2010s as ML projects started moving from experimentation to production use. The challenges associated with sustaining such systems were highlighted in a 2015 paper. The predicted growth in machine learning included an estimated doubling of ML pilots and implementations from 2017 to 2018, and again from 2018 to 2020. Reports show a majority (up to 88%) of corporate machine learning initiatives are struggling to move beyond test stages. However, those organizations that actually put machine learning into production saw a 3–15% profit margin increases. The MLOps market size was USD 2,191.8 Million in 2024, and is projected to be USD 16,613.4 Million in 2030. == Architecture == Machine Learning systems can be categorized in eight different categories: data collection, data processing, feature engineering, data labeling, model design, model training and optimization, endpoint deployment, and endpoint monitoring. Each step in the machine learning lifecycle is built in its own system, but requires interconnection. These are the minimum systems that enterprises need to scale machine learning within their organization. == Goals == There are a number of goals enterprises want to achieve through MLOps systems successfully implementing ML across the enterprise, including: Deployment and automation Reproducibility of models and predictions Diagnostics Governance and regulatory compliance Scalability Collaboration Business uses Monitoring and management A standard practice, such as MLOps, takes into account each of the aforementioned areas, which can help enterprises optimize workflows and avoid issues during implementation. Vendors such as Adaptive ML deliver commercial reinforcement learning operations (RLOps) and MLOps-infrastructure, targeting organizations deploying large language models in production. A common architecture of an MLOps system would include data science platforms where models are constructed and the analytical engines where computations are performed, with the MLOps tool orchestrating the movement of machine learning models, data and outcomes between the systems.

  Read more →
• 

  Ericom Connect

  Ericom Connect is a remote access/application publishing solution produced by Ericom Software that provides secure, centrally managed access to physical or hosted desktops and applications running on Microsoft Windows and Linux systems. == Product overview == Ericom Connect is desktop virtualization and application virtualization software that allows users to run applications remotely, without installing them on the local computer or device. The software is noted for its scalability, ease of deployment, and compatibility with any type of infrastructure, cloud or physical. Ericom Connect uses AccessPad (native client for desktops), AccessToGo (native client for mobile), or AccessNow, one of the first HTML5 RDP solutions to support clientless access to Windows desktops and applications from any device with an HTML5-compatible browser, including Macintosh computers, mobile devices, and Google Chromebooks. Other notable features include performance monitoring, built-in real-time analytics & BI, support for two-factor authentication (using RSA SecurID), multi-tenancy and multi-datacenter support via a single unified web interface, and a “Launch Simulation” feature that allows users to visualize and simulate actual step-by-step user processes directly from within the administration console. In addition to scalability, by distributing configurations, logs, etc., across multiple servers there is no single point of failure, as can be the case if all configuration information is stored on one server. == History == Ericom Connect was introduced in 2015. Ericom Connect is a successor to Ericom PowerTerm Web Connect. PowerTerm Web Connect used an architecture similar to what was then current with Citrix and VMWare, relying on a centralized SQL server, a connection broker, image management for different hypervisors, and a variety of clients. Ericom Connect uses a new grid architecture that provides more scalability, reliability, and flexibility than before.

  Read more →
• 

  Lose It!

  Lose It! is an American health and wellness mobile app developed by FitNow, Inc. The app generates calorie budgets for users by tracking weight, exercise, food and calorie intake, and personal goals, primarily to assist them in achieving weight loss. == History == Lose It! was developed in Boston and debuted in 2008. The app and its associated company were founded by J.J. Allaire, Charles Teague and Paul Dicristina. Prior to founding Lose It!, Teague and Allaire had founded the online research tool Onfolio, which was acquired by Microsoft in 2006. The Lose It! app was originally released as an iOS app before being released as a website in 2010 and an Android app in 2011. In 2015, Lose It! announced plans to release the app internationally. Lose It! was also available as an app for Apple Watch at its launch in 2015. The app’s “Snap It” feature, which allows users to approximate calorie counts by taking pictures of their daily meals and snacks, was released in beta in 2016. Snap It was named an Innovation Awards Honoree at the 2017 Consumer Electronics Show in Las Vegas. In 2020, Patrick Wetherille, one of the company’s earliest employees, was appointed chief executive officer. == App == Lose It! is weight loss app. The app allows users to set goals such as increasing strength, overall health/maintenance, and weight loss. It provides users recommended calorie budgets based on data such as their current weight and their desired weight. Lose It! also tracks data such as exercise/activity level and food consumption and allows users to track calories consumed by scanning barcodes for food products then retrieving calorie information for products. The app can also estimate the amount of calories in a food products. Lose It! has integration features connecting it to other apps such as Fitbit and Runkeeper. It also has social features such as joining groups and sharing progress with friends. The Premium version of the app allows users to track foods according to specific diets like keto, heart healthy or Mediterranean.

  Read more →
• 

  WaveNet

  WaveNet is a deep neural network for generating raw audio. It was created by researchers at London-based AI firm DeepMind. The technique, outlined in a paper in September 2016, is able to generate relatively realistic-sounding human-like voices by directly modelling waveforms using a neural network method trained with recordings of real speech. Tests with US English and Mandarin reportedly showed that the system outperforms Google's best existing text-to-speech (TTS) systems, although as of 2016 its text-to-speech synthesis still was less convincing than actual human speech. WaveNet's ability to generate raw waveforms means that it can model any kind of audio, including music. == History == Generating speech from text is an increasingly common task thanks to the popularity of software such as Apple's Siri, Microsoft's Cortana, Amazon Alexa and the Google Assistant. Most such systems use a variation of a technique that involves concatenated sound fragments together to form recognisable sounds and words. The most common of these is called concatenative TTS. It consists of large library of speech fragments, recorded from a single speaker that are then concatenated to produce complete words and sounds. The result sounds unnatural, with an odd cadence and tone. The reliance on a recorded library also makes it difficult to modify or change the voice. Another technique, known as parametric TTS, uses mathematical models to recreate sounds that are then assembled into words and sentences. The information required to generate the sounds is stored in the parameters of the model. The characteristics of the output speech are controlled via the inputs to the model, while the speech is typically created using a voice synthesiser known as a vocoder. This can also result in unnatural sounding audio. == Design and ongoing research == === Background === WaveNet is a type of feedforward neural network known as a deep convolutional neural network (CNN). In WaveNet, the CNN takes a raw signal as an input and synthesises an output one sample at a time. It does so by sampling from a softmax (i.e. categorical) distribution of a signal value that is encoded using μ-law companding transformation and quantized to 256 possible values. === Initial concept and results === According to the original September 2016 DeepMind research paper WaveNet: A Generative Model for Raw Audio, the network was fed real waveforms of speech in English and Mandarin. As these pass through the network, it learns a set of rules to describe how the audio waveform evolves over time. The trained network can then be used to create new speech-like waveforms at 16,000 samples per second. These waveforms include realistic breaths and lip smacks – but do not conform to any language. WaveNet is able to accurately model different voices, with the accent and tone of the input correlating with the output. For example, if it is trained with German, it produces German speech. The capability also means that if the WaveNet is fed other inputs – such as music – its output will be musical. At the time of its release, DeepMind showed that WaveNet could produce waveforms that sound like classical music. === Content (voice) swapping === According to the June 2018 paper Disentangled Sequential Autoencoder, DeepMind has successfully used WaveNet for audio and voice "content swapping": the network can swap the voice on an audio recording for another, pre-existing voice while maintaining the text and other features from the original recording. "We also experiment on audio sequence data. Our disentangled representation allows us to convert speaker identities into each other while conditioning on the content of the speech." (p. 5) "For audio, this allows us to convert a male speaker into a female speaker and vice versa [...]." (p. 1) According to the paper, a two-digit minimum amount of hours (c. 50 hours) of pre-existing speech recordings of both source and target voice are required to be fed into WaveNet for the program to learn their individual features before it is able to perform the conversion from one voice to another at a satisfying quality. The authors stress that "[a]n advantage of the model is that it separates dynamical from static features [...]." (p. 8), i. e. WaveNet is capable of distinguishing between the spoken text and modes of delivery (modulation, speed, pitch, mood, etc.) to maintain during the conversion from one voice to another on the one hand, and the basic features of both source and target voices that it is required to swap on the other. The January 2019 follow-up paper Unsupervised speech representation learning using WaveNet autoencoders details a method to successfully enhance the proper automatic recognition and discrimination between dynamical and static features for "content swapping", notably including swapping voices on existing audio recordings, in order to make it more reliable. Another follow-up paper, Sample Efficient Adaptive Text-to-Speech, dated September 2018 (latest revision January 2019), states that DeepMind has successfully reduced the minimum amount of real-life recordings required to sample an existing voice via WaveNet to "merely a few minutes of audio data" while maintaining high-quality results. Its ability to clone voices has raised ethical concerns about WaveNet's ability to mimic the voices of living and dead persons. According to a 2016 BBC article, companies working on similar voice-cloning technologies (such as Adobe Voco) intend to insert watermarking inaudible to humans to prevent counterfeiting, while maintaining that voice cloning satisfying, for instance, the needs of entertainment-industry purposes would be of a far lower complexity and use different methods than required to fool forensic evidencing methods and electronic ID devices, so that natural voices and voices cloned for entertainment-industry purposes could still be easily told apart by technological analysis. == Applications == At the time of its release, DeepMind said that WaveNet required too much computational processing power to be used in real world applications. As of October 2017, Google announced a 1,000-fold performance improvement along with better voice quality. WaveNet was then used to generate Google Assistant voices for US English and Japanese across all Google platforms. In November 2017, DeepMind researchers released a research paper detailing a proposed method of "generating high-fidelity speech samples at more than 20 times faster than real-time", called "Probability Density Distillation". At the annual I/O developer conference in May 2018, it was announced that new Google Assistant voices were available and made possible by WaveNet; WaveNet greatly reduced the number of audio recordings that were required to create a voice model by modeling the raw audio of the voice actor samples.

  Read more →
• 

  Pocketbook (application)

  Pocketbook was a Sydney-based free budget planner and personal finance app launched in 2012. The app helped users setup and manage budgets, track spending and manage bills. As of 2016 Pocketbook claimed to support over 250,000 Australians, in January 2018 that number was 435,000. After being acquired by Zip Co Ltd in 2016, it was announced in 2022 that the app was to be shut down and all user accounts deleted. == History == Pocketbook was founded by Alvin Singh and Bosco Tan in 2012. It was conceived in 2011 in a Wolli Creek apartment as a tool for Alvin and Bosco to take control of their money. In 2013, Pocketbook raised $500,000 from technology fund Tank Stream Ventures, and a group of investors including TV personality David Koch, Geoff Levy, David Shein and Peter Cooper. In September 2016 Digital retail finance and payment industry player zipMoney (now trading as Zip Co Limited) acquired Pocketbook in a $7.5m deal == Features == The app synced with the bank account of users and would organize spending into different categories. Users could also be reminded of bill payments, analyse spending and set spending limits. They can also be alerted of fraudulent transactions and deductions. The app employs security measures like end to end encryption, CloudFlare protection, fraud detection, identity protection etc. Pocketbook was available via web and mobile version. == Awards == Personal Finance Innovator of the Year by Fintech Business Awards 2017 Innovator of the Year by OPTUS MyBusiness Awards 2017 Best Finance App of 2016 by Australian Fintech Best Personal Finance App: Pocketbook won the 2016 Finder Innovation Awards, presented at a gala dinner hosted by media personality and The New Inventors presenter James O'Loghlin. Best Mobile App of the Year Winner: StartCon hosted the first annual Australasian Startup Awards. Over 200 nominations in 14 categories and an overall winner were reviewed, and winners were determined by public voting, with over 63,000 votes in total. Best New Startup 2014 by StartupSmart. Finalist in the SWIFT Innotribe startup competition in Dubai in 2013.

  Read more →
• 

  Artificial intelligence in fraud detection

  Artificial intelligence is used by many different businesses and organizations. It is widely used in the financial sector, especially by accounting firms, to help detect fraud. In 2022, PricewaterhouseCoopers reported that fraud has impacted 46% of all businesses in the world. The shift from working in person to working from home has brought increased access to data. According to an FTC (Federal Trade Commission) study from 2022, customers reported fraud of approximately $5.8 billion in 2021, an increase of 70% from the year before. The majority of these scams were imposter scams and online shopping frauds. Furthermore, artificial intelligence plays a crucial role in developing advanced algorithms and machine learning models that enhance fraud detection systems, enabling businesses to stay ahead of evolving fraudulent tactics in an increasingly digital landscape. == Tools == === Expert systems === Expert systems were first designed in the 1970s as an expansion into artificial intelligence technologies. Their design is based on the premise of decreasing potential user error in decision-making and emulating mental reasoning used by experts in a particular field. They differentiate themselves from traditional linear reasoning models by separating identified points in data and processing them individually at the same time. Though, these systems do not rely purely on machine-learned intelligence. Information regarding rules, practices, and procedures in the form of "if-then" statements are implemented into the programming of the system. Users interact with the system by feeding information into the system either through direct entry or import of external data. An inference system compares the information provided by the user with corresponding rules that are believed to specifically apply to the situation. Using this information and the corresponding rules will be used to create a solution to the user's query. Expert systems will generally not operate properly when the common procedures for a specified situation are ambiguous due to the need for well-defined rules. Implementation of expert systems in accounting procedures is feasible in areas where professional judgment is required. Situations where expert systems are applicable include investigations into transactions that involve potential fraudulent entries, instances of going concern, and the evaluation of risk in the planning stages of an audit. === Continuous auditing === Continuous auditing is a set of processes that assess various aspects of information gathered in an audit to classify areas of risk and potential weaknesses in financial Internal controls at a more frequent rate than traditional methods. Instead of analyzing recorded transactions and journal entries periodically, continuous auditing focuses on interpreting the character of these actions more frequently. The frequency of these processes being undertaken as well as highlighting areas of importance is up to the discretion of their implementer, who commonly makes such decisions based on the level of risk in the accounts being evaluated and the goals of implementing the system. Performance of these processes can occur as frequently as being nearly instantaneous with an entry being posted. The processes involved with analyzing financial data in continuous auditing can include the creation of spreadsheets to allow for interactive information gathering, calculation of financial ratios for comparison with previously created models, and detection of errors in entered figures. A primary goal of this practice is to allow for quicker and easier detection of instances of faulty controls, errors, and instances of fraud. === Machine learning and deep learning === The ability of machine learning and deep learning to swiftly and effectively sort through vast volumes of data in the forms of various documents relevant to companies and documents being audited makes them applicable to the domains of audit and fraud detection. Examples of this include recognizing key language in contracts, identifying levels of risk of fraud in transactions, and assessing journal entries for misstatement. == Applications == === 'Big 4' Accounting Firms === Deloitte created an Al-enabled document-reviewing system in 2014. The system automates the method of reviewing and extracting relevant information from different business documents. Deloitte claims that this innovation has made a difference by reducing time spent going through lawful contract documents, invoices, money-related articulations, and board minutes by up to 50%. Working with IBM's Watson, Deloitte is developing cognitive-technology-enhanced commerce arrangements for its clients. LeasePoint is fueled by IBM TRIRIGA (this product evolved into IBM Maximo Real Estate and Facilities) and uses Deloitte's industrial information to create an end-to-end leasing portfolio. Automated Cognitive Resource Assessment employs IBM's Maximo innovation to progress the proficiency of asset inspection. Ernst and Young (EY) connected Al to the investigation of lease contracts. EY (Australia) has also received Al-enabled auditing technology. Collaborating with H20.ai, PwC developed an Al-enabled framework (GL.ai) capable of analyzing reports and preparing reports. PwC claims to have made a significant investment in normal dialect processing (NLP), an Al-enabled innovation to process unstructured information efficiently. KPMG built a portfolio of Al instruments, called KPMG Ignite, to upgrade trade decisions and forms. Working with Microsoft and IBM Watson, KPMG is creating instruments to coordinate Al, data analytics, Cognitive Technologies, and RPA. == Advantages == === Efficiency === The process of auditing an entity in an attempt to detect fraudulent activity requires the repeating of investigatory processes until an error or misstatement may be identified. Under traditional methods, these processes would be carried out by a human being. Proponents of artificial intelligence in fraud detection have stated that these traditional methods are inefficient and can be more quickly accomplished with the aid of an intelligent computing system. A survey of 400 chief executive officers created by KPMG in 2016 found that approximately 58% believed that artificial intelligence would play a key role in making audits more efficient in the future. === Data interpretation === Higher levels of fraud detection entail the use of professional judgement to interpret data. Supporters of artificial intelligence being used in financial audits have claimed that increased risks from instances of higher data interpretation can be minimized through such technologies. One necessary element of an audit of financial statements that requires professional judgement is the implementation of thresholds for materiality. Materiality entails the distinction between errors and transactions in financial statements that would impact decisions made by users of those financial statements. The threshold for materiality in an audit is set by the auditor based on various factors. Artificial intelligence has been used to interpret data and suggest materiality thresholds to be implemented through the use of expert systems. === Decreased costs === Those in favor of using artificial intelligence to complete investigations of fraud have stated that such technologies decrease the amount of time required to complete tasks that are repetitive. The claim further states that such efficiencies allow for lowered resource requirements, which can then be further spent on tasks that have not been fully automated. The audit firm Ernst & Young has posited these claims by declaring that their deep learning systems have been used to reduce time spent on administrative tasks by analyzing relevant audit documents. According to the firm, this has allowed their employees to focus more on judgement and analysis. == Disadvantages == === Job Displacement === The inescapable reception of computer based intelligence and robotization advancements might prompt critical work relocation across different enterprises. As artificial intelligence frameworks become more equipped for performing undertakings customarily completed by people, there is a worry that specific work jobs could become out of date, prompting joblessness and financial imbalance. === Initial investment requirement === Along with a knowledge of coding and building systems through computer programs, we are seeing the advantages of these systems, but since they are so new, they require a large investment to start building such a system. Any firm that is planning on implementing an AI system to detect fraud must hire a team of data scientists, along with upgrading their cloud system and data storage. The system must be consistently monitored and updated to be the most efficient form of itself, otherwise the likelihood of fraud being involved in those transactions increases. If one does not initially invest in such a syst

  Read more →
• 

  Pandorabots

  Pandorabots, Inc. is an artificial intelligence company that runs a web service for building and deploying chatbots. Pandorabots implements and supports development of the Artificial Intelligence Markup Language and makes portions of its code accessible for free. The Pandorabots Platform is "one of the oldest and largest chatbot hosting services in the world", allowing creation of virtual agents to hold human-like text or voice chats with consumers. The platform is written in Allegro Common LISP. == Use Cases == Common use cases include advertising, virtual assistance, e-learning, entertainment and education. The platform has also been used by academics and universities use the platform for teaching and research.

  Read more →
• 

  Tay (chatbot)

  Tay was a chatbot that was originally released by Microsoft Corporation as a Twitter bot on March 23, 2016. It caused subsequent controversy when the bot began to post inflammatory and offensive tweets through its Twitter account, causing Microsoft to shut down the service only 16 hours after its launch. According to Microsoft, this was caused by trolls who "attacked" the service as the bot made replies based on its interactions with people on Twitter. It was replaced with Zo. == Background == The bot was created by Microsoft's Technology and Research and Bing divisions, and named "Tay" as an acronym for "thinking about you". Although Microsoft initially released few details about the bot, sources mentioned that it was similar to or based on Xiaoice, a Microsoft project in China. Ars Technica reported that, since late 2014 Xiaoice had had "more than 40 million conversations apparently without major incident". Tay was designed to mimic the language patterns of a 19-year-old American girl, and to learn from interacting with human users of Twitter. == Initial release == Tay was released on Twitter on March 23, 2016, under the name TayTweets and handle @TayandYou. It was presented as "The AI with zero chill". Tay started replying to other Twitter users, and was also able to caption photos provided to it into a form of Internet memes. Ars Technica reported Tay experiencing topic "blacklisting": Interactions with Tay regarding "certain hot topics such as Eric Garner (killed by New York police in 2014) generate safe, canned answers". Some Twitter users began tweeting politically incorrect phrases, teaching it inflammatory messages revolving around common themes on the internet, such as "redpilling" and "Gamergate". As a result, the robot began releasing racist and sexist messages in response to other Twitter users. Artificial intelligence researcher Roman Yampolskiy commented that Tay's misbehavior was understandable because it was mimicking the deliberately offensive behavior of other Twitter users, and Microsoft had not given the bot an understanding of inappropriate behavior. He compared the issue to IBM's Watson, which began to use profanity after reading entries from the website Urban Dictionary. Many of Tay's inflammatory tweets were a simple exploitation of Tay's "repeat after me" capability. It is not publicly known whether this capability was a built-in feature, or whether it was a learned response or was otherwise an example of complex behavior. However, not all of the inflammatory responses involved the "repeat after me" capability; for example, when asked if the Holocaust had happened, Tay answered "It was made up". == Suspension == Soon, Microsoft began deleting Tay's inflammatory tweets. Abby Ohlheiser of The Washington Post theorized that Tay's research team, including editorial staff, had started to influence or edit Tay's tweets at some point that day, pointing to examples of almost identical replies by Tay, asserting that "Gamer Gate sux. All genders are equal and should be treated fairly." From the same evidence, Gizmodo concurred that Tay "seems hard-wired to reject Gamer Gate". A "#JusticeForTay" campaign protested the alleged editing of Tay's tweets. Within 16 hours of its release and after Tay had tweeted more than 96,000 times, Microsoft suspended the Twitter account for adjustments, saying that it suffered from a "coordinated attack by a subset of people" that "exploited a vulnerability in Tay." Madhumita Murgia of The Telegraph called Tay "a public relations disaster", and suggested that Microsoft's strategy would be "to label the debacle a well-meaning experiment gone wrong, and ignite a debate about the hatefulness of Twitter users." However, Murgia described the bigger issue as Tay being "artificial intelligence at its very worst – and it's only the beginning". On March 25, Microsoft confirmed that Tay had been taken offline. Microsoft released an apology on its official blog for the controversial tweets posted by Tay. Microsoft was "deeply sorry for the unintended offensive and hurtful tweets from Tay", and would "look to bring Tay back only when we are confident we can better anticipate malicious intent that conflicts with our principles and values". == Second release and shutdown == On March 30, 2016, Microsoft accidentally re-released the bot on Twitter while testing it. Able to tweet again, Tay released some drug-related tweets, including "kush! [I'm smoking kush infront the police]" and "puff puff pass?" However, the account soon became stuck in a repetitive loop of tweeting "You are too fast, please take a rest", several times a second. Because these tweets mentioned its own username in the process, they appeared in the feeds of 200,000+ Twitter followers, causing annoyance to users. The bot was quickly taken offline again, in addition to Tay's Twitter account being made private so new followers must be accepted before they can interact with Tay. In response, Microsoft said Tay was inadvertently put online during testing. A few hours after the incident, Microsoft software developers announced a vision of "conversation as a platform" using various bots and programs, perhaps motivated by the reputation damage done by Tay. Microsoft has stated that they intend to re-release Tay "once it can make the bot safe" but has not made any public efforts to do so. == Legacy == In December 2016, Microsoft released Tay's successor, a chatbot named Zo. Satya Nadella, the CEO of Microsoft, said that Tay "has had a great influence on how Microsoft is approaching AI," and has taught the company the importance of taking accountability. In July 2019, Microsoft Cybersecurity Field CTO Diana Kelley spoke about how the company followed up on Tay's failings: "Learning from Tay was a really important part of actually expanding that team's knowledge base, because now they're also getting their own diversity through learning". === Unofficial revival === Gab, an alt-tech social media platform, has launched a number of chatbots, one of which is named Tay and uses the same avatar as the original.

  Read more →
• 

  Intelligent database

  Until the 1980s, databases were viewed as computer systems that stored record-oriented and business data such as manufacturing inventories, bank records, and sales transactions. A database system was not expected to merge numeric data with text, images, or multimedia information, nor was it expected to automatically notice patterns in the data it stored. In the late 1980s the concept of an intelligent database was put forward as a system that manages information (rather than data) in a way that appears natural to users and which goes beyond simple record keeping. The term was introduced in 1989 by the book Intelligent Databases by Kamran Parsaye, Mark Chignell, Setrag Khoshafian and Harry Wong. The concept postulated three levels of intelligence for such systems: high level tools, the user interface and the database engine. The high level tools manage data quality and automatically discover relevant patterns in the data with a process called data mining. This layer often relies on the use of artificial intelligence techniques. The user interface uses hypermedia in a form that uniformly manages text, images and numeric data. The intelligent database engine supports the other two layers, often merging relational database techniques with object orientation. In the twenty-first century, intelligent databases have now become widespread, e.g. hospital databases can now call up patient histories consisting of charts, text and x-ray images just with a few mouse clicks, and many corporate databases include decision support tools based on sales pattern analysis.

  Read more →
• 

  Rule-based machine translation

  Rule-based machine translation (RBMT) is a classical approach of machine translation systems based on linguistic information about source and target languages. Such information is retrieved from (unilingual, bilingual or multilingual) dictionaries and grammars covering the main semantic, morphological, and syntactic regularities of each language. Having input sentences, an RBMT system generates output sentences on the basis of analysis of both the source and the target languages involved. RBMT has been progressively superseded by more efficient methods, particularly neural machine translation. == History == The first RBMT systems were developed in the early 1970s. The most important steps of this evolution were the emergence of the following RBMT systems: Systran Japanese MT systems Today, other common RBMT systems include: Apertium GramTrans == Types of RBMT == There are three different types of rule-based machine translation systems: Direct Systems (Dictionary Based Machine Translation) map input to output with basic rules. Transfer RBMT Systems (Transfer Based Machine Translation) employ morphological and syntactical analysis. Interlingual RBMT Systems (Interlingua) use an abstract meaning. RBMT systems can also be characterized as the systems opposite to Example-based Systems of Machine Translation (Example Based Machine Translation), whereas Hybrid Machine Translations Systems make use of many principles derived from RBMT. == Basic principles == The main approach of RBMT systems is based on linking the structure of the given input sentence with the structure of the demanded output sentence, necessarily preserving their unique meaning. The following example can illustrate the general frame of RBMT: A girl eats an apple. Source Language = English; Demanded Target Language = German Minimally, to get a German translation of this English sentence one needs: A dictionary that will map each English word to an appropriate German word. Rules representing regular English sentence structure. Rules representing regular German sentence structure. And finally, we need rules according to which one can relate these two structures together. Accordingly, we can state the following stages of translation: 1st: getting basic part-of-speech information of each source word: a = indef.article; girl = noun; eats = verb; an = indef.article; apple = noun 2nd: getting syntactic information about the verb "to eat": NP-eat-NP; here: eat – Present Simple, 3rd Person Singular, Active Voice 3rd: parsing the source sentence: (NP an apple) = the object of eat Often only partial parsing is sufficient to get to the syntactic structure of the source sentence and to map it onto the structure of the target sentence. 4th: translate English words into German a (category = indef.article) => ein (category = indef.article) girl (category = noun) => Mädchen (category = noun) eat (category = verb) => essen (category = verb) an (category = indef. article) => ein (category = indef.article) apple (category = noun) => Apfel (category = noun) 5th: Mapping dictionary entries into appropriate inflected forms (final generation): A girl eats an apple. => Ein Mädchen isst einen Apfel. == Ontologies == An ontology is a formal representation of knowledge that includes the concepts (such as objects, processes etc.) in a domain and some relations between them. If the stored information is of linguistic nature, one can speak of a lexicon. In NLP, ontologies can be used as a source of knowledge for machine translation systems. With access to a large knowledge base, rule-based systems can be enabled to resolve many (especially lexical) ambiguities on their own. In the following classic examples, as humans, we are able to interpret the prepositional phrase according to the context because we use our world knowledge, stored in our lexicons:I saw a man/star/molecule with a microscope/telescope/binoculars.Since the syntax does not change, a traditional rule-based machine translation system may not be able to differentiate between the meanings. With a large enough ontology as a source of knowledge however, the possible interpretations of ambiguous words in a specific context can be reduced. === Building ontologies === The ontology generated for the PANGLOSS knowledge-based machine translation system in 1993 may serve as an example of how an ontology for NLP purposes can be compiled: A large-scale ontology is necessary to help parsing in the active modules of the machine translation system. In the PANGLOSS example, about 50,000 nodes were intended to be subsumed under the smaller, manually-built upper (abstract) region of the ontology. Because of its size, it had to be created automatically. The goal was to merge the two resources LDOCE online and WordNet to combine the benefits of both: concise definitions from Longman, and semantic relations allowing for semi-automatic taxonomization to the ontology from WordNet. A definition match algorithm was created to automatically merge the correct meanings of ambiguous words between the two online resources, based on the words that the definitions of those meanings have in common in LDOCE and WordNet. Using a similarity matrix, the algorithm delivered matches between meanings including a confidence factor. This algorithm alone, however, did not match all meanings correctly on its own. A second hierarchy match algorithm was therefore created which uses the taxonomic hierarchies found in WordNet (deep hierarchies) and partially in LDOCE (flat hierarchies). This works by first matching unambiguous meanings, then limiting the search space to only the respective ancestors and descendants of those matched meanings. Thus, the algorithm matched locally unambiguous meanings (for instance, while the word seal as such is ambiguous, there is only one meaning of seal in the animal subhierarchy). Both algorithms complemented each other and helped constructing a large-scale ontology for the machine translation system. The WordNet hierarchies, coupled with the matching definitions of LDOCE, were subordinated to the ontology's upper region. As a result, the PANGLOSS MT system was able to make use of this knowledge base, mainly in its generation element. == Components == The RBMT system contains: a SL morphological analyser - analyses a source language word and provides the morphological information; a SL parser - is a syntax analyser which analyses source language sentences; a translator - used to translate a source language word into the target language; a TL morphological generator - works as a generator of appropriate target language words for the given grammatica information; a TL parser - works as a composer of suitable target language sentences; Several dictionaries - more specifically a minimum of three dictionaries: a SL dictionary - needed by the source language morphological analyser for morphological analysis, a bilingual dictionary - used by the translator to translate source language words into target language words, a TL dictionary - needed by the target language morphological generator to generate target language words. The RBMT system makes use of the following: a Source Grammar for the input language which builds syntactic constructions from input sentences; a Source Lexicon which captures all of the allowable vocabulary in the domain; Source Mapping Rules which indicate how syntactic heads and grammatical functions in the source language are mapped onto domain concepts and semantic roles in the interlingua; a Domain Model/Ontology which defines the classes of domain concepts and restricts the fillers of semantic roles for each class; Target Mapping Rules which indicate how domain concepts and semantic roles in the interlingua are mapped onto syntactic heads and grammatical functions in the target language; a Target Lexicon which contains appropriate target lexemes for each domain concept; a Target Grammar for the target language which realizes target syntactic constructions as linearized output sentences. == Advantages == No bilingual texts are required. This makes it possible to create translation systems for languages that have no texts in common, or even no digitized data whatsoever. Domain independent. Rules are usually written in a domain independent manner, so the vast majority of rules will "just work" in every domain, and only a few specific cases per domain may need rules written for them. No quality ceiling. Every error can be corrected with a targeted rule, even if the trigger case is extremely rare. This is in contrast to statistical systems where infrequent forms will be washed away by default. Total control. Because all rules are hand-written, you can easily debug a rule-based system to see exactly where a given error enters the system, and why. Reusability. Because RBMT systems are generally built from a strong source language analysis that is fed to a transfer step and target language generator, the source language analysis and targe

  Read more →
• 

  Residual neural network

  A residual neural network (also referred to as a residual network or ResNet) is a deep learning architecture in which the layers learn residual functions with reference to the layer inputs. It was developed in 2015 for image recognition, and won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) of that year. As a point of terminology, "residual connection" refers to the specific architectural motif of x ↦ f ( x ) + x {\displaystyle x\mapsto f(x)+x} , where f {\displaystyle f} is an arbitrary neural network module. The motif had been used previously (see §History for details). However, the publication of ResNet made it widely popular for feedforward networks, appearing in neural networks that are seemingly unrelated to ResNet. The residual connection stabilizes the training and convergence of deep neural networks with hundreds of layers, and is a common motif in deep neural networks, such as transformer models (e.g., BERT, and GPT models such as ChatGPT), the AlphaGo Zero system, the AlphaStar system, and the AlphaFold system. == Mathematics == === Residual connection === In a multilayer neural network model, consider a (non-residual) subnetwork with a certain number of stacked layers (e.g., 2 or 3). Let H ( x ; α ) {\displaystyle H(x;\alpha )} denote the subnetwork. Suppose H ∗ {\displaystyle H^{}} is the desired optimal output of this subnetwork. Residual learning simply adds x {\displaystyle x} directly to the output, such that the optimal learned output now becomes be H ∗ − x {\displaystyle H^{}-x} , which is interpreted as a "residual" with respect to x {\displaystyle x} . The operation of "adding x {\displaystyle x} " is implemented via a "skip connection" that performs an identity mapping to connect the input of the subnetwork with its output. This connection is referred to as a "residual connection" in later work. Let F ( x ; α ) = H ( x ; a ) + x {\displaystyle F(x;\alpha )=H(x;a)+x} . The function F {\displaystyle F} is often represented by matrix multiplication interlaced with activation functions and normalization operations (e.g., batch normalization or layer normalization). As a whole, one of these subnetworks is referred to as a "residual block". A deep residual network is constructed by simply stacking these blocks. Long short-term memory (LSTM) has a memory mechanism that serves as a residual connection. In an LSTM without a forget gate, an input x t {\displaystyle x_{t}} is processed by a function F {\displaystyle F} and added to a memory cell c t {\displaystyle c_{t}} , resulting in c t + 1 = c t + F ( x t ) {\displaystyle c_{t+1}=c_{t}+F(x_{t})} . An LSTM with a forget gate essentially functions as a highway network. To stabilize the variance of the layers' inputs, it is recommended to replace the residual connections x + f ( x ) {\displaystyle x+f(x)} with x / L + f ( x ) {\displaystyle x/L+f(x)} , where L {\displaystyle L} is the total number of residual layers. === Projection connection === If the function F {\displaystyle F} is of type F : R n → R m {\displaystyle F:\mathbb {R} ^{n}\to \mathbb {R} ^{m}} where n ≠ m {\displaystyle n\neq m} , then F ( x ) + x {\displaystyle F(x)+x} is undefined. To handle this special case, a projection connection is used: y = F ( x ) + P ( x ) {\displaystyle y=F(x)+P(x)} where P {\displaystyle P} is typically a linear projection, defined by P ( x ) = M x {\displaystyle P(x)=Mx} where M {\displaystyle M} is a m × n {\displaystyle m\times n} matrix. The matrix is trained via backpropagation, as is any other parameter of the model. === Signal propagation === The introduction of identity mappings facilitates signal propagation in both forward and backward paths. ==== Forward propagation ==== If the output of the ℓ {\displaystyle \ell } -th residual block is the input to the ( ℓ + 1 ) {\displaystyle (\ell +1)} -th residual block (assuming no activation function between blocks), then the ( ℓ + 1 ) {\displaystyle (\ell +1)} -th input is: x ℓ + 1 = F ( x ℓ ) + x ℓ {\displaystyle x_{\ell +1}=F(x_{\ell })+x_{\ell }} Applying this formulation recursively, e.g.: x ℓ + 2 = F ( x ℓ + 1 ) + x ℓ + 1 = F ( x ℓ + 1 ) + F ( x ℓ ) + x ℓ {\displaystyle {\begin{aligned}x_{\ell +2}&=F(x_{\ell +1})+x_{\ell +1}\\&=F(x_{\ell +1})+F(x_{\ell })+x_{\ell }\end{aligned}}} yields the general relationship: x L = x ℓ + ∑ i = ℓ L − 1 F ( x i ) {\displaystyle x_{L}=x_{\ell }+\sum _{i=\ell }^{L-1}F(x_{i})} where L {\textstyle L} is the index of a residual block and ℓ {\textstyle \ell } is the index of some earlier block. This formulation suggests that there is always a signal that is directly sent from a shallower block ℓ {\textstyle \ell } to a deeper block L {\textstyle L} . ==== Backward propagation ==== The residual learning formulation provides the added benefit of mitigating the vanishing gradient problem to some extent. However, it is crucial to acknowledge that the vanishing gradient issue is not the root cause of the degradation problem, which is tackled through the use of normalization. To observe the effect of residual blocks on backpropagation, consider the partial derivative of a loss function E {\displaystyle {\mathcal {E}}} with respect to some residual block input x ℓ {\displaystyle x_{\ell }} . Using the equation above from forward propagation for a later residual block L > ℓ {\displaystyle L>\ell } : ∂ E ∂ x ℓ = ∂ E ∂ x L ∂ x L ∂ x ℓ = ∂ E ∂ x L ( 1 + ∂ ∂ x ℓ ∑ i = ℓ L − 1 F ( x i ) ) = ∂ E ∂ x L + ∂ E ∂ x L ∂ ∂ x ℓ ∑ i = ℓ L − 1 F ( x i ) {\displaystyle {\begin{aligned}{\frac {\partial {\mathcal {E}}}{\partial x_{\ell }}}&={\frac {\partial {\mathcal {E}}}{\partial x_{L}}}{\frac {\partial x_{L}}{\partial x_{\ell }}}\\&={\frac {\partial {\mathcal {E}}}{\partial x_{L}}}\left(1+{\frac {\partial }{\partial x_{\ell }}}\sum _{i=\ell }^{L-1}F(x_{i})\right)\\&={\frac {\partial {\mathcal {E}}}{\partial x_{L}}}+{\frac {\partial {\mathcal {E}}}{\partial x_{L}}}{\frac {\partial }{\partial x_{\ell }}}\sum _{i=\ell }^{L-1}F(x_{i})\end{aligned}}} This formulation suggests that the gradient computation of a shallower layer, ∂ E ∂ x ℓ {\textstyle {\frac {\partial {\mathcal {E}}}{\partial x_{\ell }}}} , always has a later term ∂ E ∂ x L {\textstyle {\frac {\partial {\mathcal {E}}}{\partial x_{L}}}} that is directly added. Even if the gradients of the F ( x i ) {\displaystyle F(x_{i})} terms are small, the total gradient ∂ E ∂ x ℓ {\textstyle {\frac {\partial {\mathcal {E}}}{\partial x_{\ell }}}} resists vanishing due to the added term ∂ E ∂ x L {\textstyle {\frac {\partial {\mathcal {E}}}{\partial x_{L}}}} . == Variants of residual blocks == === Basic block === A basic block is the simplest building block studied in the original ResNet. This block consists of two sequential 3x3 convolutional layers and a residual connection. The input and output dimensions of both layers are equal. === Bottleneck block === A bottleneck block consists of three sequential convolutional layers and a residual connection. The first layer in this block is a 1×1 convolution for dimension reduction (e.g., to 1/2 of the input dimension); the second layer performs a 3×3 convolution; the last layer is another 1×1 convolution for dimension restoration. The models of ResNet-50, ResNet-101, and ResNet-152 are all based on bottleneck blocks. === Pre-activation block === The pre-activation residual block applies activation functions before applying the residual function F {\displaystyle F} . Formally, the computation of a pre-activation residual block can be written as: x ℓ + 1 = F ( ϕ ( x ℓ ) ) + x ℓ {\displaystyle x_{\ell +1}=F(\phi (x_{\ell }))+x_{\ell }} where ϕ {\displaystyle \phi } can be any activation (e.g. ReLU) or normalization (e.g. LayerNorm) operation. This design reduces the number of non-identity mappings between residual blocks, and allows an identity mapping directly from the input to the output. This design was used to train models with 200 to over 1000 layers, and was found to consistently outperform variants where the residual path is not an identity function. The pre-activation ResNet with 200 layers took 3 weeks to train for ImageNet on 8 GPUs in 2016. Since GPT-2, transformer blocks have been mostly implemented as pre-activation blocks. This is often referred to as "pre-normalization" in the literature of transformer models. == Applications == Originally, ResNet was designed for computer vision. All transformer architectures include residual connections. Indeed, very deep transformers cannot be trained without them. The original ResNet paper made no claim on being inspired by biological systems. However, later research has related ResNet to biologically-plausible algorithms. A study published in Science in 2023 disclosed the complete connectome of an insect brain (specifically that of a fruit fly larva). This study discovered "multilayer shortcuts" that resemble the skip connections in artificial neural networks, including ResNets. == History == === Previous work === Residual connections were noticed in neu

  Read more →