AI Face Upscale

AI Face Upscale — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

• 

  Structured-light 3D scanner

  A structured-light 3D scanner is a device used to capture the three-dimensional shape of an object by projecting light patterns, such as grids or stripes, onto its surface. The deformation of these patterns is recorded by cameras and processed using specialized algorithms to generate a detailed 3D model. Structured-light 3D scanning is widely employed in fields such as industrial design, quality control, cultural heritage preservation, augmented reality gaming, and medical imaging. Compared to laser-based 3D scanning, structured-light scanners use non-coherent light sources, such as LEDs or projectors, which enable faster data acquisition and eliminate potential safety concerns associated with lasers. However, the accuracy of structured-light scanning can be influenced by external factors, including ambient lighting conditions and the reflective properties of the scanned object. == Principle == Projecting a narrow band of light onto a three-dimensional surface creates a line of illumination that appears distorted when viewed from perspectives other than that of the projector. This distortion can be analyzed to reconstruct the geometry of the surface, a technique known as light sectioning. Projecting patterns composed of multiple stripes or arbitrary fringes simultaneously enables the acquisition of numerous data points at once, improving scanning speed. While various structured light projection techniques exist, parallel stripe patterns are among the most commonly used. By analyzing the displacement of these stripes, the three-dimensional coordinates of surface details can be accurately determined. === Generation of light patterns === Two major methods of stripe pattern generation have been established: Laser interference and projection. The laser interference method works with two wide planar laser beam fronts. Their interference results in regular, equidistant line patterns. Different pattern sizes can be obtained by changing the angle between these beams. The method allows for the exact and easy generation of very fine patterns with unlimited depth of field. Disadvantages are high cost of implementation, difficulties providing the ideal beam geometry, and laser typical effects like speckle noise and the possible self interference with beam parts reflected from objects. Typically, there is no means of modulating individual stripes, such as with Gray codes. The projection method uses incoherent light and basically works like a video projector. Patterns are usually generated by passing light through a digital spatial light modulator, typically based on one of the three currently most widespread digital projection technologies, transmissive liquid crystal, reflective liquid crystal on silicon (LCOS) or digital light processing (DLP; moving micro mirror) modulators, which have various comparative advantages and disadvantages for this application. Other methods of projection could be and have been used, however. Patterns generated by digital display projectors have small discontinuities due to the pixel boundaries in the displays. Sufficiently small boundaries however can practically be neglected as they are evened out by the slightest defocus. A typical measuring assembly consists of one projector and at least one camera. For many applications, two cameras on opposite sides of the projector have been established as useful. Invisible (or imperceptible) structured light uses structured light without interfering with other computer vision tasks for which the projected pattern will be confusing. Example methods include the use of infrared light or of extremely high framerates alternating between two exact opposite patterns. === Calibration === Geometric distortions by optics and perspective must be compensated by a calibration of the measuring equipment, using special calibration patterns and surfaces. A mathematical model is used for describing the imaging properties of projector and cameras. Essentially based on the simple geometric properties of a pinhole camera, the model also has to take into account the geometric distortions and optical aberration of projector and camera lenses. The parameters of the camera as well as its orientation in space can be determined by a series of calibration measurements, using photogrammetric bundle adjustment. === Analysis of stripe patterns === There are several depth cues contained in the observed stripe patterns. The displacement of any single stripe can directly be converted into 3D coordinates. For this purpose, the individual stripe has to be identified, which can for example be accomplished by tracing or counting stripes (pattern recognition method). Another common method projects alternating stripe patterns, resulting in binary Gray code sequences identifying the number of each individual stripe hitting the object. An important depth cue also results from the varying stripe widths along the object surface. Stripe width is a function of the steepness of a surface part, i.e. the first derivative of the elevation. Stripe frequency and phase deliver similar cues and can be analyzed by a Fourier transform. Finally, the wavelet transform has recently been discussed for the same purpose. In many practical implementations, series of measurements combining pattern recognition, Gray codes and Fourier transform are obtained for a complete and unambiguous reconstruction of shapes. Another method also belonging to the area of fringe projection has been demonstrated, utilizing the depth of field of the camera. It is also possible to use projected patterns primarily as a means of structure insertion into scenes, for an essentially photogrammetric acquisition. === Precision and range === The optical resolution of fringe projection methods depends on the width of the stripes used and their optical quality. It is also limited by the wavelength of light. An extreme reduction of stripe width proves inefficient due to limitations in depth of field, camera resolution and display resolution. Therefore, the phase shift method has been widely established: A number of at least 3, typically about 10 exposures are taken with slightly shifted stripes. The first theoretical deductions of this method relied on stripes with a sine wave shaped intensity modulation, but the methods work with "rectangular" modulated stripes, as delivered from LCD or DLP displays as well. By phase shifting, surface detail of e.g. 1/10 the stripe pitch can be resolved. Current optical stripe pattern profilometry hence allows for detail resolutions down to the wavelength of light, below 1 micrometer in practice or, with larger stripe patterns, to approx. 1/10 of the stripe width. Concerning level accuracy, interpolating over several pixels of the acquired camera image can yield a reliable height resolution and also accuracy, down to 1/50 pixel. Arbitrarily large objects can be measured with accordingly large stripe patterns and setups. Practical applications are documented involving objects several meters in size. Typical accuracy figures are: Planarity of a 2-foot (0.61 m) wide surface, to 10 micrometres (0.00039 in). Shape of a motor combustion chamber to 2 micrometres (7.9×10−5 in) (elevation), yielding a volume accuracy 10 times better than with volumetric dosing. Shape of an object 2 inches (51 mm) large, to about 1 micrometre (3.9×10−5 in) Radius of a blade edge of e.g. 10 micrometres (0.00039 in), to ±0.4 μm === Navigation === As the method can measure shapes from only one perspective at a time, complete 3D shapes have to be combined from different measurements in different angles. This can be accomplished by attaching marker points to the object and combining perspectives afterwards by matching these markers. The process can be automated, by mounting the object on a motorized turntable on robotic inspection cell, or CNC positioning device. Markers can as well be applied on a positioning device instead of the object itself. The 3D data gathered can be used to retrieve CAD (computer aided design) data and models from existing components (reverse engineering), hand formed samples or sculptures, natural objects or artifacts. === Challenges === As with all optical methods, reflective or transparent surfaces raise difficulties. Reflections cause light to be reflected either away from the camera or right into its optics. In both cases, the dynamic range of the camera can be exceeded. Transparent or semi-transparent surfaces also cause major difficulties. In these cases, coating the surfaces with a thin opaque lacquer just for measuring purposes is a common practice. A recent method handles highly reflective and specular objects by inserting a 1-dimensional diffuser between the light source (e.g., projector) and the object to be scanned. Alternative optical techniques have been proposed for handling perfectly transparent and specular objects. Double reflections and inter-reflections can cause the stripe pattern to be overlaid with unwanted ligh

  Read more →
• 

  Top 10 Conversational AI Platforms Compared (2026)

  In search of the best conversational AI platform? An conversational AI platform is software that uses machine learning to help you get more done — it turns a rough idea into a polished result in seconds. When choosing one, weigh output quality, pricing, export formats, and how well it fits the tools you already use. Whether you are a beginner or a pro, the right conversational AI platform slots into your workflow and pays for itself fast. We tested the leading options and ranked them by quality, value, and ease of use.

  Read more →
• 

  Isolation forest

  Isolation forest is an unsupervised learning algorithm for anomaly detection that works on the principle of isolating anomalies, instead of the most common techniques of profiling normal points. In statistics, an anomaly (a.k.a. outlier) is an observation or event that deviates so much from other events to arouse suspicion it was generated by a different mean. For example, the graph in Fig.1 represents ingress traffic to a web server, expressed as the number of requests in 3-hours intervals, for a period of one month. It is quite evident by simply looking at the picture that some points (marked with a red circle) are unusually high, to the point of inducing suspect that the web server might have been under attack at that time. On the other hand, the flat segment indicated by the red arrow also seems unusual and might possibly be a sign that the server was down during that time period. Anomalies in a big dataset may follow very complicated patterns, which are difficult to detect "by eye" in the great majority of cases. This is the reason why the field of anomaly detection is well suited for the application of machine learning techniques. The most common techniques employed for anomaly detection are based on the construction of a profile of what is "normal": anomalies are reported as those instances in the dataset that do not conform to the normal profile. Isolation Forest uses a different approach: instead of trying to build a model of normal instances, it explicitly isolates anomalous points in the dataset. The main advantage of this approach is the possibility of exploiting sampling techniques to an extent that is not allowed to the profile-based methods, creating a very fast algorithm with a low memory demand. == History == The Isolation Forest (iForest) algorithm was initially proposed by Fei Tony Liu, Kai Ming Ting and Zhi-Hua Zhou in 2008. The authors took advantage of two quantitative properties of anomalous data points in a sample, that is: they are the minority consisting of fewer instances and they have attribute-values that are very different from those of normal instances Since anomalies are typically few and very different from the other points in the sample, they must be easier to "isolate" compared to normal points. On the basis of this principle, Isolation Forest builds an ensemble of "Isolation Trees" (iTrees) for the data set and marks as anomalies the points that have short average path lengths on the iTrees. In a later paper, published in 2012 the same authors described a set of experiments to prove that iForest: has a low linear time complexity and a small memory requirement is able to deal with high dimensional data with irrelevant attributes can be trained with or without anomalies in the training set can provide detection results with different levels of granularity without re-training In 2013 Zhiguo Ding and Minrui Fei proposed a framework based on iForest to resolve the problem of detecting anomalies in streaming data. More application of iForest to streaming data are described in papers by Swee Chuan Tan et al., G. A. Susto et al. and Yu Weng et al. One of the main problems of the application of iForest to anomaly detection was not with the model itself, but rather in the way the "anomaly score" was computed. This problem was highlighted by Sahand Hariri, Matias Carrasco Kind and Robert J. Brunner in a 2018 paper, wherein they proposed an improved iForest model named Extended Isolation Forest (EIF). In the same paper the authors describe the improvements made to the original model and how they are able to enhance the consistency and reliability of the anomaly score produced for a given data point. == Algorithm == At the basis of the Isolation Forest algorithm there is the tendency of anomalous instances in a dataset to be easier to separate from the rest of the sample (isolate), compared to normal points. In order to isolate a data point the algorithm recursively generates partitions on the sample by randomly selecting an attribute and then randomly selecting a split value for the attribute, between the minimum and maximum values allowed for that attribute. An example of random partitioning in a 2D dataset of normally distributed points is given in Fig. 2 for a non-anomalous point and Fig. 3 for a point that's more likely to be an anomaly. It is apparent from the pictures how anomalies require fewer random partitions to be isolated, compared to normal points. From a mathematical point of view, recursive partitioning can be represented by a tree structure named Isolation Tree, while the number of partitions required to isolate a point can be interpreted as the length of the path, within the tree, to reach a terminating node starting from the root. For example, the path length of point xi in Fig. 2 is greater than the path length of xj in Fig. 3. More formally, let X = { x1, ..., xn } be a set of d-dimensional points and X' ⊂ X a subset of X. An Isolation Tree (iTree) is defined as a data structure with the following properties: for each node T in the Tree, T is either an external-node with no child, or an internal-node with one "test" and exactly two daughter nodes (Tl, Tr) a test at node T consists of an attribute q and a split value p such that the test q < p determines the traversal of a data point to either Tl or Tr. In order to build an iTree, the algorithm recursively divides X' by randomly selecting an attribute q and a split value p, until either (i) the node has only one instance or (ii) all data at the node have the same values. When the iTree is fully grown, each point in X is isolated at one of the external nodes. Intuitively, the anomalous points are those (easier to isolate, hence) with the smaller path length in the tree, where the path length h(xi) of point x i ∈ X {\displaystyle x_{i}\in X} is defined as the number of edges xi traverses from the root node to get to an external node. A probabilistic explanation of iTree is provided in the iForest original paper. == Properties of Isolation Forest == Sub-sampling: since iForest does not need to isolate all of normal instances, it can frequently ignore the big majority of the training sample. As a consequence, iForest works very well when the sampling size is kept small, a property that is in contrast with the great majority of existing methods, where large sampling size is usually desirable. Swamping: when normal instances are too close to anomalies, the number of partitions required to separate anomalies increases, a phenomena known as swamping, which makes it more difficult for iForest to discriminate between anomalies and normal points. One of the main reasons for swamping is the presence of too many data for the purpose of anomaly detection, which implies one possible solution to the problem is sub-sampling. Since iForest respond very well to sub-sampling in terms of performance, the reduction of the number of points in the sample is also a good way to reduce the effect of swamping. Masking: when the number of anomalies is high it is possible that some of those aggregate in a dense and large cluster, making it more difficult to separate the single anomalies and, in turn, to detect such points as anomalous. Similarly to swamping, this phenomena (known as "masking") is also more likely when the number of points in the sample is big, and can be alleviated through sub-sampling. High Dimensional Data: one of the main limitation to standard, distance-based methods is their inefficiency in dealing with high dimensional datasets:. The main reason for that is, in a high dimensional space every point is equally sparse, so using a distance-based measure of separation is pretty ineffective. Unfortunately, high-dimensional data also affects the detection performance of iForest, but the performance can be vastly improved by adding a features selection test like Kurtosis to reduce the dimensionality of the sample space. Normal Instances Only: iForest performs well even if the training set does not contain any anomalous point, the reason being that iForest describes data distributions in such a way that high values of the path length h(xi) correspond to the presence of data points. As a consequence, the presence of anomalies is pretty irrelevant to iForest's detection performance. == Anomaly Detection with Isolation Forest == Anomaly detection with Isolation Forest is a process composed of two main stages: in the first stage, a training dataset is used to build iTrees as described in previous sections. in the second stage, each instance in test set is passed through the iTrees build in the previous stage, and a proper "anomaly score" is assigned to the instance using the algorithm described below Once all the instances in the test set have been assigned an anomaly score, it is possible to mark as "anomaly" any point whose score is greater than a predefined threshold, which depends on the domain the analysis is being applied to. === Anomaly Score === Th

  Read more →
• 

  AI Video Generators Reviews: What Actually Works in 2026

  Comparing the best AI video generator? An AI video generator is software that uses machine learning to help you get more done — it lowers the barrier so anyone can produce professional output. Privacy matters too: check whether your data trains the model and whether a no-log or enterprise tier is available. Whether you are a beginner or a pro, the right AI video generator slots into your workflow and pays for itself fast. We tested the leading options and ranked them by quality, value, and ease of use.

  Read more →
• 

  Artificial intelligence of things

  Artificial Intelligence of Things (AIoT) is the combination of artificial intelligence (AI) technologies with the Internet of things (IoT) infrastructure to create systems capable of sensing, learning, and acting on data without continuous human intervention. While IoT focuses on connectivity and sensor data collection, AI enables IoT devices to analyse data in real time and produce actionable outputs, including automated decisions at the edge. == Applications == === Manufacturing and predictive maintenance === Manufacturing accounts for the largest share of AIoT adoption by industry vertical. A common application is predictive maintenance, where sensors measuring vibration, temperature, current draw, and acoustic emissions feed machine learning models trained to detect signatures that precede equipment failure. These systems can flag developing faults weeks or months in advance, and in more advanced deployments can autonomously adjust machine parameters such as motor speed or cooling cycles to delay or prevent failure. === Other industries === In healthcare, AIoT enables remote patient monitoring through wearable devices that collect vital signs and apply AI models to detect anomalies or predict deterioration. In logistics, GPS and telematics sensors combined with AI models support real-time route optimisation, vehicle maintenance prediction, and fuel cost forecasting. Smart building systems use occupancy, temperature, and energy sensors with AI to dynamically adjust HVAC and lighting, reducing energy consumption. == Architecture == AIoT systems typically operate across three layers: a device layer of sensors and actuators that collect data, a connectivity layer that transmits data via protocols such as MQTT or HTTP, and a compute layer where AI models process the data either in the cloud or at the edge. The trend toward edge-based processing, where inference runs on low-cost processors near the data source rather than in a centralised cloud, has accelerated as hardware costs have fallen and applications increasingly require sub-second response times. == Market == Market sizing estimates for AIoT vary significantly depending on scope and definition. Fortune Business Insights valued the AIoT market at USD 35.65 billion in 2023, projecting growth to USD 253.86 billion by 2030 at a compound annual growth rate of 32.4%. Grand View Research estimated the broader market at USD 171.4 billion in 2024 with a CAGR of 31.7% through 2030, reflecting a wider definition that includes AI-integrated hardware components. North America accounted for approximately 40% of global market share in 2024, with the Asia-Pacific region projected as the fastest-growing market.

  Read more →
• 

  Deepset

  deepset is an enterprise software vendor that provides developers with the tools to build production-ready Artificial Intelligence (AI) and natural language processing (NLP) systems, using architectures such as agents, retrieval augmented generation (RAG) and multimodal AI. It was founded in 2018 in Berlin by Milos Rusic, Malte Pietsch, and Timo Möller. deepset authored and maintains the open source software Haystack and its commercial SaaS and self-hosted (VPC, on-prem, air gapped) offering, Haystack Enterprise Platform. (formerly known as deepset Cloud and deepset AI Platform) == History == In June 2018, Milos Rusic, Malte Pietsch, and Timo Möller co-founded deepset in Berlin, Germany. In the same year, the company served first customers who wanted to implement NLP services by tailoring BERT language models to their domain. In July 2019, the company released the initial version of the open source software FARM. In November 2019, the company released the initial version of the open source software Haystack. Throughout 2020 and 2021 deepset published several applied research papers at EMNLP, COLING and ACL, the leading conferences in the area of NLP. In 2020, the research contributions comprised German language models named GBERT and GELECTRA, and a question answering dataset addressing the COVID-19 pandemic called COVID-QA, which was created in collaboration with Intel and has been annotated by biomedical experts. In 2021, the research contributions comprised German models and datasets for question answering and passage retrieval named GermanQuAD and GermanDPR, a semantic answer similarity metric, and an approach for multimodal retrieval of texts and tables to enable question answering on tabular data. Haystack contains implementations of all three contributions, enabling the use of the research through the open source framework. In November 2021, the development of the FARM framework was discontinued and its main features were integrated into the Haystack framework. In April 2022, the company announced its commercial SaaS offering deepset Cloud, which was rebranded in 2025 as Haystack Enterprise Platform supporting SaaS and on-premise deployment options. As of August 2023, the most popular finetuned language model created by deepset was downloaded more than 52 million times. In 2024, deepset was named a Gartner Cool Vendor in AI Engineering. In 2025, deepset was recognized for its growth by WirtschaftsWoche and Sifted and shared partnership integrations and announcements with Meta Llama Stack, MongoDB, NVIDIA, Amazon Web Services (AWS), and PwC. As of September 2025, the Haystack open source AI orchestration framework has more than 24,000 GitHub stars. == Products and applications == Haystack is an open source Python AI Orchestration framework for building custom AI agents and applications with large language models. With its modular building block components, software developers and AI engineers can implement pipelines to build and customize various AI architectures over large document and multimodal data collections, such as agents, retrieval augmented generation (RAG), intelligent document processing (IDP), text-to-SQL as well as document retrieval, semantic search, text generation, question answering, or summarization. Haystack emphasizes context engineering, an approach to AI system design that focuses on explicit control over how contextual information is retrieved, structured, routed to language models, and evaluated after generation. This allows developers to build AI systems with transparent data flow, tool usage, and configurable reasoning processes. Haystack integrates with 90+ model and technology providers including Hugging Face Transformers, Elasticsearch, OpenSearch, OpenAI, Cohere, Anthropic, Mistral and others. Developers can extend these integrations with their own custom components. The framework has an active community on Discord with more than 4k members and GitHub, where so far more than 300 people have contributed to its continuous development, and engage on Meetup. Thousands of organizations use the framework, including public sector leaders like the European Commission and Global 500 enterprises like Airbus, Intel, NVIDIA, Lufthansa, Netflix, Apple, Infineon, Alcatel-Lucent Enterprise, BetterUp, Etalab, Sooth.ai, and Lego. On top of the Haystack open source framework, deepset offers two enterprise offerings to organizations. Haystack Enterprise Starter provides enterprise support on the open source framework from the Haystack engineering team as well as a private GitHub repository with production use case templates and Kubernetes deployment guides. The Haystack Enterprise Platform supports customers at building scalable AI applications by covering the entire process of prototyping, experimentation, deployment, monitoring, and governance. It is built on the Haystack open source framework and is available for hosting in the cloud and self-hosted via VPC, on-premise, or air gapped environments. deepset's enterprise tools are used by organizations including The European Commission, The Economist, Oxford University Press, the German Federal Ministry of Research, Technology, and Space (BMFTR), Manz Verlag, and the German Armed Forces. FARM was an earlier framework for adapting representation models. One of its core concepts was the implementation of adaptive models, which comprised language models and an arbitrary number of prediction heads. FARM supported domain-adaptation and finetuning of these models with advanced options, for example gradient accumulation, cross-validation or automatic mixed-precision training. Its main features were integrated into Haystack in November 2021, and its development was discontinued at that time. == Funding == On August 9, 2023, deepset announced a Series B investment round of $30 million led by Balderton Capital and including participation from existing investors GV, System.One, Lunar Ventures and Harpoon Ventures. On April 28, 2022, deepset announced a Series A investment round of $14 million led by GV, with the participation of Harpoon Ventures, Acequia Capital and a team of experienced commercial open source software and machine learning founders, such as Alex Ratner (Snorkel AI), Mustafa Suleyman (Deepmind), Spencer Kimball (Cockroach Labs), Jeff Hammerbacher (Cloudera) and Emil Eifrem (Neo4j). A previous pre-seed investment round of $1.6 million on March 8, 2021, was led by System.One and Lunar Ventures, who also participated in the subsequent Series A round.

  Read more →
• 

  Unsupervised learning

  Unsupervised learning is a framework in machine learning where, in contrast to supervised learning, algorithms learn patterns exclusively from unlabeled data. Other frameworks in the spectrum of supervisions include weak- or semi-supervision, where a small portion of the data is tagged, and self-supervision. Some researchers consider self-supervised learning a form of unsupervised learning. Conceptually, unsupervised learning divides into the aspects of data, training, algorithm, and downstream applications. Typically, the dataset is harvested cheaply "in the wild", such as massive text corpus obtained by web crawling, with only minor filtering (such as Common Crawl). This compares favorably to supervised learning, where the dataset (such as the ImageNet1000) is typically constructed manually, which is much more expensive. There are algorithms designed specifically for unsupervised learning, such as clustering algorithms like k-means, dimensionality reduction techniques like principal component analysis (PCA), Boltzmann machine learning, and autoencoders. After the rise of deep learning, most large-scale unsupervised learning has been done by training general-purpose neural network architectures by gradient descent, adapted to performing unsupervised learning by designing an appropriate training procedure. Sometimes a trained model can be used as-is, but more often they are modified for downstream applications. For example, the generative pretraining method trains a model to generate a textual dataset, before finetuning it for other applications, such as text classification. As another example, autoencoders are trained to produce good features, which can then be used as a module for other models, such as in a latent diffusion model. == Tasks == Tasks are often categorized as discriminative (recognition) or generative (imagination). Often but not always, discriminative tasks use supervised methods and generative tasks use unsupervised (see Venn diagram); however, the separation is very hazy. For example, object recognition favors supervised learning but unsupervised learning can also cluster objects into groups. Furthermore, as progress marches onward, some tasks employ both methods, and some tasks swing from one to another. For example, image recognition started off as heavily supervised, but became hybrid by employing unsupervised pre-training, and then moved towards supervision again with the advent of dropout, ReLU, and adaptive learning rates. A typical generative task is as follows. At each step, a datapoint is sampled from the dataset, and part of the data is removed, and the model must infer the removed part. This is particularly clear for the denoising autoencoders and BERT. == Neural network architectures == === Training === During the learning phase, an unsupervised network tries to mimic the data it is given and uses the error in its mimicked output to correct itself (i.e. correct its weights and biases). Sometimes the error is expressed as a low probability that the erroneous output occurs, or it might be expressed as an unstable high energy state in the network. In contrast to supervised methods' dominant use of backpropagation, unsupervised learning also employs other methods including: Hopfield learning rule, Boltzmann learning rule, Contrastive Divergence, Wake Sleep, Variational Inference, Maximum Likelihood, Maximum A Posteriori, Gibbs Sampling, and backpropagating reconstruction errors or hidden state reparameterizations. See the table below for more details. === Energy === An energy function is a macroscopic measure of a network's activation state. In Boltzmann machines, it plays the role of the Cost function. This analogy with physics is inspired by Ludwig Boltzmann's analysis of a gas' macroscopic energy from the microscopic probabilities of particle motion p ∝ e − E / k T {\displaystyle p\propto e^{-E/kT}} , where k is the Boltzmann constant and T is temperature. In the RBM network the relation is p = e − E / Z {\displaystyle p=e^{-E}/Z} , where p {\displaystyle p} and E {\displaystyle E} vary over every possible activation pattern and Z = ∑ All Patterns e − E ( pattern ) {\displaystyle \textstyle {Z=\sum _{\scriptscriptstyle {\text{All Patterns}}}e^{-E({\text{pattern}})}}} . To be more precise, p ( a ) = e − E ( a ) / Z {\displaystyle p(a)=e^{-E(a)}/Z} , where a {\displaystyle a} is an activation pattern of all neurons (visible and hidden). Hence, some early neural networks bear the name Boltzmann Machine. Paul Smolensky calls − E {\displaystyle -E\,} the Harmony. A network seeks low energy which is high Harmony. === Networks === This table shows connection diagrams of various unsupervised networks, the details of which will be given in the section Comparison of Networks. Circles are neurons and edges between them are connection weights. As network design changes, features are added on to enable new capabilities or removed to make learning faster. For instance, neurons change between deterministic (Hopfield) and stochastic (Boltzmann) to allow robust output, weights are removed within a layer (RBM) to hasten learning, or connections are allowed to become asymmetric (Helmholtz). Of the networks bearing people's names, only Hopfield worked directly with neural networks. Boltzmann and Helmholtz came before artificial neural networks, but their work in physics and physiology inspired the analytical methods that were used. === History === === Specific Networks === Here, we highlight some characteristics of select networks. The details of each are given in the comparison table below. Hopfield Network Ferromagnetism inspired Hopfield networks. A neuron corresponds to an iron domain with binary magnetic moments Up and Down, and neural connections correspond to the domain's influence on each other. Symmetric connections enable a global energy formulation. During inference the network updates each state using the standard activation step function. Symmetric weights and the right energy functions guarantees convergence to a stable activation pattern. Asymmetric weights are difficult to analyze. Hopfield nets are used as Content Addressable Memories (CAM). Boltzmann Machine These are stochastic Hopfield nets. Their state value is sampled from this pdf as follows: suppose a binary neuron fires with the Bernoulli probability p(1) = 1/3 and rests with p(0) = 2/3. One samples from it by taking a uniformly distributed random number y, and plugging it into the inverted cumulative distribution function, which in this case is the step function thresholded at 2/3. The inverse function = { 0 if x <= 2/3, 1 if x > 2/3 }. Sigmoid Belief Net Introduced by Radford Neal in 1992, this network applies ideas from probabilistic graphical models to neural networks. A key difference is that nodes in graphical models have pre-assigned meanings, whereas Belief Net neurons' features are determined after training. The network is a sparsely connected directed acyclic graph composed of binary stochastic neurons. The learning rule comes from Maximum Likelihood on p(X): Δwij ∝ {\displaystyle \propto } sj (si - pi), where pi = 1 / ( 1 + eweighted inputs into neuron i ). sj's are activations from an unbiased sample of the posterior distribution and this is problematic due to the Explaining Away problem raised by Judea Perl. Variational Bayesian methods uses a surrogate posterior and blatantly disregard this complexity. Deep Belief Network Introduced by Hinton, this network is a hybrid of RBM and Sigmoid Belief Network. The top 2 layers is an RBM and the second layer downwards form a sigmoid belief network. One trains it by the stacked RBM method and then throw away the recognition weights below the top RBM. As of 2009, 3-4 layers seems to be the optimal depth. Helmholtz machine These are early inspirations for the Variational Auto Encoders. Its 2 networks combined into one—forward weights operates recognition and backward weights implements imagination. It is perhaps the first network to do both. Helmholtz did not work in machine learning but he inspired the view of "statistical inference engine whose function is to infer probable causes of sensory input". the stochastic binary neuron outputs a probability that its state is 0 or 1. The data input is normally not considered a layer, but in the Helmholtz machine generation mode, the data layer receives input from the middle layer and has separate weights for this purpose, so it is considered a layer. Hence this network has 3 layers. Variational autoencoder These are inspired by Helmholtz machines and combines probability network with neural networks. An Autoencoder is a 3-layer CAM network, where the middle layer is supposed to be some internal representation of input patterns. The encoder neural network is a probability distribution qφ(z given x) and the decoder network is pθ(x given z). The weights are named phi & theta rather than W and V as in Helmholtz—a cosmetic difference. These 2 networks h

  Read more →
• 

  AI Resume Builders Reviews: What Actually Works in 2026

  Shopping for the best AI resume builder? An AI resume builder is software that uses machine learning to help you get more done — it keeps getting smarter as the underlying models improve. Pricing, accuracy, and the size of the model behind the tool are the three factors that most affect daily usefulness. Whether you are a beginner or a pro, the right AI resume builder slots into your workflow and pays for itself fast. We tested the leading options and ranked them by quality, value, and ease of use.

  Read more →
• 

  EditDV

  EditDV was a video editing software released by Radius, Inc. in late 1997 as an evolution of their earlier Radius Edit product. EditDV was one of the first products providing professional-quality editing of the then new DV format at a relatively affordable cost ($999 including Radius FireWire capture card) and was named "The Best Video Tool of 1998". Originally EditDV was available for Macintosh only but in February 2000 EditDV 2.0 for Windows was released. With version 3.0 EditDV's name was changed to CineStream. == Features == Originally bundled with a FireWire card, EditDV 1.5 got updated into a less expensive software only package for use with the newer PowerMac G3 that came with a FireWire interface. Later, a scaled down version named EditDV 1.6.1 Unplugged was released as a freeware version next to EditDV 2.0. Unlike many other applications at the time which transcoded video to M-JPEG for editing, EditDV provided lossless native editing of the DV format. Only transitions (such as dissolves or wipes), effects (such as rotating or scaling the video, adjusting the audio level, or adding titles) and filters (such as changing the brightness or color balance) needed to be rendered. This also had the disadvantage to not work with analogue video capture. EditDV was built on top of QuickTime and supported QuickTime filters as well as its own built-in effects and transitions. Effects could be animated using keyframes. EditDV 2.0 worked natively with Quicktime MOV format. For Microsoft Windows users, where the standard was AVI, this required the use of a provided external conversion tool afterwards when AVI was wanted. The user interface had a Project window for organising clips into bins, a Sequence window with a multi-track timeline for arranging clips into a program using three-point editing, and Source and Program monitor windows. A finished program could either be exported as a QuickTime movie or written back to DV tape using the "print to video" command. Version 3.0, then renamed CineStream, shifted towards web designers who wanted to add video streaming interactivity to a website. The new feature called EventStream allowed setting clickable hot spots to link to another location, either to another page with a URL or to another video. This feature distinguished CineStream from the rest of the competition. == Products == The EditDV product family included a number of related products, all sharing a similar name: EditDV Video editing software (Mac and Windows) SoftDV A QuickTime software codec for playing DV media, included as part of EditDV (Mac and Windows) MotoDV PCI-based FireWire interface with DV capture software (Mac and Windows) PhotoDV Software to capture high-quality stills from a DV tape using MotoDV hardware (Mac and Windows) RotoDV Software for rotoscoping (painting over video), released in Sept 1999 (Macintosh only) == Name changes and eventual demise == In 1999, the company Radius Inc. changed its name to Digital Origin. In 2000, Digital Origin Inc (and EditDV) was bought by Media 100. In early 2001, Media 100 released an updated version of EditDV under the new name CineStream 3.0. Later that year (October 2001) Media 100 was bought by Autodesk's Discreet Division. CineStream for Macintosh required classic Mac OS. It was never ported to Mac OS X and faced increasing competition on that platform from Apple's own Final Cut Pro application. Development of EditDV/Cinestream was officially discontinued in 2002.

  Read more →
• 

  Best AI Text-to-image Tools in 2026

  Trying to pick the best AI text-to-image tool? An AI text-to-image tool is software that uses machine learning to help you get more done — it scales effortlessly from a single task to thousands. The best picks balance beginner-friendly simplicity with the depth power users need, and they ship updates often. Whether you are a beginner or a pro, the right AI text-to-image tool slots into your workflow and pays for itself fast. This guide breaks down the top picks, their pros and cons, and who each one is best for.

  Read more →
• 

  Localization Industry Standards Association

  Localization Industry Standards Association or LISA was a Swiss-based trade body concerning the translation of computer software (and associated materials) into multiple natural languages, which existed from 1990 to February 2011. It counted among its members most of the large information technology companies of the period, including Adobe, Cisco, Hewlett-Packard, IBM, McAfee, Nokia, Novell and Xerox. LISA played a significant role in representing its partners at the International Organization for Standardization (ISO), and the TermBase eXchange (TBX) standard developed by LISA was submitted to ISO in 2007 and became ISO 30042:2008. LISA also had a presence at the W3C. A number of the LISA standards are used by the OASIS Open Architecture for XML Authoring and Localization framework. LISA shut down on 28 February 2011, and its website went offline shortly afterwards. In the wake of the closure of LISA, the European Telecommunications Standards Institute started an Industry Specification Group (ISG) for localization. The ISG has five work items: Term-Base eXchange (TBX) / ISO 30042:2008 Translation Memory eXchange (TMX), with GALA Segmentation Rules eXchange (SRX) / ISO/CD 24621) Global information management Metrics eXchange – Volume (GMX-V); Another organization that was formed in response to the closure of LISA is Terminology for Large Organizations (TerminOrgs), a consortium of terminology professionals who promote terminology management best practices.

  Read more →
• 

  OCR-A

  OCR-A is a font issued in 1966 and first implemented in 1968. A special font was needed in the early days of computer optical character recognition, when there was a need for a font that could be recognized not only by the computers of that day, but also by humans. OCR-A uses simple, thick strokes to form recognizable characters. The font is monospaced (fixed-width), with the printer required to place glyphs 0.254 cm (0.10 inch) apart, and the reader required to accept any spacing between 0.2286 cm (0.09 inch) and 0.4572 cm (0.18 inch). == Standardization == The OCR-A font was standardized by the American National Standards Institute (ANSI) as ANSI X3.17-1981. X3.4 has since become the INCITS and the OCR-A standard is now called ISO 1073-1:1976. == Implementations == In 1968, American Type Founders produced OCR-A, one of the first optical character recognition typefaces to meet the criteria set by the U.S. Bureau of Standards. The design is simple so that it can be easily read by a machine, but it is more difficult for the human eye to read. As metal type gave way to computer-based typesetting, Tor Lillqvist used Metafont to describe the OCR-A font. That definition was subsequently improved by Richard B. Wales. Their work is available from CTAN. To make the free version of the font more accessible to users of Microsoft Windows, John Sauter converted the Metafont definitions to TrueType using potrace and FontForge in 2004. In 2007, Gürkan Sengün created a Debian package from this implementation. In 2008. Luc Devroye corrected the vertical positioning in John Sauter's implementation, and fixed the name of lower case z. Independently, Matthew Skala used mftrace to convert the Metafont definitions to TrueType format in 2006. In 2011 he released a new version created by rewriting the Metafont definitions to work with METATYPE1, generating outlines directly without an intermediate tracing step. On September 27, 2012, he updated his implementation to version 0.2. In addition to these free implementations of OCR-A, there are also implementations sold by several vendors. As a joke, Tobias Frere-Jones in 1995 created Estupido-Espezial, a redesign with swashes and a long s. It was used in a "technology"-themed section of Rolling Stone. Maxitype designed the OCR-X typeface—based on the OCR-A typeface with OpenType features, alien/technology-themed dingbats and available in six weights (Thin, Light, Regular, Medium, Bold, Black). Japanese typeface foundry Visual Design Laboratory (VDL) designed two typefaces based on the OCR-A typeface: one for Simplified Chinese characters named Jieyouti and one for Japanese characters named Yota G (ヨタG) , both available in five weights (Light, Regular, Medium, Semi Bold, Bold). == Use == Although optical character recognition technology has advanced to the point where such simple fonts are no longer necessary, the OCR-A font has remained in use. Its usage remains widespread in the encoding of checks around the world. Some lock box companies still insist that the account number and amount owed on a bill return form be printed in OCR-A. Also, because of its unusual look, it is sometimes used in advertising and display graphics. Notably, it is used for the subtitles in films and television series such as Blacklist and for the main titles in The Pretender. Additionally, OCR-A is used in the titles and subtitles for the films 13 Hours: The Secret Soldiers of Benghazi and Hoppers (film). It was also used for the logo, branding, and marketing material of the children's toy line Hexbug. == Code points == A font is a set of character shapes, or glyphs. For a computer to use a font, each glyph must be assigned a code point in a character set. When OCR-A was being standardized the usual character coding was the American Standard Code for Information Interchange or ASCII. Not all of the glyphs of OCR-A fit into ASCII, and for five of the characters there were alternate glyphs, which might have suggested the need for a second font. However, for convenience and efficiency all of the glyphs were expected to be accessible in a single font using ASCII coding, with the additional characters placed at coding points that would otherwise have been unused. The modern descendant of ASCII is Unicode, also known as ISO 10646. Unicode contains ASCII and has special provisions for OCR characters, so some implementations of OCR-A have looked to Unicode for guidance on character code assignments. === Pre-Unicode standard representation === The ISO standard ISO 2033:1983, and the corresponding Japanese Industrial Standard JIS X 9010:1984 (originally JIS C 6229–1984), define character encodings for OCR-A, OCR-B and E-13B. For OCR-A, they define a modified 7-bit ASCII set (also known by its ISO-IR number ISO-IR-91) including only uppercase letters, digits, a subset of the punctuation and symbols, and some additional symbols. Codes which are redefined relative to ASCII, as opposed to simply omitted, are listed below: Additionally, the long vertical mark () is encoded at 0x7C, corresponding to the ASCII vertical bar (|). === Dedicated OCR-A characters in Unicode === The following characters have been defined for control purposes and are now in the "Optical Character Recognition" Unicode range 2440–245F: === Space, digits, and unaccented letters === All implementations of OCR-A use U+0020 for space, U+0030 through U+0039 for the decimal digits, U+0041 through U+005A for the unaccented upper case letters, and U+0061 through U+007A for the unaccented lower case letters. === Regular characters === In addition to the digits and unaccented letters, many of the characters of OCR-A have obvious code points in ASCII. Of those that do not, most, including all of OCR-A's accented letters, have obvious code points in Unicode. === Remaining characters === Linotype coded the remaining characters of OCR-A as follows: === Additional characters === The fonts that descend from the work of Tor Lillqvist and Richard B. Wales define four characters not in OCR-A to fill out the ASCII character set. These shapes use the same style as the OCR-A character shapes. They are: Linotype also defines additional characters. === Exceptions === Some implementations do not use the above code point assignments for some characters. ==== PrecisionID ==== The PrecisionID implementation of OCR-A has the following non-standard code points: OCR Hook at U+007E OCR Chair at U+00C1 OCR Fork at U+00C2 Euro Sign at U+0080 ==== Barcodesoft ==== The Barcodesoft implementation of OCR-A has the following non-standard code points: OCR Hook at U+0060 OCR Chair at U+007E OCR Fork at U+005F Long Vertical Mark at U+007C (agrees with Linotype) Character Erase at U+0008 ==== Morovia ==== The Morovia implementation of OCR-A has the following non-standard code points: OCR Hook at U+007E (agrees with PrecisionID) OCR Chair at U+00F0 OCR Fork at U+005F (agrees with Barcodesoft) Long Vertical Mark at U+007C (agrees with Linotype) ==== IDAutomation ==== The IDAutomation implementation of OCR-A has the following non-standard code points: OCR Hook at U+007E (agrees with PrecisionID) OCR Chair at U+00C1 (agrees with PrecisionID) OCR Fork at U+00C2 (agrees with PrecisionID) OCR Belt Buckle at U+00C3 == Sellers of font standards == Hardcopy of ISO 1073-1:1976, distributed through ANSI, from Amazon.com ISO 1073-1 is also available from Techstreet, who distributes standards for ANSI and ISO

  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 →
• 

  AI Background Removers: Free vs Paid (2026)

  Looking for the best AI background remover? An AI background remover is software that uses machine learning to help you get more done — it can save you hours every week by automating repetitive work. Most options offer a generous free tier, with paid plans unlocking higher limits, faster processing, and team features. Whether you are a beginner or a pro, the right AI background remover slots into your workflow and pays for itself fast. Read on for hands-on impressions, pricing tiers, and the standout features that matter.

  Read more →
• 

  Unsupervised learning

  Unsupervised learning is a framework in machine learning where, in contrast to supervised learning, algorithms learn patterns exclusively from unlabeled data. Other frameworks in the spectrum of supervisions include weak- or semi-supervision, where a small portion of the data is tagged, and self-supervision. Some researchers consider self-supervised learning a form of unsupervised learning. Conceptually, unsupervised learning divides into the aspects of data, training, algorithm, and downstream applications. Typically, the dataset is harvested cheaply "in the wild", such as massive text corpus obtained by web crawling, with only minor filtering (such as Common Crawl). This compares favorably to supervised learning, where the dataset (such as the ImageNet1000) is typically constructed manually, which is much more expensive. There are algorithms designed specifically for unsupervised learning, such as clustering algorithms like k-means, dimensionality reduction techniques like principal component analysis (PCA), Boltzmann machine learning, and autoencoders. After the rise of deep learning, most large-scale unsupervised learning has been done by training general-purpose neural network architectures by gradient descent, adapted to performing unsupervised learning by designing an appropriate training procedure. Sometimes a trained model can be used as-is, but more often they are modified for downstream applications. For example, the generative pretraining method trains a model to generate a textual dataset, before finetuning it for other applications, such as text classification. As another example, autoencoders are trained to produce good features, which can then be used as a module for other models, such as in a latent diffusion model. == Tasks == Tasks are often categorized as discriminative (recognition) or generative (imagination). Often but not always, discriminative tasks use supervised methods and generative tasks use unsupervised (see Venn diagram); however, the separation is very hazy. For example, object recognition favors supervised learning but unsupervised learning can also cluster objects into groups. Furthermore, as progress marches onward, some tasks employ both methods, and some tasks swing from one to another. For example, image recognition started off as heavily supervised, but became hybrid by employing unsupervised pre-training, and then moved towards supervision again with the advent of dropout, ReLU, and adaptive learning rates. A typical generative task is as follows. At each step, a datapoint is sampled from the dataset, and part of the data is removed, and the model must infer the removed part. This is particularly clear for the denoising autoencoders and BERT. == Neural network architectures == === Training === During the learning phase, an unsupervised network tries to mimic the data it is given and uses the error in its mimicked output to correct itself (i.e. correct its weights and biases). Sometimes the error is expressed as a low probability that the erroneous output occurs, or it might be expressed as an unstable high energy state in the network. In contrast to supervised methods' dominant use of backpropagation, unsupervised learning also employs other methods including: Hopfield learning rule, Boltzmann learning rule, Contrastive Divergence, Wake Sleep, Variational Inference, Maximum Likelihood, Maximum A Posteriori, Gibbs Sampling, and backpropagating reconstruction errors or hidden state reparameterizations. See the table below for more details. === Energy === An energy function is a macroscopic measure of a network's activation state. In Boltzmann machines, it plays the role of the Cost function. This analogy with physics is inspired by Ludwig Boltzmann's analysis of a gas' macroscopic energy from the microscopic probabilities of particle motion p ∝ e − E / k T {\displaystyle p\propto e^{-E/kT}} , where k is the Boltzmann constant and T is temperature. In the RBM network the relation is p = e − E / Z {\displaystyle p=e^{-E}/Z} , where p {\displaystyle p} and E {\displaystyle E} vary over every possible activation pattern and Z = ∑ All Patterns e − E ( pattern ) {\displaystyle \textstyle {Z=\sum _{\scriptscriptstyle {\text{All Patterns}}}e^{-E({\text{pattern}})}}} . To be more precise, p ( a ) = e − E ( a ) / Z {\displaystyle p(a)=e^{-E(a)}/Z} , where a {\displaystyle a} is an activation pattern of all neurons (visible and hidden). Hence, some early neural networks bear the name Boltzmann Machine. Paul Smolensky calls − E {\displaystyle -E\,} the Harmony. A network seeks low energy which is high Harmony. === Networks === This table shows connection diagrams of various unsupervised networks, the details of which will be given in the section Comparison of Networks. Circles are neurons and edges between them are connection weights. As network design changes, features are added on to enable new capabilities or removed to make learning faster. For instance, neurons change between deterministic (Hopfield) and stochastic (Boltzmann) to allow robust output, weights are removed within a layer (RBM) to hasten learning, or connections are allowed to become asymmetric (Helmholtz). Of the networks bearing people's names, only Hopfield worked directly with neural networks. Boltzmann and Helmholtz came before artificial neural networks, but their work in physics and physiology inspired the analytical methods that were used. === History === === Specific Networks === Here, we highlight some characteristics of select networks. The details of each are given in the comparison table below. Hopfield Network Ferromagnetism inspired Hopfield networks. A neuron corresponds to an iron domain with binary magnetic moments Up and Down, and neural connections correspond to the domain's influence on each other. Symmetric connections enable a global energy formulation. During inference the network updates each state using the standard activation step function. Symmetric weights and the right energy functions guarantees convergence to a stable activation pattern. Asymmetric weights are difficult to analyze. Hopfield nets are used as Content Addressable Memories (CAM). Boltzmann Machine These are stochastic Hopfield nets. Their state value is sampled from this pdf as follows: suppose a binary neuron fires with the Bernoulli probability p(1) = 1/3 and rests with p(0) = 2/3. One samples from it by taking a uniformly distributed random number y, and plugging it into the inverted cumulative distribution function, which in this case is the step function thresholded at 2/3. The inverse function = { 0 if x <= 2/3, 1 if x > 2/3 }. Sigmoid Belief Net Introduced by Radford Neal in 1992, this network applies ideas from probabilistic graphical models to neural networks. A key difference is that nodes in graphical models have pre-assigned meanings, whereas Belief Net neurons' features are determined after training. The network is a sparsely connected directed acyclic graph composed of binary stochastic neurons. The learning rule comes from Maximum Likelihood on p(X): Δwij ∝ {\displaystyle \propto } sj (si - pi), where pi = 1 / ( 1 + eweighted inputs into neuron i ). sj's are activations from an unbiased sample of the posterior distribution and this is problematic due to the Explaining Away problem raised by Judea Perl. Variational Bayesian methods uses a surrogate posterior and blatantly disregard this complexity. Deep Belief Network Introduced by Hinton, this network is a hybrid of RBM and Sigmoid Belief Network. The top 2 layers is an RBM and the second layer downwards form a sigmoid belief network. One trains it by the stacked RBM method and then throw away the recognition weights below the top RBM. As of 2009, 3-4 layers seems to be the optimal depth. Helmholtz machine These are early inspirations for the Variational Auto Encoders. Its 2 networks combined into one—forward weights operates recognition and backward weights implements imagination. It is perhaps the first network to do both. Helmholtz did not work in machine learning but he inspired the view of "statistical inference engine whose function is to infer probable causes of sensory input". the stochastic binary neuron outputs a probability that its state is 0 or 1. The data input is normally not considered a layer, but in the Helmholtz machine generation mode, the data layer receives input from the middle layer and has separate weights for this purpose, so it is considered a layer. Hence this network has 3 layers. Variational autoencoder These are inspired by Helmholtz machines and combines probability network with neural networks. An Autoencoder is a 3-layer CAM network, where the middle layer is supposed to be some internal representation of input patterns. The encoder neural network is a probability distribution qφ(z given x) and the decoder network is pθ(x given z). The weights are named phi & theta rather than W and V as in Helmholtz—a cosmetic difference. These 2 networks h

  Read more →