The Cloud Native Computing Foundation (CNCF) is a subsidiary of the Linux Foundation founded in 2015 to support cloud-native computing. == History == It was announced alongside Kubernetes 1.0, an open source container cluster manager, which was contributed to the Linux Foundation by Google as a seed technology. Founding members include Google, CoreOS, Mesosphere, Red Hat, Twitter, Huawei, Intel, RX-M, Cisco, IBM, Docker, Univa, and VMware. Today, CNCF is supported by over 450 members. In August 2018 Google announced that it was handing over operational control of Kubernetes to the community. == Projects == Argo is a collection of tools for getting work done with Kubernetes. Among its main features are Workflows and Events. It was accepted to CNCF on March 26, 2020 at the Incubating maturity level and then moved to the Graduated maturity level on December 6, 2022. cert-manager provisions and manages TLS certificates in Kubernetes. It was accepted to CNCF on November 10, 2020, moved to the Incubating maturity level on September 19, 2022, and then moved to the Graduated maturity level on September 29, 2024. Cilium provides networking, security, and observability for Kubernetes deployments using eBPF technology. It joined the CNCF at incubation level in October 2021 and the CNCF announced its graduation in October 2023. containerd is an industry-standard core container runtime. It is currently available as a daemon for Linux and Windows, which can manage the complete container lifecycle of its host system. In 2015, Docker donated the OCI Specification to The Linux Foundation with a reference implementation called runc. Since February 28, 2019 it is an official CNCF project. Its general availability and intention to donate the project to CNCF was announced by Docker in 2017. CoreDNS is a DNS server that chains plugins. Its graduation was announced in 2019. Dapr, the distributed application runtime, provides APIs for building secure and reliable microservices and agentic AI systems. Dapr was donated to the CNCF in November 2021 and joined at incubation level. The CNCF announced its graduation in November 2024. Envoy: Originally built at Lyft to move their architecture away from a monolith, Envoy is a high-performance open source edge and service proxy that makes the network transparent to applications. Lyft contributed Envoy to Cloud Native Computing Foundation in September 2017. etcd is a distributed key value store, providing a method of storing data across a cluster of machines. It became a CNCF incubating project in 2018 at KubeCon+CloudNativeCon North America in Seattle that year. Falco is an open source and cloud native runtime security initiative. It is the "de facto Kubernetes threat detection engine". It became an incubating project in January 2020 and graduated in February 2024. Flux is an open source project for powering GitOps in Kubernetes clusters. It provides the GitOps Toolkit, a set of Kubernetes APIs that allow you to define how configuration source code is securely pulled into your cluster and deployed by popular Kubernetes manifests rendering engines like Kustomize and Helm. The most recommended source mechanism is the OCIRepository API, which provides enhanced security and benefits from container image tooling out there. Flux has also notification integrations with popular services like Prometheus Alertmanager, PagerDuty, Slack and so on. Flux has graduated in CNCF in 2022. Harbor is an "open source trusted cloud native registry project that stores, signs, and scans content." It became an incubating project in September 2019 and graduated in June 2020. Helm is a package manager that helps developers "easily manage and deploy applications onto the Kubernetes cluster." It joined the incubating level in June 2018 and graduated in April 2020. Istio is a service mesh technology. It was accepted by CNCF in September 2022 and graduated on July 12, 2023. Jaeger, Created by Uber Engineering, Jaeger is an open source distributed tracing system inspired by Google Dapper paper and OpenZipkin community. It can be used for tracing microservice-based architectures, including distributed context propagation, distributed transaction monitoring, root cause analysis, service dependency analysis, and performance/latency optimization. The Cloud Native Computing Foundation Technical Oversight Committee voted to accept Jaeger as the 12th hosted project in September 2017 and became a graduated project in 2019. In 2020 it became an approved and fully integrated part of the CNCF ecosystem. Kubernetes is an open source framework for automating deployment and managing applications in a containerized and clustered environment. "It aims to provide better ways of managing related, distributed components across the varied infrastructure." It was originally designed by Google and donated to The Linux Foundation to form the Cloud Native Computing Foundation with Kubernetes as the seed technology. The "large and diverse" community supporting the project has made its staying power more robust than other, older technologies of the same ilk. In January 2020, the CNCF annual report showed significant growth in interest, training, event attendance and investment related to Kubernetes. Linkerd is CNCF's fifth member project, and the project that coined the term "service mesh". Linkerd adds observability, security, and reliability features to applications by adding them to the platform rather than the application layer, and features a "micro-proxy" to maximize speed and security of its data plane. Linkerd graduated from CNCF in July 2021. Open Policy Agent (OPA) is "an open source general-purpose policy engine and language for cloud infrastructure." It became a CNCF incubating project in April 2019. OPA graduated from CNCF in February 2021. Prometheus is a cloud monitoring tool sponsored by SoundCloud in early iterations. In August 2018, the tool was designated a graduated project by the Cloud Native Computing Foundation. It is now a Cloud Native Computing Foundation member project. Rook is CNCF's first cloud native storage project. It became an incubation level project in 2018 and graduated in October 2020. SPIFFE is an open standard and framework for workload identity, much the same way that OAuth is an open standard and framework for human identity. It is built from the ground up to accommodate modern computing environments, which operate with systems scale and velocity (as opposed to human scale and velocity), while still maintaining interoperability with existing technologies like OAuth and X.509 Public key infrastructure. Unlike other identity standards, SPIFFE supports multiple credential types for a single identity, ensuring that the highly varied needs of production environments are consistently met without compromise. SPIFFE joined the CNCF as a sandbox project in 2018, was accepted to incubation in 2020, and graduated in 2022. SPIRE is an open source identity provider for workloads based on the SPIFFE framework. It is highly pluggable, and fills the attestation and issuance needs required by any workload identity solution. The plugin interfaces it exposes allows users to write integrations with in-house systems, build internal self-service portals, and more. It is a very powerful building block for issuing short-lived identity credentials to dynamic cloud workloads. SPIRE became a CNCF Graduated project in 2022. The Update Framework (TUF) helps developers to secure new or existing software update systems, which are often found to be vulnerable to many known attacks. TUF addresses this widespread problem by providing a comprehensive, flexible security framework that developers can integrate with any software update system. TUF was CNCF's first security-focused project and the ninth project overall to graduate from the foundation's hosting program. TiKV provides a distributed key–value database. Vitess is a database clustering system for horizontal scaling of MySQL, first created for internal use by YouTube. It became a CNCF project in 2018 and graduated in November 2019. Contour is a management server for Envoy that can direct the management of Kubernetes' traffic. Contour also provides routing features that are more advanced than Kubernetes' out-of-the-box Ingress specification. VMWare contributed the project to CNCF in July 2020. Cortex offers horizontally scalable, multi-tenant, long-term storage for Prometheus and works alongside Amazon DynamoDB, Google Bigtable, Cassandra, S3, GCS, and Microsoft Azure. It was introduced into the ecosystem incubator alongside Thanos in August 2020. CRI-O is an Open Container Initiative (OCI) based "implementation of Kubernetes Container Runtime Interface". CRI-O allows Kubernetes to be container runtime-agnostic. It became an incubating project in 2019. gRPC is a "modern open source high performance RPC framework that can run in any environment." The project was formed in 2015 when Google decided to open sou
CinePlayer
CinePlayer is a software based media player used to review Digital Cinema Packages (DCP) without the need for a digital cinema server by Doremi Labs. CinePlayer can play back any DCP, not just those created by Doremi Mastering products. In addition to playing DCPs, CinePlayer can also playback JPEG2000 image sequences and many popular multimedia file types. There are two versions of CinePlayer available, standard and Pro. The standard version supports playback of non-encrypted, 2D DCP's up to 2K resolution. The Pro version supports playback of encrypted, 2D or 3D DCP's with subtitles up to 4K resolution. == Supported formats == === Containers === AVI MOV MXF MPG TS WMV M2TS MTS MP4 MKV === Video codecs === JPEG2000 ProRes 422 DNxHD YUV Uncompressed 8-10 bits DIVX XVID MPEG4 AVC / H-264 VC-1 MPEG2 === Supported image sequences === BMP TIFF TGA DPX JPG J2C === Supported audio files === WAV MP3 WMA MP2
Simulation decomposition
SimDec, or Simulation decomposition, is a hybrid uncertainty and sensitivity analysis method, for visually examining the relationships between the output and input variables of a computational model. SimDec maps multivariable scenarios onto the distribution of the model output. This visual analytics approach exposes the underlying nature of the model behavior, including its nonlinear and multivariate interaction effects. SimDec can be used in any range of science, engineering, and social domains. Existing applications include business and environmental issues. == Method == SimDec operates on Monte Carlo simulation (or measured) data where both output and input values are recorded. At least one thousand observations (or simulated iterations) are typically recommended to preserve the readability of the resulting histograms. An outline of the decomposition algorithm, which is readily available in multiple programming languages, proceeds as follows: Select the input variables for decomposition. One can use sensitivity indices (see variance-based sensitivity analysis) to define the most influential variables for decomposition or choose them manually according to the decision-problem context (for example, only those input variables that the decision-maker can act upon). Two to three input variables, ordered by decreasing value of their sensitivity indices, usually provide the most meaningful decomposition results. Divide the inputs into states. The numeric ranges of the inputs are split into several intervals with an equal number of observations in each. For categorical variables, the categories represent states. Form scenarios. All combinations of states of the selected input variables produce unique scenarios or subsets of the data. For example, if the range of X2 is divided into low, medium and high, and X3 takes values of 1 or 2, six scenarios are formed: (i) X2 low & X3 = 1, (ii) X2 low & X3 = 2, (iii) X2 medium & X3 = 1, (iv) X2 medium & X3 = 2, (v) X2 high & X3 = 1, and (vi) X2 high & X3 = 2. Assign scenarios to each output value. The simulation data is used to define the scenario index for each simulation run. For example, if an X2 value falls into the low state and X3 is equal to 2, the corresponding scenario, defined in Step 3, is (ii). Color-code the output distribution. When all output values are assigned scenario indices, they are plotted as series in a stacked histogram, visually separated by color-coding. For ease of visual perception, the states of the most influential input variable are assigned distinct colors, and all the remaining partitions take shades of those colors (see Figure). All of these steps can be run automatically on the given data using the open-source SimDec packages currently available in Python, R, Julia, and Matlab. A SimDec template in Excel runs a Monte Carlo simulation of a spreadsheet model but possesses only a manual option for input selection. == How to read SimDec == === Histogram === Histogram is an approximate representation of the distribution of numerical data. Its horizontal axis shows the range of the variable of interest, and its vertical axis denotes count, also called frequency, or, if divided by the total number of data points, probability. The distribution alone can supply only limited information about the data – its minimum, maximum, and shape (where the most of data occurs). === Judging the importance of inputs === If an input variable has no effect on the output, its states (e.g., low & high) would lie on top of each other on the SimDec histogram, occupying fully overlapping ranges of the output. If an input variable has a strong effect and explains most of the variance of the output, the border between its states on the SimDec histogram would be vertical. Such visualization has an important decision-making implication – e.g., if the high state of X can be achieved, it would guarantee a certain range of Y. All cases in-between with low-to-strong effects would show a diagonal border between the states. The less they overlap, the larger the effect of X on Y. While the horizontal displacement of sub-distributions on the SimDec histogram is the key to interpreting the results, the vertical disposition of sub-distributions is just a technical matter of the order of plotting the series of the stacked histogram. === Exploring the interaction of inputs === When two or more input variables are used for decomposition, it becomes possible to examine their joint effects. A schematic visualization portrays how different types of joint effects of input variables on the output appear on SimDec visualization. Understanding the nature of interaction effects in a computational model and its behavior in general is crucial for effective decision-making. == Limitations == The SimDec method has several limitations: It is based on Monte Carlo simulation and thus requires running a computational model a thousand of times or more. To models that take hours to evaluate once, it would be impossible to use SimDec (unless a supercomputer and/or large of time are available). SimDec is based on a histogram, thus, for binary or categorical output variables, the visualization would be very limited (e.g., only a few bins). The more input variables one selects for the decomposition, the less readable the histogram becomes. Only cases with two and three input variables are presented in.
National Security Memorandum on Artificial Intelligence
The Memorandum on Advancing the United States' Leadership in Artificial Intelligence; Harnessing Artificial Intelligence to Fulfill National Security Objectives; and Fostering the Safety, Security, and Trustworthiness of Artificial Intelligence is a memorandum signed by U.S. president Joe Biden. The memorandum is described as seeking to advance U.S. leadership in the development of safe, secure, and trustworthy artificial intelligence (AI); enable the U.S. government to use AI for national security; and contribute to international AI governance.
Rule-based system
In computer science, a rule-based system is a computer system in which domain-specific knowledge is represented in the form of rules and general-purpose reasoning is used to solve problems in the domain. Two different kinds of rule-based systems emerged within the field of artificial intelligence in the 1970s: Production systems, which use if-then rules to derive actions from conditions. Logic programming systems, which use conclusion if conditions rules to derive conclusions from conditions. The differences and relationships between these two kinds of rule-based system has been a major source of misunderstanding and confusion. Both kinds of rule-based systems use either forward or backward chaining, in contrast with imperative programs, which execute commands listed sequentially. However, logic programming systems have a logical interpretation, whereas production systems do not. == Production system rules == A classic example of a production rule-based system is the domain-specific expert system that uses rules to make deductions or choices. For example, an expert system might help a doctor choose the correct diagnosis based on a cluster of symptoms, or select tactical moves to play a game. Rule-based systems can be used to perform lexical analysis to compile or interpret computer programs, or in natural language processing. Rule-based programming attempts to derive execution instructions from a starting set of data and rules. This is a more indirect method than that employed by an imperative programming language, which lists execution steps sequentially. === Construction === A typical rule-based system has four basic components: A list of rules or rule base, which is a specific type of knowledge base. An inference engine or semantic reasoner, which infers information or takes action based on the interaction of input and the rule base. The interpreter executes a production system program by performing the following match-resolve-act cycle: Match: In this first phase, the condition sides of all productions are matched against the contents of working memory. As a result a set (the conflict set) is obtained, which consists of instantiations of all satisfied productions. An instantiation of a production is an ordered list of working memory elements that satisfies the condition side of the production. Conflict-resolution: In this second phase, one of the production instantiations in the conflict set is chosen for execution. If no productions are satisfied, the interpreter halts. Act: In this third phase, the actions of the production selected in the conflict-resolution phase are executed. These actions may change the contents of working memory. At the end of this phase, execution returns to the first phase. Temporary working memory, which is a database of facts. A user interface or other connection to the outside world through which input and output signals are received and sent. Whereas the matching phase of the inference engine has a logical interpretation, the conflict resolution and action phases do not. Instead, "their semantics is usually described as a series of applications of various state-changing operators, which often gets quite involved (depending on the choices made in deciding which ECA rules fire, when, and so forth), and they can hardly be regarded as declarative". == Logic programming rules == The logic programming family of computer systems includes the programming language Prolog, the database language Datalog and the knowledge representation and problem-solving language Answer Set Programming (ASP). In all of these languages, rules are written in the form of clauses: A :- B1, ..., Bn. and are read as declarative sentences in logical form: A if B1 and ... and Bn. In the simplest case of Horn clauses (or "definite" clauses), which are a subset of first-order logic, all of the A, B1, ..., Bn are atomic formulae. Although Horn clause logic programs are Turing complete, for many practical applications, it is useful to extend Horn clause programs by allowing negative conditions, implemented by negation as failure. Such extended logic programs have the knowledge representation capabilities of a non-monotonic logic. == Differences and relationships between production rules and logic programming rules == The most obvious difference between the two kinds of systems is that production rules are typically written in the forward direction, if A then B, and logic programming rules are typically written in the backward direction, B if A. In the case of logic programming rules, this difference is superficial and purely syntactic. It does not affect the semantics of the rules. Nor does it affect whether the rules are used to reason backwards, Prolog style, to reduce the goal B to the subgoals A, or whether they are used, Datalog style, to derive B from A. In the case of production rules, the forward direction of the syntax reflects the stimulus-response character of most production rules, with the stimulus A coming before the response B. Moreover, even in cases when the response is simply to draw a conclusion B from an assumption A, as in modus ponens, the match-resolve-act cycle is restricted to reasoning forwards from A to B. Reasoning backwards in a production system would require the use of an entirely different kind of inference engine. In his Introduction to Cognitive Science, Paul Thagard includes logic and rules as alternative approaches to modelling human thinking. He does not consider logic programs in general, but he considers Prolog to be, not a rule-based system, but "a programming language that uses logic representations and deductive techniques" (page 40). He argues that rules, which have the form IF condition THEN action, are "very similar" to logical conditionals, but they are simpler and have greater psychological plausibility (page 51). Among other differences between logic and rules, he argues that logic uses deduction, but rules use search (page 45) and can be used to reason either forward or backward (page 47). Sentences in logic "have to be interpreted as universally true", but rules can be defaults, which admit exceptions (page 44). He does not observe that all of these features of rules apply to logic programming systems.
15.ai
15.ai was a free non-commercial web application and research project that uses artificial intelligence to generate text-to-speech voices of fictional characters from popular media. Created by a pseudonymous artificial intelligence researcher known as 15, who began developing the technology as a freshman during their undergraduate research at the Massachusetts Institute of Technology (MIT), the application allows users to make characters from video games, television shows, and movies speak custom text with emotional inflections. The platform is able to generate convincing voice output using minimal training data; the name "15.ai" references the creator's statement that a voice can be cloned with just 15 seconds of audio. It was an early example of an application of generative artificial intelligence during the initial stages of the AI boom. Launched in March 2020, 15.ai became an Internet phenomenon in early 2021 when content utilizing it went viral on social media and quickly gained widespread use among Internet fandoms, such as the My Little Pony: Friendship Is Magic, Team Fortress 2, and SpongeBob SquarePants fandoms. The service featured emotional context through emojis, precise pronunciation control, and multi-speaker capabilities. Critics praised 15.ai's accessibility and emotional control but criticized its technical limitations in prosody options and non-English language support, with mixed results depending on character complexity. 15.ai is credited as the first platform to popularize AI voice cloning in memes and content creation. Voice actors and industry professionals debated 15.ai's implications, raising concerns about employment impacts, voice-related fraud, and potential misuse. In January 2022, it was discovered that a company called Voiceverse had generated voice lines using 15.ai without attribution, promoted them as the byproduct of their own technology, and sold them as non-fungible tokens (NFT) without permission. News publications universally characterized this incident as the company having "stolen" from 15.ai. The service went offline in September 2022 due to legal issues surrounding artificial intelligence and copyright. Its shutdown was followed by the emergence of commercial alternatives whose founders have acknowledged 15.ai's pioneering influence in the field of deep learning speech synthesis. On May 18, 2025, 15 launched 15.dev as the sequel to 15.ai. == History == === Background === The field of speech synthesis underwent a significant transformation with the introduction of deep learning approaches. In 2016, DeepMind's publication of the WaveNet paper marked a shift toward neural network-based speech synthesis, which enabled higher audio quality via causal convolutional neural networks. Previously, concatenative synthesis—which worked by stitching together pre-recorded segments of human speech—was the predominant method for generating artificial speech, but it often produced robotic-sounding results at the boundaries of sentences. In 2018, Google AI's Tacotron 2 showed that neural networks could produce highly natural speech synthesis but required substantial training data (typically tens of hours of audio) to achieve acceptable quality. When trained on two hours of training data, the output quality degraded while still being able to maintain intelligible speech; with 24 minutes of training data, Tacotron 2 failed to produce intelligible speech. The same year saw the emergence of HiFi-GAN, a generative adversarial network (GAN)-based vocoder that improved the efficiency of waveform generation while producing high-fidelity speech, followed by Glow-TTS, which introduced a flow-based approach that allowed for both fast inference and voice style transfer capabilities. Chinese tech companies like Baidu and ByteDance also made contributions to the field by developing breakthroughs that further advanced the technology. === 2016–2020: Conception and development === 15.ai was conceived in 2016 as a research project in deep learning speech synthesis by a developer known as 15 (at the age of 18) during their freshman year at MIT as part of its Undergraduate Research Opportunities Program. 15 was inspired by DeepMind's WaveNet paper, with development continuing through their studies as Google AI released Tacotron 2 the following year. By 2019, they had demonstrated at MIT their ability to replicate WaveNet and Tacotron 2's results using 75% less training data than previously required. The name "15.ai" is a reference to the developer's statement that a voice can be cloned with as little as 15 seconds of data. 15 had originally planned to pursue a PhD based on their undergraduate research, but opted to work in the tech industry instead after their startup was accepted into the Y Combinator accelerator in 2019. After their departure in early 2020, 15 returned to their voice synthesis research and began implementing it as a web application. According to a post on X from 15, instead of using conventional voice datasets like LJSpeech that contained simple, monotone recordings, they sought out more challenging voice samples that could demonstrate the model's ability to handle complex speech patterns and emotional undertones. During this phase, 15 discovered the Pony Preservation Project, a collaborative project started by /mlp/, the My Little Pony board on 4chan. Contributors of the project had manually trimmed, denoised, transcribed, and emotion-tagged thousands of voice lines from My Little Pony: Friendship Is Magic and had compiled them into a dataset that provided ideal training material for 15.ai. === 2020–2022: Release and operation === 15.ai was released on March 2, 2020 as a free and non-commercial web application that did not require user registration to use, but did require the user to accept its terms of service before proceeding. At the time of its launch, the platform had a limited selection of available characters, including those from My Little Pony: Friendship Is Magic and Team Fortress 2. Users were permitted to create any content with the synthesized voices under two conditions: they had to properly credit 15.ai by including "15.ai" in any posts, videos, or projects using the generated audio; and they were prohibited from mixing 15.ai outputs with other text-to-speech outputs in the same work to prevent misrepresentation of the technology's capabilities. On March 8, 2020, Tyler McVicker of Valve News Network uploaded a YouTube video showcasing 15.ai. More voices were added to the website in the following months. In late 2020, 15 implemented a multi-speaker embedding in the deep neural network, which enabled the simultaneous training of multiple voices. Following this, the website's roster expanded from eight to over fifty characters. In addition, this implementation allowed the deep learning model to recognize common emotional patterns across different characters, even when certain emotions were missing from the characters' training data. By May 2020, the site had served over 4.2 million audio files to users. In early 2021, the application gained popularity after skits, memes, and fan content created using 15.ai went viral on Twitter, TikTok, Reddit, Twitch, Facebook, and YouTube. At its peak, the platform incurred operational costs of US$12,000 per month from AWS infrastructure needed to handle millions of daily voice generations; despite receiving offers from companies to acquire 15.ai and its underlying technology, the website remained independent and was funded out of the personal previous startup earnings of the developer. === 2022: Voiceverse NFT controversy === On January 14, 2022, 15 discovered that a blockchain-based company called Voiceverse had generated voice lines using 15.ai, falsely showcased them on Twitter as a demonstration of their own voice technology without permission or attribution, and sold them as NFTs. This came shortly after 15 had stated in December 2021 that they had no interest in incorporating NFTs into their work. A screenshot of the log files posted by 15 showed that Voiceverse had generated audio of characters from My Little Pony: Friendship Is Magic using 15.ai and pitched them up to make them sound unrecognizable, a violation of 15.ai's terms of service, which explicitly prohibited commercial use and required proper attribution. When confronted with evidence, Voiceverse stated that their marketing team had used 15.ai without proper attribution while rushing to create a demo. In response, 15 tweeted "Go fuck yourself," which went viral, amassing hundreds of thousands of retweets and likes on Twitter in support of the developer. The tweets showcasing the stolen voices were subsequently deleted. ==== Aftermath ==== The controversy raised concerns about NFT projects, which, according to critics, were frequently associated with intellectual property theft and questionable business practices. The incident was documented in the AI Incident Database (AIID) and the AI, Alg
Catastrophic interference
Catastrophic interference, also known as catastrophic forgetting, is the tendency of an artificial neural network to abruptly and drastically forget previously learned information upon learning new information. Neural networks are an important part of the connectionist approach to cognitive science. The issue of catastrophic interference when modeling human memory with connectionist models was originally brought to the attention of the scientific community by research from McCloskey and Cohen (1989), and Ratcliff (1990). It is a radical manifestation of the 'sensitivity-stability' dilemma or the 'stability-plasticity' dilemma. Specifically, these problems refer to the challenge of making an artificial neural network that is sensitive to, but not disrupted by, new information. Lookup tables and connectionist networks lie on the opposite sides of the stability plasticity spectrum. The former remains completely stable in the presence of new information but lacks the ability to generalize, i.e. infer general principles, from new inputs. On the other hand, connectionist networks like the standard backpropagation network can generalize to unseen inputs, but they are sensitive to new information. Backpropagation models can be analogized to human memory insofar as they have a similar ability to generalize, but these networks often exhibit less stability than human memory. Notably, these backpropagation networks are susceptible to catastrophic interference. This is an issue when modelling human memory, because unlike these networks, humans typically do not show catastrophic forgetting. == Discovery == The term catastrophic interference was originally coined by McCloskey and Cohen (1989) but was also brought to the attention of the scientific community by research from Ratcliff (1990). === The Sequential Learning Problem: McCloskey and Cohen (1989) === McCloskey and Cohen (1989) noted the problem of catastrophic interference during two different experiments with backpropagation neural network modelling. Experiment 1: Learning the ones and twos addition facts In their first experiment they trained a standard backpropagation neural network on a single training set consisting of 17 single-digit ones problems (i.e., 1 + 1 through 9 + 1, and 1 + 2 through 1 + 9) until the network could represent and respond properly to all of them. The error between the actual output and the desired output steadily declined across training sessions, which reflected that the network learned to represent the target outputs better across trials. Next, they trained the network on a single training set consisting of 17 single-digit twos problems (i.e., 2 + 1 through 2 + 9, and 1 + 2 through 9 + 2) until the network could represent, respond properly to all of them. They noted that their procedure was similar to how a child would learn their addition facts. Following each learning trial on the twos facts, the network was tested for its knowledge on both the ones and twos addition facts. Like the ones facts, the twos facts were readily learned by the network. However, McCloskey and Cohen noted the network was no longer able to properly answer the ones addition problems even after one learning trial of the twos addition problems. The output pattern produced in response to the ones facts often resembled an output pattern for an incorrect number more closely than the output pattern for a correct number. This is considered to be a drastic amount of error. Furthermore, the problems 2+1 and 1+2, which were included in both training sets, even showed dramatic disruption during the first learning trials of the twos facts. Experiment 2: Replication of Barnes and Underwood (1959) study In their second connectionist model, McCloskey and Cohen attempted to replicate the study on retroactive interference in humans by Barnes and Underwood (1959). They trained the model on A-B and A-C lists and used a context pattern in the input vector (input pattern), to differentiate between the lists. Specifically the network was trained to respond with the right B response when shown the A stimulus and A-B context pattern and to respond with the correct C response when shown the A stimulus and the A-C context pattern. When the model was trained concurrently on the A-B and A-C items then the network readily learned all of the associations correctly. In sequential training the A-B list was trained first, followed by the A-C list. After each presentation of the A-C list, performance was measured for both the A-B and A-C lists. They found that the amount of training on the A-C list in Barnes and Underwood study that lead to 50% correct responses, lead to nearly 0% correct responses by the backpropagation network. Furthermore, they found that the network tended to show responses that looked like the C response pattern when the network was prompted to give the B response pattern. This indicated that the A-C list apparently had overwritten the A-B list. This could be likened to learning the word dog, followed by learning the word stool and then finding that you think of the word stool when presented with the word dog. McCloskey and Cohen tried to reduce interference through a number of manipulations including changing the number of hidden units, changing the value of the learning rate parameter, overtraining on the A-B list, freezing certain connection weights, changing target values 0 and 1 instead 0.1 and 0.9. However, none of these manipulations satisfactorily reduced the catastrophic interference exhibited by the networks. Overall, McCloskey and Cohen (1989) concluded that: at least some interference will occur whenever new learning alters the weights involved in representing old learning the greater the amount of new learning, the greater the disruption in old knowledge interference was catastrophic in the backpropagation networks when learning was sequential but not concurrent === Constraints Imposed by Learning and Forgetting Functions: Ratcliff (1990) === Ratcliff (1990) used multiple sets of backpropagation models applied to standard recognition memory procedures, in which the items were sequentially learned. After inspecting the recognition performance models he found two major problems: Well-learned information was catastrophically forgotten as new information was learned in both small and large backpropagation networks. Even one learning trial with new information resulted in a significant loss of the old information, paralleling the findings of McCloskey and Cohen (1989). Ratcliff also found that the resulting outputs were often a blend of the previous input and the new input. In larger networks, items learned in groups (e.g. AB then CD) were more resistant to forgetting than were items learned singly (e.g. A then B then C...). However, the forgetting for items learned in groups was still large. Adding new hidden units to the network did not reduce interference. Discrimination between the studied items and previously unseen items decreased as the network learned more. This finding contradicts studies on human memory, which indicated that discrimination increases with learning. Ratcliff attempted to alleviate this problem by adding 'response nodes' that would selectively respond to old and new inputs. However, this method did not work as these response nodes would become active for all inputs. A model which used a context pattern also failed to increase discrimination between new and old items. == Proposed solutions == The main cause of catastrophic interference seems to be overlap in the representations at the hidden layer of distributed neural networks. In a distributed representation, each input tends to create changes in the weights of many of the nodes. Catastrophic forgetting occurs because when many of the weights where "knowledge is stored" are changed, it is unlikely for prior knowledge to be kept intact. During sequential learning, the inputs become mixed, with the new inputs being superimposed on top of the old ones. Another way to conceptualize this is by visualizing learning as a movement through a weight space. This weight space can be likened to a spatial representation of all of the possible combinations of weights that the network could possess. When a network first learns to represent a set of patterns, it finds a point in the weight space that allows it to recognize all of those patterns. However, when the network then learns a new set of patterns, it will move to a place in the weight space for which the only concern is the recognition of the new patterns. To recognize both sets of patterns, the network must find a place in the weight space suitable for recognizing both the new and the old patterns. Below are a number of techniques which have empirical support in successfully reducing catastrophic interference in backpropagation neural networks: === Orthogonality === Many of the early techniques in reducing representational overlap involved making either the input vecto