Evolutionary algorithms (EA) reproduce essential elements of biological evolution in a computer algorithm in order to solve "difficult" problems, at least approximately, for which no exact or satisfactory solution methods are known. They are metaheuristics and population-based bio-inspired algorithms and evolutionary computation, which itself are part of the field of computational intelligence. The mechanisms of biological evolution that an EA mainly imitates are reproduction, mutation, recombination and selection. Candidate solutions to the optimization problem play the role of individuals in a population, and the fitness function determines the quality of the solutions (see also loss function). Evolution of the population then takes place after the repeated application of the above operators. Evolutionary algorithms often perform well approximating solutions to all types of problems because they ideally do not make any assumption about the underlying fitness landscape. Techniques from evolutionary algorithms applied to the modeling of biological evolution are generally limited to explorations of microevolution (microevolutionary processes) and planning models based upon cellular processes. In most real applications of EAs, computational complexity is a prohibiting factor. In fact, this computational complexity is due to fitness function evaluation. Fitness approximation is one of the solutions to overcome this difficulty. However, seemingly simple EA can solve often complex problems; therefore, there may be no direct link between algorithm complexity and problem complexity. == Generic definition == The following is an example of a generic evolutionary algorithm: Randomly generate the initial population of individuals, the first generation. Evaluate the fitness of each individual in the population. Check, if the goal is reached and the algorithm can be terminated. Select individuals as parents, preferably of higher fitness. Produce offspring with optional crossover (mimicking reproduction). Apply mutation operations on the offspring. Select individuals preferably of lower fitness for replacement with new individuals (mimicking natural selection). Return to 2 == Types == Similar techniques differ in genetic representation and other implementation details, and the nature of the particular applied problem. Genetic algorithm – This is the most popular type of EA. One seeks the solution of a problem in the form of strings of numbers (traditionally binary, although the best representations are usually those that reflect something about the problem being solved), by applying operators such as recombination and mutation (sometimes one, sometimes both). This type of EA is often used in optimization problems. Genetic programming – Here the solutions are in the form of computer programs, and their fitness is determined by their ability to solve a computational problem. There are many variants of Genetic Programming: Cartesian genetic programming Gene expression programming Grammatical evolution Linear genetic programming Multi expression programming Evolutionary programming – Similar to evolution strategy, but with a deterministic selection of all parents. Evolution strategy (ES) – Works with vectors of real numbers as representations of solutions, and typically uses self-adaptive mutation rates. The method is mainly used for numerical optimization, although there are also variants for combinatorial tasks. CMA-ES Natural evolution strategy Differential evolution – Based on vector differences and is therefore primarily suited for numerical optimization problems. Coevolutionary algorithm – Similar to genetic algorithms and evolution strategies, but the created solutions are compared on the basis of their outcomes from interactions with other solutions. Solutions can either compete or cooperate during the search process. Coevolutionary algorithms are often used in scenarios where the fitness landscape is dynamic, complex, or involves competitive interactions. Neuroevolution – Similar to genetic programming but the genomes represent artificial neural networks by describing structure and connection weights. The genome encoding can be direct or indirect. Learning classifier system – Here the solution is a set of classifiers (rules or conditions). A Michigan-LCS evolves at the level of individual classifiers whereas a Pittsburgh-LCS uses populations of classifier-sets. Initially, classifiers were only binary, but now include real, neural net, or S-expression types. Fitness is typically determined with either a strength or accuracy based reinforcement learning or supervised learning approach. Quality–Diversity algorithms – QD algorithms simultaneously aim for high-quality and diverse solutions. Unlike traditional optimization algorithms that solely focus on finding the best solution to a problem, QD algorithms explore a wide variety of solutions across a problem space and keep those that are not just high performing, but also diverse and unique. == Theoretical background == The following theoretical principles apply to all or almost all EAs. === No free lunch theorem === The no free lunch theorem of optimization states that all optimization strategies are equally effective when the set of all optimization problems is considered. Under the same condition, no evolutionary algorithm is fundamentally better than another. This can only be the case if the set of all problems is restricted. This is exactly what is inevitably done in practice. Therefore, to improve an EA, it must exploit problem knowledge in some form (e.g. by choosing a certain mutation strength or a problem-adapted coding). Thus, if two EAs are compared, this constraint is implied. In addition, an EA can use problem specific knowledge by, for example, not randomly generating the entire start population, but creating some individuals through heuristics or other procedures. Another possibility to tailor an EA to a given problem domain is to involve suitable heuristics, local search procedures or other problem-related procedures in the process of generating the offspring. This form of extension of an EA is also known as a memetic algorithm. Both extensions play a major role in practical applications, as they can speed up the search process and make it more robust. === Convergence === For EAs in which, in addition to the offspring, at least the best individual of the parent generation is used to form the subsequent generation (so-called elitist EAs), there is a general proof of convergence under the condition that an optimum exists. Without loss of generality, a maximum search is assumed for the proof: From the property of elitist offspring acceptance and the existence of the optimum it follows that per generation k {\displaystyle k} an improvement of the fitness F {\displaystyle F} of the respective best individual x ′ {\displaystyle x'} will occur with a probability P > 0 {\displaystyle P>0} . Thus: F ( x 1 ′ ) ≤ F ( x 2 ′ ) ≤ F ( x 3 ′ ) ≤ ⋯ ≤ F ( x k ′ ) ≤ ⋯ {\displaystyle F(x'_{1})\leq F(x'_{2})\leq F(x'_{3})\leq \cdots \leq F(x'_{k})\leq \cdots } I.e., the fitness values represent a monotonically non-decreasing sequence, which is bounded due to the existence of the optimum. From this follows the convergence of the sequence against the optimum. Since the proof makes no statement about the speed of convergence, it is of little help in practical applications of EAs. But it does justify the recommendation to use elitist EAs. However, when using the usual panmictic population model, elitist EAs tend to converge prematurely more than non-elitist ones. In a panmictic population model, mate selection (see step 4 of the generic definition) is such that every individual in the entire population is eligible as a mate. In non-panmictic populations, selection is suitably restricted, so that the dispersal speed of better individuals is reduced compared to panmictic ones. Thus, the general risk of premature convergence of elitist EAs can be significantly reduced by suitable population models that restrict mate selection. === Virtual alphabets === With the theory of virtual alphabets, David E. Goldberg showed in 1990 that by using a representation with real numbers, an EA that uses classical recombination operators (e.g. uniform or n-point crossover) cannot reach certain areas of the search space, in contrast to a coding with binary numbers. This results in the recommendation for EAs with real representation to use arithmetic operators for recombination (e.g. arithmetic mean or intermediate recombination). With suitable operators, real-valued representations are more effective than binary ones, contrary to earlier opinion. == Comparison to other concepts == === Biological processes === A possible limitation of many evolutionary algorithms is their lack of a clear genotype–phenotype distinction. In nature, the fertilized egg cell undergoes a complex process known as embryogenesis to become a mature p
Coupled pattern learner
Coupled Pattern Learner (CPL) is a machine learning algorithm which couples the semi-supervised learning of categories and relations to forestall the problem of semantic drift associated with boot-strap learning methods. == Coupled Pattern Learner == Semi-supervised learning approaches using a small number of labeled examples with many unlabeled examples are usually unreliable as they produce an internally consistent, but incorrect set of extractions. CPL solves this problem by simultaneously learning classifiers for many different categories and relations in the presence of an ontology defining constraints that couple the training of these classifiers. It was introduced by Andrew Carlson, Justin Betteridge, Estevam R. Hruschka Jr. and Tom M. Mitchell in 2009. == CPL overview == CPL is an approach to semi-supervised learning that yields more accurate results by coupling the training of many information extractors. Basic idea behind CPL is that semi-supervised training of a single type of extractor such as ‘coach’ is much more difficult than simultaneously training many extractors that cover a variety of inter-related entity and relation types. Using prior knowledge about the relationships between these different entities and relations CPL makes unlabeled data as a useful constraint during training. For e.g., ‘coach(x)’ implies ‘person(x)’ and ‘not sport(x)’. == CPL description == === Coupling of predicates === CPL primarily relies on the notion of coupling the learning of multiple functions so as to constrain the semi-supervised learning problem. CPL constrains the learned function in two ways. Sharing among same-arity predicates according to logical relations Relation argument type-checking === Sharing among same-arity predicates === Each predicate P in the ontology has a list of other same-arity predicates with which P is mutually exclusive. If A is mutually exclusive with predicate B, A’s positive instances and patterns become negative instances and negative patterns for B. For example, if ‘city’, having an instance ‘Boston’ and a pattern ‘mayor of arg1’, is mutually exclusive with ‘scientist’, then ‘Boston’ and ‘mayor of arg1’ will become a negative instance and a negative pattern respectively for ‘scientist.’ Further, Some categories are declared to be a subset of another category. For e.g., ‘athlete’ is a subset of ‘person’. === Relation argument type-checking === This is a type checking information used to couple the learning of relations and categories. For example, the arguments of the ‘ceoOf’ relation are declared to be of the categories ‘person’ and ‘company’. CPL does not promote a pair of noun phrases as an instance of a relation unless the two noun phrases are classified as belonging to the correct argument types. === Algorithm description === Following is a quick summary of the CPL algorithm. Input: An ontology O, and a text corpus C Output: Trusted instances/patterns for each predicate for i=1,2,...,∞ do foreach predicate p in O do EXTRACT candidate instances/contextual patterns using recently promoted patterns/instances; FILTER candidates that violate coupling; RANK candidate instances/patterns; PROMOTE top candidates; end end ==== Inputs ==== A large corpus of Part-Of-Speech tagged sentences and an initial ontology with predefined categories, relations, mutually exclusive relationships between same-arity predicates, subset relationships between some categories, seed instances for all predicates, and seed patterns for the categories. ==== Candidate extraction ==== CPL finds new candidate instances by using newly promoted patterns to extract the noun phrases that co-occur with those patterns in the text corpus. CPL extracts, Category Instances Category Patterns Relation Instances Relation Patterns ==== Candidate filtering ==== Candidate instances and patterns are filtered to maintain high precision, and to avoid extremely specific patterns. An instance is only considered for assessment if it co-occurs with at least two promoted patterns in the text corpus, and if its co-occurrence count with all promoted patterns is at least three times greater than its co-occurrence count with negative patterns. ==== Candidate ranking ==== CPL ranks candidate instances using the number of promoted patterns that they co-occur with so that candidates that occur with more patterns are ranked higher. Patterns are ranked using an estimate of the precision of each pattern. ==== Candidate promotion ==== CPL ranks the candidates according to their assessment scores and promotes at most 100 instances and 5 patterns for each predicate. Instances and patterns are only promoted if they co-occur with at least two promoted patterns or instances, respectively. == Meta-Bootstrap Learner == Meta-Bootstrap Learner (MBL) was also proposed by the authors of CPL. Meta-Bootstrap learner couples the training of multiple extraction techniques with a multi-view constraint, which requires the extractors to agree. It makes addition of coupling constraints on top of existing extraction algorithms, while treating them as black boxes, feasible. MBL assumes that the errors made by different extraction techniques are independent. Following is a quick summary of MBL. Input: An ontology O, a set of extractors ε Output: Trusted instances for each predicate for i=1,2,...,∞ do foreach predicate p in O do foreach extractor e in ε do Extract new candidates for p using e with recently promoted instances; end FILTER candidates that violate mutual-exclusion or type-checking constraints; PROMOTE candidates that were extracted by all extractors; end end Subordinate algorithms used with MBL do not promote any instance on their own, they report the evidence about each candidate to MBL and MBL is responsible for promoting instances. == Applications == In their paper authors have presented results showing the potential of CPL to contribute new facts to existing repository of semantic knowledge, Freebase
Mealy machine
In the theory of computation, a Mealy machine is a finite-state machine whose output values are determined both by its current state and the current inputs. This is in contrast to a Moore machine, whose output values are determined solely by its current state. A Mealy machine is a deterministic finite-state transducer: for each state and input, at most one transition is possible. == History == The Mealy machine is named after George H. Mealy, who presented the concept in a 1955 paper, "A Method for Synthesizing Sequential Circuits". == Formal definition == A Mealy machine is a 6-tuple ( S , S 0 , Σ , Λ , T , G ) {\displaystyle (S,S_{0},\Sigma ,\Lambda ,T,G)} consisting of the following: a finite set of states S {\displaystyle S} a start state (also called initial state) S 0 {\displaystyle S_{0}} which is an element of S {\displaystyle S} a finite set called the input alphabet Σ {\displaystyle \Sigma } a finite set called the output alphabet Λ {\displaystyle \Lambda } a transition function T : S × Σ → S {\displaystyle T:S\times \Sigma \rightarrow S} mapping pairs of a state and an input symbol to the corresponding next state. an output function G : S × Σ → Λ {\displaystyle G:S\times \Sigma \rightarrow \Lambda } mapping pairs of a state and an input symbol to the corresponding output symbol. In some formulations, the transition and output functions are coalesced into a single function T : S × Σ → S × Λ {\displaystyle T:S\times \Sigma \rightarrow S\times \Lambda } . "Evolution across time" is realized in this abstraction by having the state machine consult the time-changing input symbol at discrete "timer ticks" t 0 , t 1 , t 2 , . . . {\displaystyle t_{0},t_{1},t_{2},...} and react according to its internal configuration at those idealized instants, or else having the state machine wait for a next input symbol (as on a FIFO) and react whenever it arrives. == Comparison of Mealy machines and Moore machines == Mealy machines tend to have fewer states: Different outputs on arcs (n2) rather than states (n). When implemented as electronic circuits (rather than as mathematical abstractions or code): Moore machines are safer to use than Mealy machines: Outputs change at the clock edge (always one cycle later). In Mealy machines, input change can cause output change as soon as logic is done — a big problem when two machines are interconnected – asynchronous feedback may occur if one isn't careful. Mealy machines react faster to inputs: React in the same cycle—they don't need to wait for the clock. In Moore machines, more logic may be necessary to decode state into outputs—more gate delays after clock edge. == Diagram == The state diagram for a Mealy machine associates an output value with each transition edge, in contrast to the state diagram for a Moore machine, which associates an output value with each state. When the input and output alphabet are both Σ, one can also associate to a Mealy automata a Helix directed graph (S × Σ, (x, i) → (T(x, i), G(x, i))). This graph has as vertices the couples of state and letters, each node is of out-degree one, and the successor of (x, i) is the next state of the automata and the letter that the automata output when it is instate x and it reads letter i. This graph is a union of disjoint cycles if the automaton is bireversible. == Examples == === Simple === A simple Mealy machine has one input and one output. Each transition edge is labeled with the value of the input (shown in red) and the value of the output (shown in blue). The machine starts in state Si. (In this example, the output is the exclusive-or of the two most-recent input values; thus, the machine implements an edge detector, outputting a 1 every time the input flips and a 0 otherwise.) === Complex === More complex Mealy machines can have multiple inputs as well as multiple outputs. == Applications == Mealy machines provide a rudimentary mathematical model for cipher machines. Considering the input and output alphabet the Latin alphabet, for example, then a Mealy machine can be designed that given a string of letters (a sequence of inputs) can process it into a ciphered string (a sequence of outputs). However, although a Mealy model could be used to describe the Enigma, the state diagram would be too complex to provide feasible means of designing complex ciphering machines. Moore/Mealy machines are DFAs that have also output at any tick of the clock. Modern CPUs, computers, cell phones, digital clocks and basic electronic devices/machines have some kind of finite state machine to control it. Simple software systems, particularly ones that can be represented using regular expressions, can be modeled as finite state machines. There are many such simple systems, such as vending machines or basic electronics. By finding the intersection of two finite state machines, one can design in a very simple manner concurrent systems that exchange messages for instance. For example, a traffic light is a system that consists of multiple subsystems, such as the different traffic lights, that work concurrently.
Phraselator
The Phraselator is a weatherproof handheld language translation device developed by Applied Data Systems and VoxTec, a former division of the military contractor Marine Acoustics, located in Annapolis, Maryland, USA. It was designed to serve as a handheld computer device that translates English into one of 40 different languages. == The device == The Phraselator is a small speech translation PDA-sized device designed to aid in interpretation. The device does not produce synthesized speech like that utilized by Stephen Hawking; instead, it plays pre-recorded foreign language MP3 files. Users can select the phrase they wish to convey from an English list on the screen or speak into the device. It then uses speech recognition technology called DynaSpeak, developed by SRI International, to play the proper sound file. The accuracy of the speech recognition software is over 70 percent according to software developer Jack Buchanan. The device can also record replies for translation later. Pre-recorded phrases are stored on Secure Digital flash memory cards. A 128 MB card can hold up to 12,000 phrases in four or five languages. Users can download phrase modules from the official website, which contained over 300,000 phrases as of March 2005. Users can also construct their own custom phrase modules. Earlier devices were known to have run on an SA-1110 Strong Arm 206 MHz CPU with 32MB SDRAM and 32MB onboard Flash RAM. A newer model, the P2, was released in 2004 and developed according to feedback from U.S. soldiers. It translates one way from English to approximately 60 other languages. It has a directional microphone, a larger library of phrases and a longer battery life. The 2004 release was created by and utilizes a computer board manufactured by InHand Electronics, Inc. In the future, the device will be able to display pictures so users can ask questions such as "Have you seen this person?" Developer Ace Sarich notes that the device is inferior to human interpreter. Conclusions derived from a Nepal field test conducted by U.S. and Nepal based NGO Himalayan Aid in 2004 seemed to confirm Sarich's comparisons: The very concept of using a machine as a communication point between individuals seemed to actually encourage a more limited form of interaction between tester and respondent. Usually, when limited language skills are present between parties, the genuine struggle and desire to communicate acts as a display of good will – we openly display our weakness in this regard – and the result is a more relaxed and human encounter. This was not necessarily present with the Phraselator as all parties abandoned learning about each other and instead focused on learning how to work with the device. As a tool for bridging any cultural differences or communicating effectively at any length, the Phraselator would not be recommended. This device, at least in the form tested, would best be used in large-scale operations where there is no time for language training and there is a need to communicate fixed ideas, quickly, over the greatest distance by employing large amounts of unskilled users. Large humanitarian or natural disasters in remote areas of third-world countries might be an effective example. == Origin == The original idea for the device came from Lee Morin, a Navy doctor in Operation Desert Storm. To communicate with patients, he played Arabic audio files from his laptop. He informed Ace Sarich, the vice president of VoxTec, about the idea. VoxTec won a DARPA Small Business Innovation Research grant in early 2001 to develop a military-grade handheld phrase translator. During its development, the Phraselator was tested and evaluated by scientists from the Army Research Laboratory. The device was first field tested in Afghanistan in 2001. By 2002, about 500 Phraselators were built for soldiers around the world with another 250 ordered by the U.S. Special Forces. The device cost $2000 to develop and could convert spoken English into one of 200,000 recorded commands and questions in 30 languages. However, the device could only translate one-way. At the time, the only existing two-way voice translator that could convert speech back and forth between languages was the Audio Voice Translation Guide System, or TONGUES, which was developed by Carnegie Mellon University for Lockheed Martin. As part of a DARPA program known as the Spoken Language Communication and Translation System for Tactical Use, SRI International has further developed two-way translation software for use in Iraq called IraqComm in 2006 which contains a vocabulary of 40,000 English words and 50,000 words in Iraqi Arabic. == Notable users == The handheld translator was recently used by U.S. troops while providing relief to tsunami victims in early 2005. About 500 prototypes of the device were provided to U.S. military forces in Operation Enduring Freedom. Units loaded with Haitian dialects have been provided to U.S. troops in Haiti. Army military police have used it in Kandahar to communicate with POWs. In late 2004, the U.S. Navy began to augment some ships with a version of the device attached to large speakers in order to broadcast clear voice instructions up to 400 yards (370 m) away. Corrections officers and law enforcement in Oneida County, New York, have tested the device. Hospital emergency rooms and health departments have also evaluated it. Several Native American tribes such as the Choctaw Nation, the Ponca, and the Comanche Nation have also used the device to preserve their dying languages. Various law enforcement agencies, such as the Los Angeles Police Department, also use the phraselator in their patrol cars. == Awards == In March 2004, DARPA director Dr. Tony Tether presented the Small Business Innovative Research Award to the VoxTec division of Marine Acoustics at DARPATech 2004 in Anaheim, CA. The device was recently listed as one of "Ten Emerging Technologies That Will Change Your World" in MIT's Technology Review. == Pop culture == Software developer Jack Buchanan believes that building a device similar to the fictional universal translator seen in Star Trek would be harder than building the Enterprise. The device was mentioned in a list of "Top 10 Star Trek Tech" on Space.com.
Co-occurrence
In linguistics, co-occurrence or cooccurrence (in older texts often shown with diacritic as coöccurrence) is an above-chance frequency of ordered occurrence of two adjacent terms in a text corpus. Co-occurrence in this linguistic sense can be interpreted as an indicator of semantic proximity or an idiomatic expression. Corpus linguistics and its statistical analyses can reveal (regularity of) patterns of co-occurrences within a language and enable the working out of typical collocations for its lexical items. A co-occurrence restriction is identified when linguistic elements never occur together. Analysis of these restrictions can lead to discoveries about the structure and development of a language. Co-occurrence can be seen an extension of word counting in higher dimensions. Co-occurrence can be quantitatively described using measures like a massive correlation or mutual information. Co-occurrence information and knowledge of co-occurring words may be relevant in analysis of language for the purposes of large language models, part of the emerging field of artificial intelligence, and helpful in word games such as scrabble.
Compute (machine learning)
In machine learning and deep learning, compute is the amount of computing power or computational resources required to train machine learning models and large language models. More broadly, compute is the computational power or resources necessary for a computer or computer program to function. == Definition == Compute is commonly defined as the amount of computing power or computational resources required to train machine learning and large language models. The term "compute" has also been more broadly applied to cloud computing, referencing processing power, memory, networking, storage, and other resources required for the computation of any program. Compute is measured in petaflop/s-days and is used to document AI training. A petaflop/s-day (pfs-day) consists of performing 1015 neural net operations per second for one day, or a total of about 1020 operations. The compute-time product serves as a mental convenience, similar to kilowatt-hour for energy. An amount of compute is meant to give an idea of the number of actual operations performed. == History == In a 2018 analysis titled "AI and compute", artificial intelligence company OpenAI introduced the concept of compute. OpenAI identified two eras of training AI systems in terms of compute-usage. From 1959 to 2012, compute roughly followed Moore’s law. Between 2012 and 2018, the amount of compute used in the largest AI training runs increased exponentially, growing by more than 300,000 times — roughly doubling every 3.4 months. By comparison, Moore’s Law doubled every two years over the same period. One of the largest models, released in 2020, used 600,000 times more computing power than the 2012 model. After 2020, compute growth began to slow down, with the compute needed for the largest AI models continuing to slow down in 2023. The notion of compute has become increasingly used from the mid-2020s onwards. == Compute growth and AI progress == Larger AI models trained on more data and using more computational resources, tend to perform better. This happens even if the algorithms themselves remain unchanged. As early as 2018, OpenAI noted the exponential increase in compute to be have a key role in AI progress. OpenAI considers three factors drive the advance of AI: algorithmic innovation, data, and the amount of compute available for training. AI models with more compute not only improve in the tasks they were trained on but can develop emergent abilities. Incremental improvements can lead to more abrupt leaps in capabilities. AI provider SpaceXAI said in 2026 that their AI progress is driven by compute and used it a key metric in the AI training of its supercomputer Colossus, the which contains 1 million GPUs. Anthropic has a contract of $1.25 billion per month with SpaceXAI to buy all the compute capacity at Colossus 1 data center. === Criticism and policy === Increasing, promoting or constraining progress in artificial intelligence has often be done via controlling the amount of compute. Policymarkers have enacted policies and provided support to make compute resources more accessible to domestic AI researchers. In a January 2022 report, the Center for Security and Emerging Technology (CSET) suggested to institutions that increasingly powerful and generalizable AI (AGI) will likely require other strategies than maximizing compute. Some AI researchers are also concerned that government might exclusively focus on scaling compute instead of other strategies. The CSET has reported on the various bottlenecks which could explain why deep learning needs for compute have slow down: training is expensive and training extremely large models generates traffic jams across many processors that are difficult to manage. there is a limited supply of AI chips (see AI chip memory shortage). CSET advances that the main resource is human capital, specifically talented researchers — according to a 2023 published survey of more than 400 AI researchers, academic and private sector workers. The survey found that AI researchers are not primarily or exclusively constrained by compute access. However, both academic and industry AI researchers equally report concerns that insufficient compute could prevent them from contributing meaningfully to AI research in the future. High compute users are more concerned about compute access. When asked about which resource provided by the government would be the most useful to them, some AI researchers select compute, other prefer grant funding. For this goal, CSET advised policymakers to ensure that even researchers with smaller budgets could effectively contribute to AI research. Other proposed strategies include using contemporary AI algorithms, managing modern AI infrastructure or focusing on interdisciplinary work between the AI field and other fields of computer science. A 2024 study on compute access found that academic-only AI research teams often have less compute intensive research topics, especially foundation models, compared to industry AI labs. As a consequence, academia is likely to play a smaller role in advancing such techniques. The researchers suggest nationally-sponsored computing infrastructure as well as open science initiatives to boost academic compute access. === Data === A 2022 study found that current large language models are significantly under-trained, a consequence of focusing on scaling language models whilst keeping the amount of training data constant. By training over 400 language models of various parameter and token size, they found that "for compute-optimal training", the model size and the number of training tokens should ideally be scaled equally: for every doubling of model size the number of training tokens should also be doubled.
Interactive machine translation
Interactive machine translation (IMT), is a specific sub-field of computer-aided translation. Under this translation paradigm, the computer software that assists the human translator attempts to predict the text the user is going to input by taking into account all the information it has available. Whenever such prediction is wrong and the user provides feedback to the system, a new prediction is performed considering the new information available. Such process is repeated until the translation provided matches the user's expectations. Interactive machine translation is specially interesting when translating texts in domains where it is not admissible to output a translation containing errors, hence requiring a human user to amend the translations provided by the system. In such cases, interactive machine translation has been proved to provide benefit to potential users. Nevertheless, there are few commercial software that implements interactive machine translation and work done in the field is mostly restrained to academic research. == History == Historically, interactive machine translation is born as an evolution of the computer-aided translation paradigm, where the human translator and the machine translation system were intended to work as a tandem. This first work was extended within the TransType research project, funded by the Canadian government. In this project, the human interaction was aimed towards producing the target text for the first time by embedding data-driven machine translation techniques within the interactive translation environment with the goal of achieving the best of both actors: the efficiency of the automatic system and the reliability of human translators. Later, a larger-scale research project, TransType2, funded by the European Commission extended such work by analyzing the incorporation of a complete machine translation system into the process, with the goal of producing a complete translation hypothesis, which the human user is allowed to amend or accept. If the user decides to amend the hypothesis, the system then attempts to make the best use of such feedback in order to produce a new translation hypothesis that takes into account the modifications introduced by the user. More recently, CASMACAT, also funded by the European Commission, aimed at developing novel types of assistance to human translators and integrated them into a new workbench, consisting of an editor, a server, and analysis and visualisation tools. The workbench was designed in a modular fashion and can be combined with existing computer aided translation tools. Furthermore, the CASMACAT workbench can learn from the interaction with the human translator by updating and adapting its models instantly based on the translation choices of the user. Recent work on involving an extensive evaluation with human users revealed the fact that interactive machine translation may even be used by users that do not speak the source language in order to achieve near professional translation quality. Moreover, it also elucidated the fact that an interactive scenario is more beneficial than a classic post-edition scenario. The previously described approaches rely on a tightly coupled underlying corpus-based machine translation system (usually, a Statistical machine translation system) that is used as a glass box, therefore inheriting the shortcomings of the translation systems and limiting the usage of interactive machine translation for some scenarios. For this reason, an approach that uses any kind of bilingual resource (not limited to machine translation) as a black-box to provide interactive machine translation was developed. This approach is not able to extract as much information from the bilingual resources used, due to the black-box nature of the interaction, but can use any resource available to the user. Forecat is a black-box interactive machine translation implementation that is available both as a web application (that includes a webpage and a web services interface) and as a plugin for OmegaT (Forecat-OmegaT). == Process == The interactive machine translation process starts with the system suggesting a translation hypothesis to the user. Then, the user may accept the complete sentence as correct, or may modify it if he considers there is some error. Typically, when modifying a given word, it is assumed that the prefix until that word is correct, leading to a left-to-right interaction scheme. Once the user has changed the word considered incorrect, the system then proposes a new suffix, i.e. the remainder of the sentence. Such process continues until the translation provided satisfies the user. Although explained at the word level, the previous process may also be implemented at the character level, and hence the system provides a suffix whenever the human translator types in a single character. In addition, there is ongoing effort towards changing the typical left-to-right interaction scheme in order to make human-machine interaction easier. A similar approach is used in the Caitra translation tool. == Evaluation == Evaluation is a difficult issue in interactive machine translation. Ideally, evaluation should take place in experiments involving human users. However, given the high monetary cost this would imply, this is seldom the case. Moreover, even when considering human translators in order to perform a true evaluation of interactive machine translation techniques, it is not clear what should be measured in such experiments, since there are many different variables that should be taken into account and cannot be controlled, as is for instance the time the user takes in order to get used to the process. In the CASMACAT project, some field trials have been carried out to study some of these variables. For quick evaluations in laboratory conditions, interactive machine translation is measured by using the key stroke ratio or the word stroke ratio. Such criteria attempt to measure how many key-strokes or words did the user need to introduce before producing the final translated document. == Differences with classical computer-aided translation == Although interactive machine translation is a sub-field of computer-aided translation, the main attractive of the former with respect to the latter is the interactivity. In classical computer-aided translation, the translation system may suggest one translation hypothesis in the best case, and then the user is required to post-edit such hypothesis. In contrast, in interactive machine translation the system produces a new translation hypothesis each time the user interacts with the system, i.e. after each word (or letter) has been introduced.