In computer science, a suffix automaton is an efficient data structure for representing the substring index of a given string which allows the storage, processing, and retrieval of compressed information about all its substrings. The suffix automaton of a string S {\displaystyle S} is the smallest directed acyclic graph with a dedicated initial vertex and a set of "final" vertices, such that paths from the initial vertex to final vertices represent the suffixes of the string. In terms of automata theory, a suffix automaton is the minimal partial deterministic finite automaton that recognizes the set of suffixes of a given string S = s 1 s 2 … s n {\displaystyle S=s_{1}s_{2}\dots s_{n}} . The state graph of a suffix automaton is called a directed acyclic word graph (DAWG), a term that is also sometimes used for any deterministic acyclic finite state automaton. Suffix automata were introduced in 1983 by a group of scientists from the University of Denver and the University of Colorado Boulder. They suggested a linear time online algorithm for its construction and showed that the suffix automaton of a string S {\displaystyle S} having length at least two characters has at most 2 | S | − 1 {\textstyle 2|S|-1} states and at most 3 | S | − 4 {\textstyle 3|S|-4} transitions. Further works have shown a close connection between suffix automata and suffix trees, and have outlined several generalizations of suffix automata, such as compacted suffix automaton obtained by compression of nodes with a single outgoing arc. Suffix automata provide efficient solutions to problems such as substring search and computation of the largest common substring of two and more strings. == History == The concept of suffix automaton was introduced in 1983 by a group of scientists from University of Denver and University of Colorado Boulder consisting of Anselm Blumer, Janet Blumer, Andrzej Ehrenfeucht, David Haussler and Ross McConnell, although similar concepts had earlier been studied alongside suffix trees in the works of Peter Weiner, Vaughan Pratt and Anatol Slissenko. In their initial work, Blumer et al. showed a suffix automaton built for the string S {\displaystyle S} of length greater than 1 {\displaystyle 1} has at most 2 | S | − 1 {\displaystyle 2|S|-1} states and at most 3 | S | − 4 {\displaystyle 3|S|-4} transitions, and suggested a linear algorithm for automaton construction. In 1983, Mu-Tian Chen and Joel Seiferas independently showed that Weiner's 1973 suffix-tree construction algorithm while building a suffix tree of the string S {\displaystyle S} constructs a suffix automaton of the reversed string S R {\textstyle S^{R}} as an auxiliary structure. In 1987, Blumer et al. applied the compressing technique used in suffix trees to a suffix automaton and invented the compacted suffix automaton, which is also called the compacted directed acyclic word graph (CDAWG). In 1997, Maxime Crochemore and Renaud Vérin developed a linear algorithm for direct CDAWG construction. In 2001, Shunsuke Inenaga et al. developed an algorithm for construction of CDAWG for a set of words given by a trie. == Definitions == Usually when speaking about suffix automata and related concepts, some notions from formal language theory and automata theory are used, in particular: "Alphabet" is a finite set Σ {\displaystyle \Sigma } that is used to construct words. Its elements are called "characters"; "Word" is a finite sequence of characters ω = ω 1 ω 2 … ω n {\displaystyle \omega =\omega _{1}\omega _{2}\dots \omega _{n}} . "Length" of the word ω {\displaystyle \omega } is denoted as | ω | = n {\displaystyle |\omega |=n} ; "Formal language" is a set of words over given alphabet; "Language of all words" is denoted as Σ ∗ {\displaystyle \Sigma ^{}} (where the "" character stands for Kleene star), "empty word" (the word of zero length) is denoted by the character ε {\displaystyle \varepsilon } ; "Concatenation of words" α = α 1 α 2 … α n {\displaystyle \alpha =\alpha _{1}\alpha _{2}\dots \alpha _{n}} and β = β 1 β 2 … β m {\displaystyle \beta =\beta _{1}\beta _{2}\dots \beta _{m}} is denoted as α ⋅ β {\displaystyle \alpha \cdot \beta } or α β {\displaystyle \alpha \beta } and corresponds to the word obtained by writing β {\displaystyle \beta } to the right of α {\displaystyle \alpha } , that is, α β = α 1 α 2 … α n β 1 β 2 … β m {\displaystyle \alpha \beta =\alpha _{1}\alpha _{2}\dots \alpha _{n}\beta _{1}\beta _{2}\dots \beta _{m}} ; "Concatenation of languages" A {\displaystyle A} and B {\displaystyle B} is denoted as A ⋅ B {\displaystyle A\cdot B} or A B {\displaystyle AB} and corresponds to the set of pairwise concatenations A B = { α β : α ∈ A , β ∈ B } {\displaystyle AB=\{\alpha \beta :\alpha \in A,\beta \in B\}} ; If the word ω ∈ Σ ∗ {\displaystyle \omega \in \Sigma ^{}} may be represented as ω = α γ β {\displaystyle \omega =\alpha \gamma \beta } , where α , β , γ ∈ Σ ∗ {\displaystyle \alpha ,\beta ,\gamma \in \Sigma ^{}} , then words α {\displaystyle \alpha } , β {\displaystyle \beta } and γ {\displaystyle \gamma } are called "prefix", "suffix" and "subword" (substring) of the word ω {\displaystyle \omega } correspondingly; If T = T 1 … T n {\displaystyle T=T_{1}\dots T_{n}} and T l T l + 1 … T r = S {\displaystyle T_{l}T_{l+1}\dots T_{r}=S} (with 1 ≤ l ≤ r ≤ n {\displaystyle 1\leq l\leq r\leq n} ) then S {\displaystyle S} is said to "occur" in T {\displaystyle T} as a subword. Here l {\displaystyle l} and r {\displaystyle r} are called left and right positions of occurrence of S {\displaystyle S} in T {\displaystyle T} correspondingly. == Automaton structure == Formally, deterministic finite automaton is determined by 5-tuple A = ( Σ , Q , q 0 , F , δ ) {\displaystyle {\mathcal {A}}=(\Sigma ,Q,q_{0},F,\delta )} , where: Σ {\displaystyle \Sigma } is an "alphabet" that is used to construct words, Q {\displaystyle Q} is a set of automaton "states", q 0 ∈ Q {\displaystyle q_{0}\in Q} is an "initial" state of automaton, F ⊂ Q {\displaystyle F\subset Q} is a set of "final" states of automaton, δ : Q × Σ ↦ Q {\displaystyle \delta :Q\times \Sigma \mapsto Q} is a partial "transition" function of automaton, such that δ ( q , σ ) {\displaystyle \delta (q,\sigma )} for q ∈ Q {\displaystyle q\in Q} and σ ∈ Σ {\displaystyle \sigma \in \Sigma } is either undefined or defines a transition from q {\displaystyle q} over character σ {\displaystyle \sigma } . Most commonly, deterministic finite automaton is represented as a directed graph ("diagram") such that: Set of graph vertices corresponds to the state of states Q {\displaystyle Q} , Graph has a specific marked vertex corresponding to initial state q 0 {\displaystyle q_{0}} , Graph has several marked vertices corresponding to the set of final states F {\displaystyle F} , Set of graph arcs corresponds to the set of transitions δ {\displaystyle \delta } , Specifically, every transition δ ( q 1 , σ ) = q 2 {\textstyle \delta (q_{1},\sigma )=q_{2}} is represented by an arc from q 1 {\displaystyle q_{1}} to q 2 {\displaystyle q_{2}} marked with the character σ {\displaystyle \sigma } . This transition also may be denoted as q 1 σ ⟶ q 2 {\textstyle q_{1}{\begin{smallmatrix}{\sigma }\\[-5pt]{\longrightarrow }\end{smallmatrix}}q_{2}} . In terms of its diagram, the automaton recognizes the word ω = ω 1 ω 2 … ω m {\displaystyle \omega =\omega _{1}\omega _{2}\dots \omega _{m}} only if there is a path from the initial vertex q 0 {\displaystyle q_{0}} to some final vertex q ∈ F {\displaystyle q\in F} such that concatenation of characters on this path forms ω {\displaystyle \omega } . The set of words recognized by an automaton forms a language that is set to be recognized by the automaton. In these terms, the language recognized by a suffix automaton of S {\displaystyle S} is the language of its (possibly empty) suffixes. === Automaton states === "Right context" of the word ω {\displaystyle \omega } with respect to language L {\displaystyle L} is a set [ ω ] R = { α : ω α ∈ L } {\displaystyle [\omega ]_{R}=\{\alpha :\omega \alpha \in L\}} that is a set of words α {\displaystyle \alpha } such that their concatenation with ω {\displaystyle \omega } forms a word from L {\displaystyle L} . Right contexts induce a natural equivalence relation [ α ] R = [ β ] R {\displaystyle [\alpha ]_{R}=[\beta ]_{R}} on the set of all words. If language L {\displaystyle L} is recognized by some deterministic finite automaton, there exists unique up to isomorphism automaton that recognizes the same language and has the minimum possible number of states. Such an automaton is called a minimal automaton for the given language L {\displaystyle L} . Myhill–Nerode theorem allows it to define it explicitly in terms of right contexts: In these terms, a "suffix automaton" is the minimal deterministic finite automaton recognizing the language of suffixes of the word S = s 1 s 2 … s n {\displaystyle S=s_{1}s_{2}\dots s_{n}} . The right context of the word ω {\displaystyle \omeg
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.
Embodied agent
In artificial intelligence, an embodied agent, also sometimes referred to as an interface agent, is an intelligent agent that interacts with the environment through a physical body within that environment. Agents that are represented graphically with a body, for example a human or a cartoon animal, are also called embodied agents, although they have only virtual, not physical, embodiment. A branch of artificial intelligence focuses on empowering such agents to interact autonomously with human beings and the environment. Mobile robots are one example of physically embodied agents; Ananova and Microsoft Agent are examples of graphically embodied agents. Embodied conversational agents are embodied agents (usually with a graphical front-end as opposed to a robotic body) that are capable of engaging in conversation with one another and with humans employing the same verbal and nonverbal means that humans do (such as gesture, facial expression, and so forth). == Embodied conversational agents == Embodied conversational agents are a form of intelligent user interface. Graphically embodied agents aim to unite gesture, facial expression and speech to enable face-to-face communication with users, providing a powerful means of human-computer interaction. == Advantages == Face-to-face communication allows communication protocols that give a much richer communication channel than other means of communicating. It enables pragmatic communication acts such as conversational turn-taking, facial expression of emotions, information structure and emphasis, visualization and iconic gestures, and orientation in a three-dimensional environment. This communication takes place through both verbal and non-verbal channels such as gaze, gesture, spoken intonation and body posture. Research has found that users prefer a non-verbal visual indication of an embodied system's internal state to a verbal indication, demonstrating the value of additional non-verbal communication channels. As well as this, the face-to-face communication involved in interacting with an embodied agent can be conducted alongside another task without distracting the human participants, instead improving the enjoyment of such an interaction. Furthermore, the use of an embodied presentation agent results in improved recall of the presented information. Embodied agents also provide a social dimension to the interaction. Humans willingly ascribe social awareness to computers, and thus interaction with embodied agents follows social conventions, similar to human to human interactions. This social interaction both raises the believably and perceived trustworthiness of agents, and increases the user's engagement with the system. Rickenberg and Reeves found that the presence of an embodied agent on a website increased the level of user trust in that website, but also increased users' anxiety and affected their performance, as if they were being watched by a real human. Another effect of the social aspect of agents is that presentations given by an embodied agent are perceived as being more entertaining and less difficult than similar presentations given without an agent. Research shows that perceived enjoyment, followed by perceived usefulness and ease of use, is the major factor influencing user adoption of embodied agents. A study in January 2004 by Byron Reeves at Stanford demonstrated how digital characters could "enhance online experiences" through explaining how virtual characters essentially add a sense of familiarity to the user experience and make it more approachable. This increase in likability in turn helps make the products better, which benefits both the end users and those creating the product. === Applications === The rich style of communication that characterizes human conversation makes conversational interaction with embodied conversational agents ideal for many non-traditional interaction tasks. A familiar application of graphically embodied agents is computer games; embodied agents are ideal for this setting because the richer communication style makes interacting with the agent enjoyable. Embodied conversational agents have also been used in virtual training environments, portable personal navigation guides, interactive fiction and storytelling systems, interactive online characters and automated presenters and commentators. Major virtual assistants like Siri, Amazon Alexa and Google Assistant do not come with any visual embodied representation, which is believed to limit the sense of human presence by users. The U.S. Department of Defense utilizes a software agent called SGT STAR on U.S. Army-run Web sites and Web applications for site navigation, recruitment and propaganda purposes. Sgt. Star is run by the Army Marketing and Research Group, a division operated directly from The Pentagon. Sgt. Star is based upon the ActiveSentry technology developed by Next IT, a Washington-based information technology services company. Other such bots in the Sgt. Star "family" are utilized by the Federal Bureau of Investigation and the Central Intelligence Agency for intelligence gathering purposes.
World model (artificial intelligence)
A world model in artificial intelligence is a machine learning system that builds an internal representation of an environment. The model predicts how that environment changes over time in response to actions. Researchers design world models to help agents plan, reason, and act without constant real-world trial and error. World models differ from systems that merely classify or generate outputs. They simulate dynamics such as physics, object interactions, and causality. Early ideas date to the 1990s. Modern versions power robots, autonomous driving, and interactive video generation. == History == Jürgen Schmidhuber introduced the term world model in machine learning in 1990. He proposed recurrent neural networks that predict future states from observations and use those predictions to train agents. David Ha and Schmidhuber revived the concept in a 2018 paper. Their agents learned to drive virtual cars and play video games inside self-generated simulations. Yann LeCun advanced the idea in a 2022 position paper titled "A Path Towards Autonomous Machine Intelligence". He argued that intelligence requires predictive models of the world rather than pure pattern matching. LeCun proposed the joint embedding predictive architecture (JEPA) as a practical foundation. LeCun and collaborators developed several JEPA variants. V-JEPA 2 reached state-of-the-art performance on video understanding and physical reasoning at the time. It supports zero-shot robot control in unfamiliar environments. Introduced in March 2026, LeWorldModel trains stably end-to-end from raw pixels and uses two loss terms and avoids hand-crafted heuristics. LeCun founded Advanced Machine Intelligence Labs in 2026 to further develop world models. Google DeepMind introduced Genie in 2024. The model learned interactive environments from unlabeled internet videos. Genie 2 followed in late 2024 and added three-dimensional generation. The Genie series set benchmarks for general-purpose simulation. Genie 3 was introduced in August 2025. It produces photorealistic, real-time interactive worlds from text prompts which are displayed at 24 frames per second and explored in real time with text or image prompts. The model supports persistent three-dimensional worlds and real-time interaction. Waymo adopted Genie 3 in February 2026 and used it to create a specialized world model for autonomous driving simulation, called the Waymo World Model. It produces synchronized camera and lidar outputs and creates edge cases that real robotaxis rarely encounter. The edge cases were reported to be unusual by PCMag. General Intuition announced a $133.7 million seed round. World Labs raised $1 billion. AMI raised $1.03 billion. In April 2026, Alibaba announced Happy Oyster, its world model designed for real-time and “flowy” world model. It includes a directing mode for world building based on text and image prompts and a wandering mode for exploring the resulting world. It can generate 3-minute in-world video clips. Also in April, World Labs, co-founded by Li Fei Fei, unveiled Spark 2.0, an open-source 3D Gaussian splatting rendering engine that targets smartphone-class devices. In June 2026, Nvidia released Cosmos 3, a family of open-weight models. It combines previously independent physical reasoning, world simulation, and action generation. Cosmos 3 integrates can process and generate text, image, video, audio, and action sequences. The model employs a Mixture-of-Transformers" (MoT) approach. An autoregressive (AR) transformer handles reasoning and next-token prediction, while a diffusion transformer (DT) does multimodal generation. Encoders (ViT for vision, VAE for visual/audio, and domain-specific for actions) and generate a shared representation space using 3D multi-dimensional rotary position embedding (mRoPE) for spatial and temporal information. The family includes Cosmos3-Nano (16B parameters) for workstations; Cosmos3-Super (64B parameters) for research. == Architecture == World models process raw sensory data such as video frames or lidar scans. They compress this input into compact latent representations. The system then predicts future representations rather than pixel-by-pixel reconstructions. Many modern world models use joint embedding predictive architecture (JEPA). An encoder turns observations into embeddings. A predictor estimates one or a suite of embeddings from the current one and an action. In some cases a critic chooses one embedding as the best result. A regularizer keeps embeddings well-behaved. The model trains by minimizing prediction error in embedding space. This approach avoids the high cost of generating every detail. Some architectures add explicit components. A fast reactive path handles immediate responses. A slower deliberative path performs longer-horizon planning. Video prediction accuracy or robot success rates are key metrics, but do not always predict real-world performance. Generative world models such as Genie 3 combine these with a simulator. They accept text prompts or layouts and output consistent video, lidar, or three-dimensional scenes. World models often train with self-supervised learning. They use large unlabeled datasets of video or robot interactions. Self-supervised learning can speed learning. Reinforcement learning can fine-tune a model for specific tasks. == Applications == World models support robot learning. Agents train inside simulations and transfer skills to the physical world. This reduces the need for dangerous or expensive real-world trials. Autonomous vehicles use world models to test rare events. Waymo's system simulates tornadoes or unusual pedestrian behavior. Companies train planners without putting vehicles on public roads. Interactive entertainment benefits from world models. Genie 3 lets users generate playable environments from simple descriptions. Game studios prototype levels faster. Scientific simulation gains from these models. Researchers model physical systems or biological processes at scale. Planners in logistics or urban design test strategies inside accurate digital twins. == Comparison with large language models == Both world models and large language models (LLMs) use inferencing on their inputs to make predictions. LLMs operate on textual inputs. They predict the next token in text sequences. They excel at language-oriented tasks such as translation or summarization. However, they lack understanding of physics. World models operate on sensor inputs such as pixels. They predict state changes in that data in latent space. This design supports planning and causal reasoning. LLMs generate fluent text but often fail at consistent physical predictions. Their architecture employs transformers with refinements such as mixture of experts. World models divide an inferencing task into work performed by encoders, predictors, simulators, and other pieces. They typically handle multimodal inputs such as video, lidar, radar, and audio, guided by textual prompting. LLMs power chatbots and code assistants. World models drive embodied agents that act in dynamic environments, such as autonomous driving. The two may be combined in hybrid systems. For example, a LLM handles instructions, while a world model manages low-level control. World model proponents such as LeCun claim that because LLMs are trained only on text, they have no ability to predict anything beyond text, such as real-world events. == Benchmarks == World model benchmarks test physical understanding, long-term consistency, planning, and generalization from sensor data. Meta introduced three benchmarks for V-JEPA 2. IntPhys 2 measures a model's ability to detect physics violations. It presents pairs of videos that diverge when one breaks physical rules. Humans score near 100% accuracy. V-JEPA 2 achieves little better than random chance on many conditions. Minimal Video Pairs (MVPBench) tests physical understanding through multiple-choice questions based on short video clips. It probes object interactions and causality. Something-Something tests action recognition. Epic-Kitchens-100 tests human action anticipation. DeepMind benchmark: Interactive evaluation measures consistency over minutes of interaction, memory of off-screen objects, and response to user actions or text prompts. Waymo benchmark: Output generation quality: Metrics include realism, controllability (via text prompts), and usefulness for training planners in simulated worlds. However, pixel reconstruction error rate with episodic rewards often fails. Other: Epic-Kitchens-100 (often measured with Recall@5) Ego4D 50 Salads, Breakfast, etc. Potential benchmarks: Zero-shot transfer to robots Long-horizon planning Implausible prediction rate
MobileNet
MobileNet is a family of convolutional neural network (CNN) architectures designed for image classification, object detection, and other computer vision tasks. They are designed for small size, low latency, and low power consumption, making them suitable for on-device inference and edge computing on resource-constrained devices like mobile phones and embedded systems. They were originally designed to be run efficiently on mobile devices with TensorFlow Lite. The need for efficient deep learning models on mobile devices led researchers at Google to develop MobileNet. As of June 2025, the family has five versions, each improving upon the previous one in terms of performance and efficiency. == Features == === V1 === MobileNetV1 was published in April 2017. Its main architectural innovation was incorporation of depthwise separable convolutions. It was first developed by Laurent Sifre during an internship at Google Brain in 2013 as an architectural variation on AlexNet to improve convergence speed and model size. The depthwise separable convolution decomposes a single standard convolution into two convolutions: a depthwise convolution that filters each input channel independently and a pointwise convolution ( 1 × 1 {\displaystyle 1\times 1} convolution) that combines the outputs of the depthwise convolution. This factorization significantly reduces computational cost. The MobileNetV1 has two hyperparameters: a width multiplier α {\displaystyle \alpha } that controls the number of channels in each layer. Smaller values of α {\displaystyle \alpha } lead to smaller and faster models, but at the cost of reduced accuracy, and a resolution multiplier ρ {\displaystyle \rho } , which controls the input resolution of the images. Lower resolutions result in faster processing but potentially lower accuracy. === V2 === MobileNetV2 was published in March 2019. It uses inverted residual layers and linear bottlenecks. Inverted residuals modify the traditional residual block structure. Instead of compressing the input channels before the depthwise convolution, they expand them. This expansion is followed by a 1 × 1 {\displaystyle 1\times 1} depthwise convolution and then a 1 × 1 {\displaystyle 1\times 1} projection layer that reduces the number of channels back down. This inverted structure helps to maintain representational capacity by allowing the depthwise convolution to operate on a higher-dimensional feature space, thus preserving more information flow during the convolutional process. Linear bottlenecks removes the typical ReLU activation function in the projection layers. This was rationalized by arguing that that nonlinear activation loses information in lower-dimensional spaces, which is problematic when the number of channels is already small. === V3 === MobileNetV3 was published in 2019. The publication included MobileNetV3-Small, MobileNetV3-Large, and MobileNetEdgeTPU (optimized for Pixel 4). They were found by a form of neural architecture search (NAS) that takes mobile latency into account, to achieve good trade-off between accuracy and latency. It used piecewise-linear approximations of swish and sigmoid activation functions (which they called "h-swish" and "h-sigmoid"), squeeze-and-excitation modules, and the inverted bottlenecks of MobileNetV2. === V4 === MobileNetV4 was published in September 2024. The publication included a large number of architectures found by NAS. Inspired by Vision Transformers, the V4 series included multi-query attention. It also unified both inverted residual and inverted bottleneck from the V3 series with the "universal inverted bottleneck", which includes these two as special cases. === V5 === MobileNetV5's architecture was published shortly after the release of Gemma 3n in June 2025. While the announcement stated a technical report on MobileNetV5 would be available soon, this has not yet materialised. The network is 10 times larger than the largest V4 variant.
Lesk algorithm
The Lesk algorithm is a classical algorithm for word sense disambiguation introduced by Michael E. Lesk in 1986. It operates on the premise that words within a given context are likely to share a common meaning. This algorithm compares the dictionary definitions of an ambiguous word with the words in its surrounding context to determine the most appropriate sense. Variations, such as the Simplified Lesk algorithm, have demonstrated improved precision and efficiency. However, the Lesk algorithm has faced criticism for its sensitivity to definition wording and its reliance on brief glosses. Researchers have sought to enhance its accuracy by incorporating additional resources like thesauruses and syntactic models. == Overview == The Lesk algorithm is based on the assumption that words in a given "neighborhood" (section of text) will tend to share a common topic. A simplified version of the Lesk algorithm is to compare the dictionary definition of an ambiguous word with the terms contained in its neighborhood. Versions have been adapted to use WordNet. An implementation might look like this: for every sense of the word being disambiguated one should count the number of words that are in both the neighborhood of that word and in the dictionary definition of that sense the sense that is to be chosen is the sense that has the largest number of this count. A frequently used example illustrating this algorithm is for the context "pine cone". The following dictionary definitions are used: PINE 1. kinds of evergreen tree with needle-shaped leaves 2. waste away through sorrow or illness CONE 1. solid body which narrows to a point 2. something of this shape whether solid or hollow 3. fruit of certain evergreen trees As can be seen, the best intersection is Pine #1 ⋂ Cone #3 = 2. == Simplified Lesk algorithm == In Simplified Lesk algorithm, the correct meaning of each word in a given context is determined individually by locating the sense that overlaps the most between its dictionary definition and the given context. Rather than simultaneously determining the meanings of all words in a given context, this approach tackles each word individually, independent of the meaning of the other words occurring in the same context. "A comparative evaluation performed by Vasilescu et al. (2004) has shown that the simplified Lesk algorithm can significantly outperform the original definition of the algorithm, both in terms of precision and efficiency. By evaluating the disambiguation algorithms on the Senseval-2 English all words data, they measure a 58% precision using the simplified Lesk algorithm compared to the only 42% under the original algorithm. Note: Vasilescu et al. implementation considers a back-off strategy for words not covered by the algorithm, consisting of the most frequent sense defined in WordNet. This means that words for which all their possible meanings lead to zero overlap with current context or with other word definitions are by default assigned sense number one in WordNet." Simplified LESK Algorithm with smart default word sense (Vasilescu et al., 2004) The COMPUTEOVERLAP function returns the number of words in common between two sets, ignoring function words or other words on a stop list. The original Lesk algorithm defines the context in a more complex way. == Criticisms == Unfortunately, Lesk’s approach is very sensitive to the exact wording of definitions, so the absence of a certain word can radically change the results. Further, the algorithm determines overlaps only among the glosses of the senses being considered. This is a significant limitation in that dictionary glosses tend to be fairly short and do not provide sufficient vocabulary to relate fine-grained sense distinctions. A lot of work has appeared offering different modifications of this algorithm. These works use other resources for analysis (thesauruses, synonyms dictionaries or morphological and syntactic models): for instance, it may use such information as synonyms, different derivatives, or words from definitions of words from definitions. == Lesk variants == Original Lesk (Lesk, 1986) Adapted/Extended Lesk (Banerjee and Pederson, 2002/2003): In the adaptive lesk algorithm, a word vector is created corresponds to every content word in the wordnet gloss. Concatenating glosses of related concepts in WordNet can be used to augment this vector. The vector contains the co-occurrence counts of words co-occurring with w in a large corpus. Adding all the word vectors for all the content words in its gloss creates the Gloss vector g for a concept. Relatedness is determined by comparing the gloss vector using the Cosine similarity measure. There are a lot of studies concerning Lesk and its extensions: Wilks and Stevenson, 1998, 1999; Mahesh et al., 1997; Cowie et al., 1992; Yarowsky, 1992; Pook and Catlett, 1988; Kilgarriff and Rosensweig, 2000; Kwong, 2001; Nastase and Szpakowicz, 2001; Gelbukh and Sidorov, 2004.
Embedding (machine learning)
In machine learning, embedding is a representation learning technique that maps complex, high-dimensional data into a lower-dimensional vector space of numerical vectors. == Technique == It also denotes the resulting representation, where meaningful patterns or relationships are preserved. As a technique, it learns these vectors from data like words, images, or user interactions, differing from manually designed methods such as one-hot encoding. This process reduces complexity and captures key features without needing prior knowledge of the domain. == Similarity == In natural language processing, words or concepts may be represented as feature vectors, where similar concepts are mapped to nearby vectors. The resulting embeddings vary by type, including word embeddings for text (e.g., Word2Vec), image embeddings for visual data, and knowledge graph embeddings for knowledge graphs, each tailored to tasks like NLP, computer vision, or recommendation systems. This dual role enhances model efficiency and accuracy by automating feature extraction and revealing latent similarities across diverse applications. To measure the distance between two embeddings, a similarity measure can be used to find the overall similarity of the concepts represented by the embeddings. If the vectors are normalized to have a magnitude of 1, then the similarity measures are proportional to cos ( θ a b ) {\displaystyle \cos \left(\theta _{ab}\right)} . The cosine similarity disregards the magnitude of the vector when determining similarity, so it is less biased towards training data that appears very frequently. The dot product includes the magnitude inherently, so it will tend to value more popular data. Generally, for high-dimensional vector spaces, vectors tend to converge in distance, so Euclidean distance becomes less reliable for large embedding vectors.