DAYDREAMER is a goal-based agent and cognitive architecture developed at the University of California, Los Angeles by Erik T. Mueller and Michael G. Dyer beginning in 1983. The system models the human stream of thought and how it is triggered and directed by emotions, simulating human daydreaming. Taking situational descriptions as input, DAYDREAMER produces English-language daydreams as output and encodes new daydreams, plans, and planning strategies for later reuse. The program comprises five components: a scenario generator based on relaxed planning, a dynamic episodic memory, a collection of personal goals and control goals, an emotion component, and domain knowledge of interpersonal relations and everyday occurrences. The source code was released under a free software license in 2015. == History == Erik Mueller began DAYDREAMER in 1983 while he was a doctoral student in the Artificial Intelligence Laboratory of the Computer Science Department at the University of California, Los Angeles, studying under Michael G. Dyer. Initial development of the project was supported by a grant from the W. M. Keck Foundation with matching funds from the UCLA School of Engineering and Applied Sciences. Additionally, Mueller was supported by an Atlantic Richfield Doctoral Fellowship and Dyer by an IBM Faculty Development Award. The first published descriptions of the program appeared in 1985 at the Ninth International Joint Conference on Artificial Intelligence in Los Angeles and at the Seventh Annual Conference of the Cognitive Science Society in Irvine. Work on the program continued, and a book, Daydreaming in Humans and Machines, was published by Ablex Publishing in 1990. The program was implemented on top of GATE, a knowledge-representation and inference substrate developed by Mueller and Uri Zernik at UCLA, and was originally written in T, a dialect of Scheme. In 2015, Mueller released the DAYDREAMER source code, version 3.5, a Common Lisp rewrite of the original T implementation, on GitHub under the GNU General Public License version 2. The release comprised approximately 12,000 lines of Common Lisp code, along with the GATE knowledge-representation substrate on which DAYDREAMER had originally been built. == Architecture == The program operates in two modes. In daydreaming mode it daydreams continuously until interrupted, while performance mode allows it to demonstrate behavior it has learned through daydreaming. === Emotion and control goals === Emotions and daydreaming form a feedback loop for DAYDREAMER. Emotions activate goals that produce daydreams, and the resulting daydreams modify existing emotions and trigger new ones, which prompt subsequent daydreaming. Recall of a goal success produces a positive emotion whereas recall of a goal failure produces a negative emotion. Emotions activate a set of goals, called control goals, which direct the course of a daydream. The program has four control goals. "Rationalization" generates reasons why an unsatisfactory outcome is in fact acceptable, in order to reduce a negative emotion and maintain self-esteem. "Revenge" is activated by anger when a failure is caused by another and reduces negative emotion through imagined retaliation. "Failure/success reversal" imagines alternative scenarios in which a failure was prevented or a success did not occur as a means of learning planning strategies for future situations. "Preparation" generates hypothetical future scenarios in order to rehearse plans and actions for events that have not yet occurred. === Scenario generator and relaxed planning === The scenario generator produces the sequence of events that make up a daydream. It operates under multiple, often conflicting personal goals rather than pursuing a single goal, applies relaxation rules that permit the generation of non-realistic scenarios, and it draws on episodic memory of past experiences both as subject matter and as a source of planning knowledge. The personal goals that guide the scenario generator include health, food, sex, friendship, love, possessions, self-esteem, social esteem, enjoyment, and achievement. These goals are organized into a goal tree that specifies their relative importance at any given time. Relaxation rules allow the program to set aside its ordinary constraints when generating a scenario. The four constraints that may be relaxed are the behavior of others, the daydreamer's own attributes, physical constraints, and social constraints. The degree of relaxation varies with the active control goal. For example a failure-reversal goal aimed at alternatives uses a low level of relaxation, whereas a revenge goal aimed at a retaliation uses a high level. === Episodic memory and analogy === DAYDREAMER's episodic memory stores its personal and vicarious experiences along with the daydreams it generates. The memory is described as dynamic because it is continually modified during daydreaming such that previously daydreamed episodes become available alongside real ones. As it daydreams, the program indexes daydreams, future plans or actions, and planning strategies into memory. Episodes are organized and retrieved using surface-level similarities, emotions, abstract themes, and Plot Units which are abstract configurations of positive and negative outcomes developed by Wendy Lehnert. A recalled episode is adapted to the current situation through analogy, which requires less effort than generating an equivalent scenario from scratch. == Sample output == In the sample experience from the source code, called LOVERS1, DAYDREAMER begins from an initial situation in which it has a job, is not romantically involved, and is at home. Starting in daydreaming mode, it activates a top-level goal to be in a romantic relationship because it is not currently in one, and a positive motivating emotion of interest becomes associated with that goal. The program then activates a goal to be entertained and pursues seeing a film as a way to achieve it. Facts asserted into memory are converted to English and produced as output, such as "I want to be going out with someone" and "I have to go see a movie". == Reception and influence == DAYDREAMER has been cited in research on computational models of creativity, emotion, and narrative. Linda Wills and Janet Kolodner cite the program as an example of work on opportunism in their study of serendipitous recognition in design. Joseph Bates, A. Bryan Loyall, and W. Scott Reilly of the Carnegie Mellon Oz Project cite DAYDREAMER among prior work in their description of an architecture combining action, emotion, and social behavior. Rafael Pérez y Pérez, Ricardo Sosa, and Christian Lemaitre cite Mueller's DAYDREAMER as one of the few computer models at the time to model daydreaming during the creative process. Jichen Zhu and D. Fox Harrell likewise cite the program in their work on imagining and agency in generative interactive narrative.
T-pose
In computer animation, a T-pose is a default posing for a humanoid 3D model's skeleton before it is animated. It is called so because of its shape: the straight legs and arms of a humanoid model combine to form a capital letter T. When the arms are angled downwards, the pose is sometimes referred to as an A-pose instead. Likewise, if the arms are angled upward, it is called a Y-pose. Generic terms encompassing all these (especially for non-humanoid models) include bind pose, blind pose, and reference pose. == Usage == The T-pose is primarily used as the default armature pose for skeletal animation in 3D software, which is then manipulated to create animation. The purpose of the T-pose relates to the important elements of the body being axis-aligned, thereby making it easier to rig the model for animation, physics, and other controls. Depending on the exact geometry of the model, other poses such as the A-pose may be more suitable for vertex deformation around areas such as the shoulders. Outside of being default poses in animation software, T-poses are typically used as placeholders for animation not yet completed, particularly in 3D animated video games. In some motion capture software, a T-pose must be assumed by the actor in the motion capture suit before motion capturing can begin. There are other poses used, but the T-pose is the most common one. == As an Internet meme == Starting in 2016 and resurfacing in 2017, the T-pose has become a widespread Internet meme due to its bizarre and somewhat comedic appearance, especially in video game glitches where a character's animation is unexpectedly supplanted by a T-pose. In a prerelease video of the game NBA Elite 11, the demo was filled with glitches, notably one unintentionally showing a T-pose in place of the proper animation for the model of player Andrew Bynum. The glitch later gained fame as the "Jesus Bynum glitch". Publisher EA eventually cancelled the game as they found it unsatisfactory. A similar occurrence happened with Cyberpunk 2077. In the 2023 Formula One season, driver George Russell performed a T-pose in the opening credits of the series' TV broadcasts. This quickly became a meme within the motorsports community. Russell repeated the pose after claiming pole position at the 2024 Canadian Grand Prix and winning the 2024 Austrian Grand Prix.
GNOWSYS
GNOWSYS (Gnowledge Networking and Organizing system) is a specification for a generic distributed network based memory/knowledge management. It is developed as an application for developing and maintaining semantic web content. It is written in Python. It is implemented as a Django app. The GNOWSYS project was launched by Nagarjuna G. in 2001, while he was working at Homi Bhabha Centre for Science Education (HBCSE). The memory of GNOWSYS is designed as a node-oriented space. A node is described by other nodes to which it has links. The nodes are organized and processed according to a complex data structure called the neighborhood. == Applications == The application can be used for web-based knowledge representation and content management projects, for developing structured knowledge bases, as a collaborative authoring tool, suitable for making electronic glossaries, dictionaries and encyclopedias, for managing large web sites or links, developing an online catalogue for a library of any thing including books, to make ontologies, classifying and networking any objects, etc. This tool is also intended to be used for an on-line tutoring system with dependency management between various concepts or software packages. For example, the dependency relations between Debian packages have been represented by the gnowledge portal Archived 2018-05-14 at the Wayback Machine. == Component Classes == The kernel is designed to provide support to persistently store highly granular nodes of knowledge representation like terms, predicates and very complex propositional systems like arguments, rules, axiomatic systems, loosely held paragraphs, and more complex structured and consistent compositions. All the component classes in GNOWSYS are classified according to complexity into three groups, where the first two groups are used to express all possible well formed formulae permissible in a first order logic. === Terms === ‘Object’, ‘Object Type’ for declarative knowledge, ‘Event’, ‘Event Type’, for temporal objects, and ‘Meta Types’ for expressing upper ontology. The objects in this group are essentially any thing about which the knowledge engineer intends to express and store in the knowledge base, i.e., they are the objects of discourse. The instances of these component classes can be stored with or without expressing ‘instance of’ or ‘sub-class of’ relations among them. === Predicates === This group consists of ‘Relation’, and ‘Relation Type’ for expressing declarative knowledge, and ‘Function’ and ‘Function Type’ for expressing procedural knowledge. This group is to express qualitative and quantitative relations among the various instances stored in the knowledge base. While instantiating the predicates can be characterized by their logical properties of relations, quantifiers and cardinality as monadic predicates of these predicate objects. === Structures === ‘System’, ‘Encapsulated Class’, ‘Program’, and ‘Process’, are other base classes for complex structures, which can be combined iteratively to produce more complex systems. The component class ‘System’ is to store in the knowledge base a set of propositions composed into ontologies, axiomatic systems, complex systems like say a human body, an artifact like a vehicle etc., with or without consistency check. An ‘Encapsulated Class’ is to com- pose declarative and behavioural objects in a flexible way to build classes. A ‘Program’ is not only to store the logic of any complete program or a component class, composed from the already available behavioural instances in the knowledge base with built-in connectives (conditions, and loops), but also execute them as web services. A ‘Process’ is to structure temporal objects with sequence, concurrency, synchronous or asynchronous specifications. Every node in the database keeps the neighbourhood information, such as its super-class, sub-class, instance-of, and other relations, in which the object has a role, in the form of predicates. This feature makes computation of drawing graphs and inferences, on the one hand, and dependency and navigation paths on the other hand very easy. All the data and metadata is indexed in a central catalogue making query and locating resources efficient.
SHRDLU
SHRDLU is an early natural-language understanding computer program that was developed by Terry Winograd at MIT in 1968–1970. In the program, the user carries on a conversation with the computer, moving objects, naming collections and querying the state of a simplified "blocks world", essentially a virtual box filled with different blocks. SHRDLU was written in the Micro Planner and Lisp programming language on the DEC PDP-6 computer and a DEC graphics terminal. Later additions were made at the computer graphics labs at the University of Utah, adding a full 3D rendering of SHRDLU's "world". The name SHRDLU was derived from ETAOIN SHRDLU, the arrangement of the letter keys on a Linotype machine, arranged in descending order of usage frequency in English. == Functionality == SHRDLU is primarily a language parser that allows user interaction using English terms. The user instructs SHRDLU to move various objects around in the "blocks world" containing various basic objects such as blocks, cones and balls. SHRDLU combined four simple ideas to make the simulation of "understanding" more convincing. One was that SHRDLU's world is so simple that the entire set of objects and locations could be described by including as few as 50 words: nouns like "block" and "cone", verbs like "place on" and "move to", and adjectives like "big" and "blue". The possible combinations of these basic language building blocks are quite simple. SHRDLU also includes a basic memory to supply context. One could ask SHRDLU to "put the green cone on the red block" and then "take the cone off"; "the cone" would be taken to mean the green cone one had just talked about. SHRDLU can search back further through the interactions to find the proper context in most cases when additional adjectives were supplied. One could also ask questions about the history; for instance, one could ask "did you pick up anything before the cone?" A side effect of this memory, and the original rules SHRDLU was supplied with, is that the program can answer questions about what was possible in the world and what was not. For instance, SHRDLU can deduce that blocks could be stacked by looking for examples, but also realize that triangles could not be stacked, after having tried it. The "world" contains basic physics to make blocks fall over, independent of the language parser. Finally, SHRDLU can also remember names given to objects, or arrangements of them. For instance one could say "a steeple is a small triangle on top of a tall rectangle"; SHRDLU can then answer questions about steeples in the blocks world, and build new ones. This is an example of a SHRDLU session: Person: Pick up a big red block. Computer: OK. Person: Grasp the pyramid. Computer: I DON'T UNDERSTAND WHICH PYRAMID YOU MEAN. Person (changing their mind): Find a block which is taller than the one you are holding and put it into the box. Computer: BY "IT", I ASSUME YOU MEAN THE BLOCK WHICH IS TALLER THAN THE ONE I AM HOLDING. Computer: OK. Person: What does the box contain? Computer: THE BLUE PYRAMID AND THE BLUE BLOCK. Person: What is the pyramid supported by? Computer: THE BOX. Person: How many blocks are not in the box? Computer: FOUR OF THEM. Person: Is at least one of them narrower than the one which i told you to pick up? Computer: YES, THE RED CUBE. Person: Is it supported? Computer: YES, BY THE TABLE. Person: Can the table pick up blocks? Computer: NO. Person: Can a pyramid be supported by a block? Computer: YES. Person: Can a pyramid support a pyramid? Computer: I DON'T KNOW. Person: Stack up two pyramids. Computer: I CAN'T. Person: Will you please stack up both of the red blocks and either a green cube or a pyramid? Computer: OK. Person: Which cube is sitting on the table? Computer: THE LARGE GREEN ONE WHICH SUPPORTS THE RED PYRAMID. Person: Is there a large block behind a pyramid? Computer: YES, THREE OF THEM: A LARGE RED ONE, A LARGE GREEN CUBE, AND THE BLUE ONE. Person: Put a small one onto the green cube which supports a pyramid. Computer: OK. Person: Put the littlest pyramid on top of it. Computer: OK. Person: Does the shortest thing the tallest pyramid's support supports support anything green? Computer: YES, THE GREEN PYRAMID. Person: What color is it? Computer: BY "IT", I ASSUME YOU MEAN THE SHORTEST THING THE TALLEST PYRAMID'S SUPPORT SUPPORTS. Computer: RED Person: How many things are on top of green cubes? Computer: I'M NOT SURE WHAT YOU MEAN BY "ON TOP OF" IN THE PHRASE "ON TOP OF GREEN CUBES" . DO YOU MEAN: 1 - DIRECTLY ON THE SURFACE 2 - ANWHERE ON TOP OF ? Person: 2 Computer: THREE OF THEM. Person: Had you touched any pyramid before you put the green one on the little cube? Computer: YES, THE GREEN ONE. Person: When did you pick it up? Computer: WHILE I WAS STACKING UP THE RED CUBE, A LARGE RED BLOCK AND A LARGE GREEN CUBE. Person: Why? Computer: TO GET RID OF IT. Person: Why did you do that? Computer: TO CLEAN OFF THE RED CUBE. Person: Why did you clear off that cube? Computer: TO PUT IT ON A LARGE GREEN CUBE. Person: Why did you do that? Computer: TO STACK UP THE RED CUBE , A LARGE RED BLOCK AND A LARGE GREEN CUBE. Person: Why did you do that? Computer: BECAUSE YOU ASKED ME TO. Person: How did you do it? Computer: BY PUTTING A LARGE RED BLOCK ON THE TABLE ; THEN LETTING GO OF IT ; THEN PUTTING A LARGE GREEN CUBE ON IT ; THEN LETTING GO OF THAT CUBE ; THEN PUTTING THE RED CUBE ON THAT CUBE; THEN LETTING GO OF THAT CUBE. Person: How many objects did you touch while you were doing it? Computer: FOUR OF THEM. Person: What did the red cube support before you started to clean it off? Computer: THE GREEN PYRAMID. Person: There were five blocks to the left of the box then. Computer: NO, ONLY FOUR OF THEM: THE RED CUBE, TWO LARGE GREEN CUBES AND A LARGE RED BLOCK. Person: Put the blue pyramid on the block in the box. Computer: OK. Person: Is there anything which is bigger than every pyramid but is not as wide as the thing that supports it? Computer: YES, THE BLUE BLOCK. Person: Does a steeple Computer (interrupting): SORRY, I DON'T KNOW THE WORD "STEEPLE". Person: A "steeple" is a stack which contains two green cubes and a pyramid. Computer: I UNDERSTAND. Person: Are there any steeples now? Computer: NO. Person: Build one. Computer: OK. Person: Call the biggest block "superblock". Computer: OK. Person: Have you picked superblock up since we began? Computer: YES. Person: Why did you drop it? Computer: BY "IT", I ASSUME YOU MEAN THE SUPERBLOCK. I CAN'T EXPLAIN A NON-EXISTENT EVENT. Person: Is there anything to the right of the red pyramid? Computer: YES, FOUR THINGS: THE BOX, THE RED CUBE, THE BLUE PYRAMID AND THE BLOCK NAMED SUPERBLOCK Person: Thank you. Computer: YOU'RE WELCOME! == Consequences == SHRDLU was considered a tremendously successful demonstration of artificial intelligence (AI). This led other AI researchers to excessive optimism which was soon lost when later systems attempted to deal with situations with a more realistic level of ambiguity and complexity. Subsequent efforts of the SHRDLU type, such as Cyc, have tended to focus on providing the program with considerably more information from which it can draw conclusions. In a 1991 interview, Winograd said about SHRDLU: [...] the famous dialogue with SHRDLU where you could pick up a block, and so on, I very carefully worked through, line by line. If you sat down in front of it, and asked it a question that wasn't in the dialogue, there was some probability it would answer it. I mean, if it was reasonably close to one of the questions that was there in form and in content, it would probably get it. But there was no attempt to get it to the point where you could actually hand it to somebody and they could use it to move blocks around. And there was no pressure for that whatsoever. Pressure was for something you could demo. Take a recent example, Negroponte's Media Lab, where instead of "perish or publish" it's "demo or die." I think that's a problem. I think AI suffered from that a lot, because it led to "Potemkin villages", things which - for the things they actually did in the demo looked good, but when you looked behind that there wasn't enough structure to make it really work more generally. Though not intentionally developed as such, SHRDLU is considered the first known formal example of interactive fiction, as the user interacts with simple commands to move objects around a virtual environment, though lacking the distinct story-telling normally present in the interactive fiction genre. The 1976-1977 game Colossal Cave Adventure is broadly considered to be the first true work of interactive fiction.
Plinian Core
Plinian Core is a set of vocabulary terms that can be used to describe different aspects of biological species information. Under "biological species Information" all kinds of properties or traits related to taxa—biological and non-biological—are included. Thus, for instance, terms pertaining descriptions, legal aspects, conservation, management, demographics, nomenclature, or related resources are incorporated. == Description == The Plinian Core is aimed to facilitate the exchange of information about the species and upper taxa. What is in scope? Species level catalogs of any kind of biological objects or data. Terminology associated with biological collection data. Striving for compatibility with other biodiversity-related standards. Facilitating the addition of components and attributes of biological data. What is not in scope? Data interchange protocols. Non-biodiversity-related data. Occurrence level data. This standard is named after Pliny the Elder, a very influential figure in the study of the biological species. Plinian Core design requirements includes: ease of use, to be self-contained, able to support data integration from multiple databases, and ability to handle different levels of granularity. Core terms can be grouped in its current version as follows: Metadata Base Elements Record Metadata Nomenclature and Classification Taxonomic description Natural history Invasive species Habitat and Distribution Demography and Threats Uses, Management and Conservation associatedParty, MeasurementOrFact, References, AncillaryData == Background == Plinian Core started as a collaborative project between Instituto Nacional de Biodiversidad and GBIF Spain in 2005. A series of iterations in which elements were defined and implanted in different projects resulted in a "Plinian Core Flat" [deprecated]. As a result, a new development was impulse to overcome them in 2012. New formal requirements, additional input and a will to better support the standard and its documentation, as well as to align it with the processes of TDWG, the world reference body for biodiversity information standards. A new version, Plinian Core v3.x.x was defined. This provides more flexibility to fully represent the information of a species in a variety of scenarios. New elements to deal with aspects such as IPR, related resources, referenced, etc. were introduced, and elements already included were better-defined and documented. Partner for the development of Plinian Core in this new phase incorporated the University of Granada (UG, Spain), the Alexander von Humboldt Institute (IAvH, Colombia), the National Commission for the Knowledge and Use of Biodiversity (Conabio, Mexico) and the University of São Paulo (USP, Brazil). A "Plinian Core Task Group" within TDWG "Interest Group on species Information" was constituted and currently working on its development. == Levels of the standard == Plinian Core is presented in to levels: the abstract model and the application profiles. The abstract model (AM), comprising the abstract model schema(xsd) and the terms' URIs, is the normative part. It is all comprehensive, and allows for different levels of granularity in describing species properties. The AM should be taken as a "menu" from which to choose terms and level of detail needed in any specific project. The subsets of the abstract model intended to be implemented in specific projects are the "application profiles" (APs). Besides containing part of the elements of the AM, APs can impose additional specifications on the included elements, such as controlled vocabularies. Some examples of APs in use follow: Application profile CONABIO Application profile INBIO Application profile GBIF.ES Application profile Banco de Datos de la Naturaleza.Spain Application profile SIB-COLOMBIA == Relation to other standards == Plinian incorporates a number of elements already defined by other standards. The following table summarizes these standards and the elements used in Plinian Core:
Contrastive Language-Image Pre-training
Contrastive Language-Image Pre-training (CLIP) is a technique for training a pair of neural network models, one for image understanding and one for text understanding, using a contrastive objective. This method has enabled broad applications across multiple domains, including cross-modal retrieval, text-to-image generation, and aesthetic ranking. == Algorithm == The CLIP method trains a pair of models contrastively. One model takes in a piece of text as input and outputs a single vector representing its semantic content. The other model takes in an image and similarly outputs a single vector representing its visual content. The models are trained so that the vectors corresponding to semantically similar text-image pairs are close together in the shared vector space, while those corresponding to dissimilar pairs are far apart. To train a pair of CLIP models, one would start by preparing a large dataset of image-caption pairs. During training, the models are presented with batches of N {\displaystyle N} image-caption pairs. Let the outputs from the text and image models be respectively v 1 , . . . , v N , w 1 , . . . , w N {\displaystyle v_{1},...,v_{N},w_{1},...,w_{N}} . Two vectors are considered "similar" if their dot product is large. The loss incurred on this batch is the multi-class N-pair loss, which is a symmetric cross-entropy loss over similarity scores: − 1 N ∑ i ln e v i ⋅ w i / T ∑ j e v i ⋅ w j / T − 1 N ∑ j ln e v j ⋅ w j / T ∑ i e v i ⋅ w j / T {\displaystyle -{\frac {1}{N}}\sum _{i}\ln {\frac {e^{v_{i}\cdot w_{i}/T}}{\sum _{j}e^{v_{i}\cdot w_{j}/T}}}-{\frac {1}{N}}\sum _{j}\ln {\frac {e^{v_{j}\cdot w_{j}/T}}{\sum _{i}e^{v_{i}\cdot w_{j}/T}}}} In essence, this loss function encourages the dot product between matching image and text vectors ( v i ⋅ w i {\displaystyle v_{i}\cdot w_{i}} ) to be high, while discouraging high dot products between non-matching pairs. The parameter T > 0 {\displaystyle T>0} is the temperature, which is parameterized in the original CLIP model as T = e − τ {\displaystyle T=e^{-\tau }} where τ ∈ R {\displaystyle \tau \in \mathbb {R} } is a learned parameter. Other loss functions are possible. For example, Sigmoid CLIP (SigLIP) proposes the following loss function: L = 1 N ∑ i , j ∈ 1 : N f ( ( 2 δ i , j − 1 ) ( e τ w i ⋅ v j + b ) ) {\displaystyle L={\frac {1}{N}}\sum _{i,j\in 1:N}f((2\delta _{i,j}-1)(e^{\tau }w_{i}\cdot v_{j}+b))} where f ( x ) = ln ( 1 + e − x ) {\displaystyle f(x)=\ln(1+e^{-x})} is the negative log sigmoid loss, and the Dirac delta symbol δ i , j {\displaystyle \delta _{i,j}} is 1 if i = j {\displaystyle i=j} else 0. == CLIP models == While the original model was developed by OpenAI, subsequent models have been trained by other organizations as well. === Image model === The image encoding models used in CLIP are typically vision transformers (ViT). The naming convention for these models often reflects the specific ViT architecture used. For instance, "ViT-L/14" means a "vision transformer large" (compared to other models in the same series) with a patch size of 14, meaning that the image is divided into 14-by-14 pixel patches before being processed by the transformer. The size indicator ranges from B, L, H, G (base, large, huge, giant), in that order. Other than ViT, the image model is typically a convolutional neural network, such as ResNet (in the original series by OpenAI), or ConvNeXt (in the OpenCLIP model series by LAION). Since the output vectors of the image model and the text model must have exactly the same length, both the image model and the text model have fixed-length vector outputs, which in the original report is called "embedding dimension". For example, in the original OpenAI model, the ResNet models have embedding dimensions ranging from 512 to 1024, and for the ViTs, from 512 to 768. Its implementation of ViT was the same as the original one, with one modification: after position embeddings are added to the initial patch embeddings, there is a LayerNorm. Its implementation of ResNet was the same as the original one, with 3 modifications: In the start of the CNN (the "stem"), they used three stacked 3x3 convolutions instead of a single 7x7 convolution, as suggested by. There is an average pooling of stride 2 at the start of each downsampling convolutional layer (they called it rect-2 blur pooling according to the terminology of ). This has the effect of blurring images before downsampling, for antialiasing. The final convolutional layer is followed by a multiheaded attention pooling. ALIGN a model with similar capabilities, trained by researchers from Google used EfficientNet, a kind of convolutional neural network. === Text model === The text encoding models used in CLIP are typically Transformers. In the original OpenAI report, they reported using a Transformer (63M-parameter, 12-layer, 512-wide, 8 attention heads) with lower-cased byte pair encoding (BPE) with 49152 vocabulary size. Context length was capped at 76 for efficiency. Like GPT, it was decoder-only, with only causally-masked self-attention. Its architecture is the same as GPT-2. Like BERT, the text sequence is bracketed by two special tokens [SOS] and [EOS] ("start of sequence" and "end of sequence"). Take the activations of the highest layer of the transformer on the [EOS], apply LayerNorm, then a final linear map. This is the text encoding of the input sequence. The final linear map has output dimension equal to the embedding dimension of whatever image encoder it is paired with. These models all had context length 77 and vocabulary size 49408. ALIGN used BERT of various sizes. == Dataset == === WebImageText === The CLIP models released by OpenAI were trained on a dataset called "WebImageText" (WIT) containing 400 million pairs of images and their corresponding captions scraped from the internet. The total number of words in this dataset is similar in scale to the WebText dataset used for training GPT-2, which contains about 40 gigabytes of text data. The dataset contains 500,000 text-queries, with up to 20,000 (image, text) pairs per query. The text-queries were generated by starting with all words occurring at least 100 times in English Wikipedia, then extended by bigrams with high mutual information, names of all Wikipedia articles above a certain search volume, and WordNet synsets. The dataset is private and has not been released to the public, and there is no further information on it. ==== Data preprocessing ==== For the CLIP image models, the input images are preprocessed by first dividing each of the R, G, B values of an image by the maximum possible value, so that these values fall between 0 and 1, then subtracting by [0.48145466, 0.4578275, 0.40821073], and dividing by [0.26862954, 0.26130258, 0.27577711]. The rationale was that these are the mean and standard deviations of the images in the WebImageText dataset, so this preprocessing step roughly whitens the image tensor. These numbers slightly differ from the standard preprocessing for ImageNet, which uses [0.485, 0.456, 0.406] and [0.229, 0.224, 0.225]. If the input image does not have the same resolution as the native resolution (224×224 for all except ViT-L/14@336px, which has 336×336 resolution), then the input image is first scaled by bicubic interpolation, so that its shorter side is the same as the native resolution, then the central square of the image is cropped out. === Others === ALIGN used over one billion image-text pairs, obtained by extracting images and their alt-tags from online crawling. The method was described as similar to how the Conceptual Captions dataset was constructed, but instead of complex filtering, they only applied a frequency-based filtering. Later models trained by other organizations had published datasets. For example, LAION trained OpenCLIP with published datasets LAION-400M, LAION-2B, and DataComp-1B. == Training == In the original OpenAI CLIP report, they reported training 5 ResNet and 3 ViT (ViT-B/32, ViT-B/16, ViT-L/14). Each was trained for 32 epochs. The largest ResNet model took 18 days to train on 592 V100 GPUs. The largest ViT model took 12 days on 256 V100 GPUs. All ViT models were trained on 224×224 image resolution. The ViT-L/14 was then boosted to 336×336 resolution by FixRes, resulting in a model. They found this was the best-performing model. In the OpenCLIP series, the ViT-L/14 model was trained on 384 A100 GPUs on the LAION-2B dataset, for 160 epochs for a total of 32B samples seen. == Applications == === Cross-modal retrieval === CLIP's cross-modal retrieval enables the alignment of visual and textual data in a shared latent space, allowing users to retrieve images based on text descriptions and vice versa, without the need for explicit image annotations. In text-to-image retrieval, users input descriptive text, and CLIP retrieves images with matching embeddings. In image-to-text retrieval, images are used to find related text content. CLIP’s ability to connect vis
Tim Houlne
Tim Houlne is an American business executive, entrepreneur, and author known for his work in outsourcing and homeshoring, remote working, and artificial intelligence (AI) in customer service. He is the founder and CEO of Humach, a company that uses human agents and AI in customer experience solutions. Previously, he was co-founder and CEO of Working Solutions, a virtual contact center company in the United States. == Early life and education == Houlne graduated from Missouri Western State University (MWSU) in 1986 with a bachelor's degree in business administration and from the University of Texas in Dallas with an MBA. In 2024, MWSU and North Central Missouri College renamed the Convergent Technology Alliance Center to the Houlne Center for Convergent Technology. The 20,000 square-foot learning laboratory provides training and applied education experiences in industries such as AI, cybersecurity, manufacturing and construction, and service technologies. == Career == In 1998, Houlne co-founded Working Solutions, a Plano, Texas-based U.S. outsourcing company that provides customer service using remote, home-based agents. As CEO, he oversaw the development of a virtual workforce model that routes service calls to either domestic or offshore agents, according to client needs and service requirements. In 2015, Houlne founded Humach, a customer experience outsourcing provider that uses human service agents with AI-based digital agents. The company derives its name from the combination of services provided by humans and machines. Its clients include Amazon, Carfax and McDonald's. The company acquired InfiniteAI in 2020, and Markets EQ in 2025. In 2013, Houlne was named a finalist for the Ernst & Young Entrepreneur of the Year Award (Southwest Region).He is the co-author of several books focused on the evolution of work, the gig economy, and the influence of AI in customer-facing roles. == Works == The New World of Work: From the Cube to the Cloud (2013) ISBN 0982562276 OCLC 813933360 The New World of Work, Second Edition: The Cube, the Cloud and What's Next (2023) ISBN 9781642258318 OCLC 1389815847 The Intelligent Workforce: How Humans & Machines Will Co-Create a Better Future (2024) ISBN 9798887501604 OCLC 1439598569