{
  "id": "zero-shot-learning",
  "title": "Zero-shot learning",
  "lang": "en",
  "source_lang": "en",
  "summary": "Zero-shot learning enables models to recognize unseen classes or perform tasks without labeled examples for those targets, using transferable representations and semantic information.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "Zero-shot machine learning",
    "ZSL"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/zero-shot-learning/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/zero-shot-learning/",
      "json": "https://aiwiki.page/zh/zero-shot-learning.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Zero-shot learning is an approach to machine learning in which a system recognizes categories or performs tasks for which it has received no task-specific labeled examples. Instead of learning each target directly from examples, the system transfers knowledge through descriptions, attributes, linguistic representations, or other relationships between familiar and unfamiliar targets. A visual classifier, for example, may recognize a zebra without having trained on labeled zebra photographs if it can connect the description “a horse-like animal with black-and-white stripes” to visual properties learned elsewhere.\n\nThe term has two closely related uses. In classical zero-shot classification, training and evaluation categories are explicitly separated, and semantic information connects the two. In contemporary large language models and multimodal systems, “zero-shot” commonly means performing a downstream task from an instruction or description without demonstrations or task-specific parameter updates. Neither usage means that the model has learned without data: substantial prior training, supervision, or human-provided knowledge usually supplies the capabilities being transferred.\n\nZero-shot learning matters because exhaustive annotation is impractical when categories are numerous, rare, changing, or defined only after a model has been trained. Its central questions concern what knowledge can transfer, how a new target is specified, and how to establish that apparent success does not depend on hidden exposure to evaluation examples."
    },
    {
      "heading": "Historical development and changing scope",
      "text": "Zero-shot learning developed within transfer learning, but it made an unusually demanding departure from standard supervised learning. A conventional classifier learns associations between inputs and labels represented in its training set. Early zero-shot research asked whether labels could instead be identified through descriptions that connect them to already learned properties. This framing shifted attention from memorizing category boundaries to representing relationships among categories.\n\nHugo Larochelle, Dumitru Erhan, and Yoshua Bengio explored zero-data learning using descriptions of tasks in 2008. At the 2009 IEEE Conference on Computer Vision and Pattern Recognition, Christoph Lampert, Hannes Nickisch, and Stefan Harmeling presented “Learning to Detect Unseen Object Classes by Between-Class Attribute Transfer.” Their work helped establish attribute-based recognition as a practical research program. Another influential 2009 contribution, Mark Palatucci and colleagues’ “Zero-Shot Learning with Semantic Output Codes,” used semantic representations to predict concepts without training examples for every target concept.\n\nThese approaches made class descriptions operational. Rather than treating “zebra” as an arbitrary numerical identifier, an attribute-based system represented the category through properties such as striped, quadrupedal, and herbivorous. Some properties are visually observable; others are semantic facts that only indirectly predict appearance. Shared attributes allowed examples of known animals to contribute evidence relevant to an unfamiliar one. This was more structured than simply applying a classifier trained on one dataset to another.\n\nDuring the early 2010s, learned semantic spaces became increasingly important. Word embeddings made it possible to derive class representations from text rather than manually specify every attribute. The 2013 DeViSE system, developed by Andrea Frome and colleagues, connected deep visual representations with a semantic space derived from language. Work on label embeddings and compatibility functions similarly treated recognition as matching an image to a semantic description. The growth of deep learning improved visual features, while also creating evaluation problems when feature extractors had already encountered supposedly unseen classes.\n\nThe field broadened further with large-scale language and image–text pretraining. GPT-3’s 2020 evaluation distinguished zero-shot, one-shot, and few-shot prompting, while CLIP, introduced in 2021, demonstrated broad image classification through natural-language class descriptions. These developments did not eliminate the classical definition, but they made the evaluation unit more varied: a class, a dataset, a language pair, or an entire task could be described as unseen. Consequently, historical comparisons require attention to what was withheld, what pretraining contained, and what information was supplied at evaluation time."
    },
    {
      "heading": "Formal problem and information available to the learner",
      "text": "In classical zero-shot classification, the labeled training data contain examples from a set of seen classes, \\(\\mathcal{Y}_s\\). Evaluation involves a disjoint set of unseen classes, \\(\\mathcal{Y}_u\\), with\n\n\\[\\mathcal{Y}_s \\cap \\mathcal{Y}_u = \\varnothing.\\]\n\nEach class \\(y\\) has an auxiliary representation \\(a(y)\\), such as an attribute vector, textual description, or semantic embedding. The training set can be written as\n\n\\[\\mathcal{D}_s=\\{(x_i,y_i)\\}_{i=1}^{N}, \\qquad y_i\\in\\mathcal{Y}_s.\\]\n\nThe learner uses \\(\\mathcal{D}_s\\) and the auxiliary information to construct predictions for classes lacking labeled training examples. Without some relationship between seen and unseen targets, identifying an arbitrary new label is generally impossible: the label itself contains no evidence about which inputs belong to it.\n\nA common decision rule assigns an input to the candidate class with the highest compatibility score:\n\n\\[\\hat y=\\underset{y\\in\\mathcal{Y}_{\\mathrm{candidate}}}{\\arg\\max} F_\\theta(x,a(y)).\\]\n\nHere, \\(F_\\theta\\) is a learned scoring function. In conventional zero-shot evaluation, the candidate set contains only unseen classes. In generalized zero-shot learning, it contains both seen and unseen classes. This distinction is consequential: a model may separate several unfamiliar animals successfully when told that every image belongs to an unfamiliar category, yet assign nearly all of them to familiar animals when those familiar labels are also available.\n\nZero-shot learning also differs from few-shot learning, which provides a small number of labeled examples for each target, and from unsupervised learning, which does not necessarily supply semantic identities for discovered groups. Domain adaptation often studies changes in input distribution while preserving a shared label set; classical zero-shot learning changes the label set itself. These problems can overlap, but their assumptions are not interchangeable.\n\nAnother distinction concerns access to unlabeled target examples. Inductive zero-shot learning does not use the target evaluation collection during fitting. Transductive zero-shot learning may exploit unlabeled examples from the target distribution, for instance to refine clusters or align feature distributions. Such access can improve performance, but it introduces information unavailable in a strictly inductive setting and must be reported separately.\n\nFor language models, the boundary is often defined at the task interface rather than by a mathematically disjoint class set. A sentiment classifier can be called zero-shot if it receives an instruction and label descriptions but no labeled sentiment examples in its prompt and no task-specific fine-tuning. Its pretraining may nevertheless contain reviews, sentiment labels, or similar instructions. Thus, zero-shot describes a learning or evaluation protocol, not an intrinsic property of a model independent of its data history."
    },
    {
      "heading": "Semantic descriptions and the bridge between classes",
      "text": "The defining resource in classical zero-shot learning is a semantic bridge between examples and unfamiliar labels. A class identifier such as an integer provides no transferable meaning. By contrast, a vector encoding “has wings,” “has a beak,” and “lives in water” places a target in relation to properties that can be learned from other classes. The quality of this bridge determines whether the target is distinguishable and whether the relevant knowledge is available at all.\n\nHuman-defined attributes are one of the most interpretable bridges. An animal dataset may describe classes by coloration, body parts, habitat, and behavior. A bird dataset can use finer properties such as bill shape, wing color, or breast pattern. Attribute values may be binary, continuous, or expressed as estimated proportions. Class-level attributes describe typical members of a category, whereas instance-level attributes describe a particular image. Confusing these levels can obscure variation: a species-level color description does not imply that every body region is visible in every photograph.\n\nAttributes make assumptions explicit, but they require substantial annotation and can encode information that images do not reveal. “Nocturnal” or “eats insects” may help distinguish species conceptually while offering little direct visual evidence. Correlations learned from the seen classes can also fail on unseen ones. A model that associates aquatic habitat with a particular body shape may transfer poorly to an unfamiliar animal with a different morphology. An attribute inventory therefore acts as an inductive bias, privileging some relationships and ignoring others.\n\nTextual representations reduce the need for hand-built attribute lists. A class name can be represented using word2vec or another embedding method; longer descriptions can be encoded by a transformer. Text can express appearance, function, behavior, and context in one representation. However, linguistic similarity is not identical to perceptual similarity. Two animals may occur in similar sentences because they share a habitat while looking quite different. Polysemy also matters: a short label such as “crane” can refer to a bird or a machine unless the description supplies context.\n\nStructured resources provide another route. A knowledge graph can relate classes through taxonomic, functional, or part–whole connections, while WordNet supplies lexical and hierarchical relations used in visual recognition research. Such structure may propagate evidence from a parent category to a child or connect a new object to known components. Its usefulness depends on whether those relations predict the target observations. A taxonomic hierarchy is not automatically a map of visual resemblance.\n\nThe semantic description is therefore part of the supervision, even when no target images are labeled. Elaborate expert-written descriptions, externally curated attributes, and automatically generated text contribute different amounts and kinds of information. Comparisons between systems are meaningful only when they account for these resources rather than treating every absence of labeled target examples as an equivalent experimental condition."
    },
    {
      "heading": "Attribute prediction and compatibility-based methods",
      "text": "Early attribute-based systems decomposed recognition into reusable intermediate decisions. Direct attribute prediction learns detectors for properties from seen-class images, then combines their outputs according to each candidate class’s attribute signature. If the model predicts that an image depicts something striped, four-legged, and horse-shaped, a matching class description can receive a high score even though no labeled example of that class was used in fitting.\n\nA simplified probabilistic implementation combines predicted attribute probabilities across a candidate’s required properties. Such multiplication is convenient but often assumes conditional independence that does not hold: “has wings,” “has feathers,” and “can fly” are correlated. It can also amplify errors when several uncertain predictions are multiplied. Indirect attribute prediction instead estimates seen-class probabilities and transfers them through relationships between seen classes, attributes, and unseen targets. Both families illustrate how an intermediate semantic vocabulary can replace direct target-specific classifiers.\n\nCompatibility-based methods learn a shared relationship more directly. A common bilinear form is\n\n\\[F(x,y)=\\phi(x)^{\\mathsf T}W a(y),\\]\n\nwhere \\(\\phi(x)\\) is an input representation, \\(a(y)\\) is a class representation, and \\(W\\) is learned from seen classes. The input representation may come from a convolutional neural network or another feature extractor. Learning \\(W\\) aligns visual dimensions with semantic dimensions, allowing the same scoring rule to evaluate new class vectors at inference time.\n\nTraining may use a ranking loss function that rewards the correct class for scoring above an incorrect class:\n\n\\[L_i=\\max\\left(0,\\, m-F(x_i,y_i)+F(x_i,y^-)\\right),\\]\n\nwith margin \\(m\\) and a negative label \\(y^-\\). Alternative objectives include classification losses, regression to semantic prototypes, and structured prediction losses. Models such as ALE, DEVISE, and SJE explored related compatibility-learning approaches with different representations and optimization choices. Their shared principle is that training shapes a transferable matching function rather than independently estimating every future classifier.\n\nThe direction of projection matters. A system may map images into a semantic space, class descriptions into a visual space, or both into a common latent space. These choices alter neighborhood structure and the consequences of prediction error. Projecting many diverse images toward a small number of class prototypes can compress variation, while projecting descriptions into visual space requires extrapolating the appearance of classes not represented by training examples.\n\nNonlinear mappings increase expressive capacity but can also fit accidental relationships among seen classes. Regularization, careful class-held-out validation, and suitable feature representations are therefore central. Improving training accuracy on seen categories does not necessarily improve zero-shot transfer. The relevant objective is to learn relationships that remain valid when class identity changes, rather than merely reproduce the distinctions present in the training partition."
    },
    {
      "heading": "Generative approaches and synthetic training examples",
      "text": "Generative zero-shot learning addresses unfamiliar classes by constructing representations of examples that could belong to them. Instead of using semantic descriptions only to score a test input, a conditional generator learns how visual features vary with class semantics. Given an unseen class description, it produces synthetic features, which can then serve as training data for an ordinary classifier. The target still has no real labeled training examples under the zero-shot protocol; its synthetic examples derive from transferred knowledge.\n\nA typical formulation learns a generator\n\n\\[\\tilde h=G_\\psi(z,a(y)),\\]\n\nwhere \\(z\\) is a random latent variable and \\(\\tilde h\\) is a synthesized visual feature. Conditioning on \\(a(y)\\) controls class identity, while \\(z\\) is intended to represent within-class variation such as pose, viewpoint, or background. At deployment preparation time, the generator receives descriptions of unseen classes and produces a feature collection for each. A classifier can then be trained on these features, sometimes together with real features from seen classes.\n\nGenerative adversarial networks became important in this line of work. The 2018 f-CLSWGAN method combined adversarial feature generation with a classification objective intended to preserve discriminative information. Other approaches use a variational autoencoder, combine adversarial and variational objectives, or add reconstruction constraints that require generated features to recover the conditioning semantics. These objectives address different weaknesses: realism alone does not guarantee correct class identity, and semantic consistency alone does not ensure realistic feature distributions.\n\nGenerating features rather than full images has practical advantages. A pretrained visual encoder may already remove much irrelevant pixel-level variation, leaving a lower-dimensional representation easier to model. Feature generation also avoids the requirement that synthetic images look convincing to humans. Nevertheless, it inherits the encoder’s limitations: a feature space that discards a distinguishing property cannot recover it merely because a semantic description mentions it. Pixel generation and feature generation therefore solve related but different problems.\n\nGenerative methods are particularly relevant to generalized zero-shot learning because they can place seen and unseen classes in one conventional training procedure. Synthetic examples give unseen classes an explicit role in classifier fitting rather than leaving them as descriptions compared against classifiers optimized only for seen classes. The balance between real and synthetic samples, however, changes decision boundaries and can create artificial class priors. Large numbers of synthetic features do not constitute independent observations of the real target distribution.\n\nThe principal uncertainty is whether the conditional generator extrapolates correctly. It learns the association between descriptions and appearances from seen categories, and that association may fail for a novel combination of attributes. Diffusion models offer another conditional generative mechanism, but do not remove this limitation. A plausible synthetic target can still omit rare appearances, reproduce contextual biases, or collapse distinctions between neighboring classes. Evaluation must ultimately use real, independently held-out target examples rather than the generator’s own outputs."
    },
    {
      "heading": "Language models, prompting, and vision–language systems",
      "text": "In natural language processing, zero-shot operation commonly means specifying a task through language without providing demonstrations. A model might receive “Classify this review as positive or negative” followed by the review. It then generates a label or assigns scores to candidate label strings. This use differs from classical unseen-class recognition because the words “positive” and “negative,” the task structure, and even similar examples may all have appeared during broad pretraining.\n\nThe mechanism depends on capabilities acquired before the task is presented. A language model learns statistical relationships in text; instruction tuning can additionally train it to map task descriptions to suitable responses. Consequently, a model can be zero-shot on a particular evaluation dataset while having received extensive supervision on related tasks. Instruction tuning, alignment training, and task-specific adaptation are separate axes from the number of examples supplied in the current prompt.\n\nZero-shot prompting contrasts with in-context learning in its demonstration-based form, where examples appear inside the input without changing model parameters. A few-shot prompt might show several reviews and their labels before asking for a new classification. Zero-shot prompts omit these demonstrations, but may still contain a detailed rubric, definitions, constraints, or a requested reasoning procedure. The 2022 work “Large Language Models are Zero-Shot Reasoners” showed that a simple reasoning instruction could improve results on several reasoning benchmarks, illustrating that the absence of demonstrations does not imply the absence of influential task guidance.\n\nVision–language models connect this paradigm with visual recognition. CLIP trained image and text encoders using contrastive learning on 400 million image–text pairs. For classification, an image embedding is compared with embeddings of textual prompts such as “a photo of a dog.” With normalized image vector \\(v(x)\\) and text vector \\(t(y)\\), the score can be expressed as\n\n\\[s(x,y)=\\frac{v(x)^{\\mathsf T}t(y)}{\\tau},\\]\n\nwhere \\(\\tau\\) controls score scaling. Candidate classes can be changed by changing the text rather than retraining a final classification layer.\n\nThis flexibility supports open-vocabulary object detection and other recognition systems in which target names are not fixed during task-specific training. It does not mean that every possible word or concept is equally well grounded. Recognition depends on the coverage and quality of pretraining pairs, the distinction between an entire image and a localized region, and the match between prompt wording and the intended visual concept.\n\nPrompt selection is itself an experimental choice. Label wording, articles, templates, descriptions, and combinations of multiple prompts can alter results. If a template is chosen after examining target labels and examples, the resulting protocol uses more target information than a fully predetermined prompt. Extensive target-specific prompt engineering therefore needs to be distinguished from a single fixed zero-shot evaluation."
    },
    {
      "heading": "Benchmarks, metrics, and experimental validity",
      "text": "Classical benchmarks test whether semantic transfer works across deliberately separated categories. Animals with Attributes 2 contains 50 animal classes and 85 attributes, with a commonly used split of 40 seen and 10 unseen classes. Caltech–UCSD Birds-200-2011 contains 200 bird species and 312 attributes, with widely used zero-shot splits of 150 seen and 50 unseen classes. The SUN Attribute dataset includes 717 scene categories and 102 attributes, providing a different recognition problem from object- or species-centered classification.\n\nThese datasets test different kinds of transfer. Coarse animal categories often differ in conspicuous body form, while bird species can require subtle local features. Scene recognition may depend on spatial arrangement, objects, and human activity rather than a single foreground entity. Performance on one benchmark therefore does not establish a general ability to recognize arbitrary unseen concepts. Dataset size, semantic-description quality, visual similarity, and the relationship to pretraining data all affect difficulty.\n\nA major evaluation issue is exposure through feature extraction. A class can be absent from the zero-shot training partition yet present in the data used to pretrain its visual encoder. Yongqin Xian and colleagues’ 2017 “Zero-Shot Learning—The Good, the Bad and the Ugly” emphasized this problem and introduced proposed splits designed to avoid overlap between evaluation classes and the 1,000 classes used in standard ImageNet pretraining. This does not eliminate every possible source of prior exposure, but it clarifies the assumptions of the benchmark.\n\nPer-class accuracy is common because class frequencies can differ. If \\(\\mathcal{D}_c\\) contains evaluation examples from class \\(c\\), macro-averaged accuracy is\n\n\\[A=\\frac{1}{|\\mathcal{C}|} \\sum_{c\\in\\mathcal{C}} \\frac{\\text{correct predictions in }\\mathcal{D}_c}{|\\mathcal{D}_c|}.\\]\n\nFor generalized evaluation, performance is commonly reported separately on seen classes, \\(S\\), and unseen classes, \\(U\\), together with their harmonic mean:\n\n\\[H=\\frac{2SU}{S+U}.\\]\n\nA high score on only one group cannot produce a high harmonic mean, making this measure useful for revealing severe seen-class bias.\n\nModel selection must also respect the information boundary. Choosing hyperparameters using labeled unseen test examples leaks target supervision into development. A typical alternative reserves some otherwise seen classes as pseudo-unseen validation categories. The final unseen test set then remains untouched until evaluation. Transductive access to unlabeled target inputs, use of target class frequencies, and external descriptions should be documented independently.\n\nFor foundation models, benchmark data contamination is harder to exclude because pretraining collections may contain benchmark images, questions, answers, or derivatives. Exact duplicates are only one issue; near-duplicates and task-specific explanatory material can also matter. Zero-shot reports consequently require precise statements about the downstream adaptation protocol and the known training-data boundary, rather than equating “no examples in the prompt” with “no prior exposure.”"
    },
    {
      "heading": "Applications and operational constraints",
      "text": "Zero-shot learning is useful where the label inventory is too large or changes too quickly for repeated annotation. In image classification, a system can receive new category descriptions without training a separate output layer for each one. A museum collection, for example, may need categories defined by object use, historical terminology, or material rather than the labels used to train a generic vision model. Text-defined classes offer flexibility, although success depends on whether the representation captures those distinctions.\n\nIn information retrieval, shared image–text representations let a textual query retrieve visually relevant items without a task-specific labeled query–image dataset. The same embedding space can support changing search descriptions, compositional queries, and similarity-based organization. Retrieval and classification remain distinct: returning several plausible images is not equivalent to making one calibrated category decision. A retrieval system can be useful even when its top-ranked item would not satisfy a strict classification benchmark.\n\nText classification provides another application. Topic routing, intent identification, and document categorization can be formulated using candidate labels and their descriptions. A language model or language-inference model may compare a document with hypotheses such as “This request concerns a delivery delay.” This can accommodate a new taxonomy with little labeled data, but closely related labels may depend heavily on wording. A broad label such as “billing” and a narrow label such as “refund for duplicate charge” also occupy different levels of specificity, complicating direct score comparison.\n\nIn machine translation, zero-shot translation commonly denotes translation between a language pair for which no direct parallel training examples were supplied, even though each language may have participated in training with other languages. Google’s 2016 multilingual neural translation work demonstrated this possibility in a shared multilingual system. The transfer problem differs from unseen-class classification: the model must generate structured sequences, and knowledge is transferred through shared representations and training across language directions.\n\nRobotics and multimodal interfaces can use natural-language goals to refer to unfamiliar object categories or combinations of familiar actions. Here, recognizing the words or objects is only one component. Grounding a description in a physical scene, identifying an appropriate action, and respecting mechanical constraints are separate requirements. A zero-shot language interpretation result does not by itself establish reliable zero-shot physical task execution.\n\nOperational costs also vary. A fixed set of class text embeddings can be precomputed, while encoding every new candidate description at request time adds latency. Comparing an input against a very large candidate vocabulary requires efficient search or hierarchical filtering. Expanding the vocabulary can reduce accuracy by adding semantically overlapping alternatives, even though the underlying encoder is unchanged. Thus, the ease of introducing labels should not be confused with constant accuracy, constant inference cost, or automatic suitability for consequential decisions."
    },
    {
      "heading": "Failure modes, calibration, and unresolved questions",
      "text": "Zero-shot learning is limited by the information available for transfer. If two unseen classes have identical semantic descriptions and no other distinguishing information, a description-based system has no principled basis for separating them. If a target depends on a visual property never learned by the encoder, naming that property may not be sufficient. Successful transfer therefore reflects both a useful representation of the input and a semantic description that identifies the relevant distinction.\n\nOne technical difficulty is hubness in high-dimensional embedding spaces. A small number of class prototypes can become nearest neighbors of disproportionately many inputs, attracting predictions regardless of the true class. Projection choices and the geometry of semantic spaces can contribute to this behavior. Another difficulty is projection shift: a mapping learned from seen classes may not preserve the correct relationships for unseen ones. These failures show why a plausible semantic space does not guarantee a reliable visual decision rule.\n\nGeneralized zero-shot learning adds a systematic preference for seen classes. Their parameters or representations have been optimized using real labeled examples, whereas unseen classes are represented indirectly. One correction subtracts a constant from seen-class scores:\n\n\\[F'(x,y)=F(x,y)-\\gamma\\,\\mathbf{1}[y\\in\\mathcal{Y}_s].\\]\n\nVarying \\(\\gamma\\) trades seen accuracy against unseen accuracy. Such adjustment can improve balance, but the appropriate value depends on the deployment distribution and validation assumptions. It is not a universal correction for all forms of mismatch.\n\nProbability calibration is a separate issue from ranking accuracy. Similarity scores converted through a softmax become normalized numbers, but they need not represent reliable probabilities of correctness. Moreover, a classifier that must choose from supplied labels may confidently select one when none is appropriate. Zero-shot recognition is therefore not equivalent to anomaly detection or reliable unknown-class rejection. Changing the candidate list can change normalized confidence even when the image and its embedding remain fixed.\n\nBroader distribution shift exposes weaknesses that class-held-out tests may miss. A model can recognize an unseen category in familiar photographs yet fail on sketches, unusual lighting, different geographic contexts, or atypical examples. Text and image–text pretraining also inherit uneven representation of languages, objects, and social contexts. In generative systems, these mismatches may produce fluent but unsupported responses rather than an explicit classification error, connecting zero-shot task performance with problems of hallucination and factual grounding.\n\nAn unresolved conceptual issue is how to separate genuine compositional transfer from broad prior coverage. With very large pretraining datasets, success on a nominally new task may reflect a combination of reusable reasoning, recognition of familiar patterns, and partial exposure to related material. Evaluations that hold out combinations, relations, domains, or newly created tasks test different aspects of this problem. Establishing what transferred requires specifying the withheld information at a finer level than the single label “zero-shot.”"
    }
  ],
  "links": [
    "machine-learning",
    "large-language-model",
    "transfer-learning",
    "supervised-learning",
    "word-embedding",
    "deep-learning",
    "clip-model",
    "unsupervised-learning",
    "domain-adaptation",
    "fine-tuning",
    "inductive-bias",
    "word2vec",
    "transformer-architecture",
    "knowledge-graph",
    "wordnet",
    "convolutional-neural-network",
    "loss-function",
    "latent-space",
    "regularization",
    "generative-adversarial-network",
    "variational-autoencoder",
    "diffusion-model",
    "natural-language-processing",
    "language-model",
    "in-context-learning",
    "contrastive-learning",
    "open-vocabulary-object-detection",
    "prompt-engineering",
    "imagenet",
    "hyperparameter",
    "test-set",
    "benchmark-data-contamination",
    "information-retrieval",
    "machine-translation",
    "probability-calibration",
    "anomaly-detection",
    "distribution-shift",
    "ai-hallucination"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}