{
  "id": "conditional-random-field",
  "title": "Conditional random field",
  "lang": "en",
  "source_lang": "en",
  "summary": "A conditional random field is a probabilistic model that predicts interdependent outputs by defining their joint distribution conditional on observed inputs.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "CRF",
    "conditional Markov random field"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/conditional-random-field/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/conditional-random-field/",
      "json": "https://aiwiki.page/zh/conditional-random-field.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A conditional random field (CRF) is a probabilistic model for predicting a collection of related outputs from observed inputs. Rather than classify each output independently, it assigns a conditional probability distribution to complete output configurations. For example, a model can label every word in a sentence while accounting for dependencies between neighboring labels, or assign categories to image pixels while encouraging compatible spatial arrangements. CRFs belong to machine learning and are particularly associated with structured prediction, in which the prediction is a sequence, segmentation, graph, or other composite object.\n\nIntroduced by John Lafferty, Andrew McCallum, and Fernando Pereira in 2001, CRFs combine the dependency structure of undirected probabilistic models with conditional, or discriminative, learning. They became important in natural language processing, computer vision, and other fields where local observations must be interpreted consistently across a larger structure. Their defining features are conditional modeling, factorization over an output graph, and normalization over complete configurations. These features remain applicable whether the model uses hand-designed features or scores produced by a neural network."
    },
    {
      "heading": "Conditional modeling and graphical structure",
      "text": "A CRF describes \\(p_\\theta(\\mathbf{y}\\mid\\mathbf{x})\\), where \\(\\mathbf{x}\\) is an observed input, \\(\\mathbf{y}\\) is a collection of output random variables, and \\(\\theta\\) denotes model parameters. The input might be a sentence, an image, or a sequence of sensor readings. The output might contain linguistic tags, region categories, or activity labels. Unlike a model of the joint distribution \\(p(\\mathbf{x},\\mathbf{y})\\), a CRF need not specify how the observations themselves were generated. Its probability model concerns possible outputs for a given observation.\n\nFormally, let \\(G=(V,E)\\) be an undirected graph whose vertices index the output variables. Conditional on \\(\\mathbf{x}\\), the outputs satisfy the Markov property associated with that graph: each variable is conditionally independent of nonneighboring output variables given its neighbors. For a strictly positive distribution, this property permits a factorization over graph cliques:\n\n\\[p_\\theta(\\mathbf{y}\\mid\\mathbf{x}) = \\frac{1}{Z_\\theta(\\mathbf{x})} \\prod_{c\\in\\mathcal C} \\psi_{\\theta,c}(\\mathbf{y}_c,\\mathbf{x}),\\]\n\nwhere \\(\\mathcal C\\) is a collection of cliques, \\(\\psi_{\\theta,c}\\) is a positive potential, and \\(Z_\\theta(\\mathbf{x})\\) normalizes the distribution. A clique is a set of vertices that are all connected to one another. The output graph expresses conditional independence among labels, not a restriction that each potential inspect only a small part of the input.\n\nThe normalizing quantity, called the partition function, is\n\n\\[Z_\\theta(\\mathbf{x}) = \\sum_{\\mathbf{y}'\\in\\mathcal Y(\\mathbf{x})} \\prod_{c\\in\\mathcal C} \\psi_{\\theta,c}(\\mathbf{y}'_c,\\mathbf{x}),\\]\n\nfor a discrete output space \\(\\mathcal Y(\\mathbf{x})\\). It makes the probabilities of all valid output configurations sum to one. Individual potentials are compatibility functions rather than normalized probabilities: a large value indicates that a particular local assignment is favored in the context of the observed input.\n\nA CRF is therefore a conditional form of a Markov random field. It can also be represented as a factor graph, which explicitly separates variables from the functions that score groups of them. This representation is especially useful when factors connect more than two output variables. An undirected edge does not imply a causal relationship, and the graph need not match a physical network. Its role is to specify which output interactions the model represents directly and, consequently, which computational methods may be available for learning and prediction."
    },
    {
      "heading": "Log-linear potentials and feature design",
      "text": "The classical CRF uses log-linear potentials. If \\(\\mathbf{F}(\\mathbf{x},\\mathbf{y})\\) is a vector of features of the input and output together, its distribution can be written as\n\n\\[p_\\theta(\\mathbf{y}\\mid\\mathbf{x}) = \\frac{ \\exp\\!\\left(\\theta^\\top\\mathbf{F}(\\mathbf{x},\\mathbf{y})\\right) }{ Z_\\theta(\\mathbf{x}) }.\\]\n\nThe global feature vector usually decomposes into local contributions, so that \\(\\mathbf{F}=\\sum_c\\mathbf{f}_c\\). Each parameter weights a particular kind of compatibility. A positive weight increases the score of configurations exhibiting that feature, whereas a negative weight decreases it. The exponential transformation turns an unrestricted real-valued score into a positive potential.\n\nIn linguistic sequence labeling, a feature can combine an observation with a proposed label: whether the current token begins with a capital letter and has a person-name label, for example. Another feature can indicate that the previous label is a beginning-of-organization tag and the current label continues that organization. Observation features may inspect suffixes, neighboring words, punctuation, document position, or a domain-specific vocabulary. Because the input is conditioned on, overlapping observations do not require the model to assume that those observations are mutually independent.\n\nThis flexibility made feature engineering central to early CRF systems. A single token could activate word-identity features, character-pattern features, and dictionary-membership features simultaneously. Such features often produce very high-dimensional but sparse vectors: only a small subset is active for a particular position and label assignment. Feature templates provide a compact description of these families. For instance, one template might create a separate parameter for every combination of a current word and a candidate tag.\n\nDependencies on observations must be distinguished from dependencies among unknown labels. A factor involving \\(y_t\\) and \\(y_{t-1}\\) may inspect the entire input sentence without making the output graph more densely connected. By contrast, introducing a factor involving \\(y_t\\) and \\(y_{t+10}\\) changes the graph and can substantially increase inference costs. Rich input context is therefore not equivalent to a high-order output model.\n\nThe same mathematics also supports an energy-based model interpretation. Defining \\(E_\\theta(\\mathbf{x},\\mathbf{y})=-\\theta^\\top\\mathbf{F}(\\mathbf{x},\\mathbf{y})\\) gives probabilities proportional to \\(\\exp(-E_\\theta)\\). Low energy means high compatibility. More general CRFs replace the linear score with nonlinear parameterized functions. The probabilistic definition does not require handcrafted features, but important optimization guarantees of classical CRFs depend specifically on scores being linear in the learned parameters."
    },
    {
      "heading": "Linear-chain CRFs and sequence labeling",
      "text": "The best-known CRF has a linear-chain output graph. For an input sequence \\(\\mathbf{x}=(x_1,\\ldots,x_T)\\), it predicts labels \\(\\mathbf{y}=(y_1,\\ldots,y_T)\\), with direct interactions between adjacent labels. A common score is\n\n\\[S_\\theta(\\mathbf{x},\\mathbf{y}) = \\sum_{t=1}^{T}u_\\theta(t,y_t,\\mathbf{x}) + \\sum_{t=2}^{T}a_\\theta(t,y_{t-1},y_t,\\mathbf{x}),\\]\n\nwhere \\(u_\\theta\\) is a unary score and \\(a_\\theta\\) is a transition score. Start and end scores can be included separately. The conditional probability is \\(\\exp(S_\\theta)/Z_\\theta(\\mathbf{x})\\). In many implementations, transitions are entries in a learned matrix independent of the input; the more general definition allows input-dependent transitions.\n\nA chain CRF is useful when neighboring decisions constrain one another. In named-entity recognition, labels often use a beginning–inside–outside scheme, usually abbreviated BIO. The phrase “New York” might receive B-LOC followed by I-LOC, while unrelated words receive O. A model that predicts tags separately can produce an I-LOC tag after an incompatible person tag. A CRF can penalize such transitions or exclude them through a fixed validity constraint. It then selects labels jointly rather than repairing independent predictions afterward.\n\nConsider the sentence “Jordan joined Acme.” The first word could denote a person or a location in different contexts. A unary score can incorporate the word itself, capitalization, and the surrounding verb; transition scores can favor coherent entity boundaries. The CRF does not simply replace uncertain decisions with common label sequences. Its decision balances observation-dependent evidence against label interactions across the entire sentence. Strong input evidence can outweigh a generally preferred transition unless that transition has been prohibited.\n\nThe term “first-order” refers to the direct dependence on one preceding label, not to the size of the observable context. A first-order CRF can use a contextual representation informed by every word in a sentence. A second-order chain includes factors over triples of consecutive labels. Such a model can distinguish patterns that pairwise transitions cannot express directly, but its standard dynamic program costs \\(O(TK^3)\\) rather than \\(O(TK^2)\\), where \\(K\\) is the number of labels.\n\nSequences need not consist of words. The same structure can label characters for word segmentation, acoustic frames for sound categories, or time windows for activity recognition. What makes a chain appropriate is the assumed dependency structure among outputs. If the task requires nested spans, branching structures, or substantial nonlocal label interactions, a simple linear chain may be an inadequate representation even when its input encoder is highly expressive."
    },
    {
      "heading": "Exact inference and decoding",
      "text": "CRFs require several distinct computations. Partition-function computation sums the scores of all possible outputs after exponentiation. Marginal inference calculates probabilities of individual labels or groups of labels. Maximum a posteriori decoding finds the single highest-probability complete output:\n\n\\[\\hat{\\mathbf{y}} = \\arg\\max_{\\mathbf{y}} p_\\theta(\\mathbf{y}\\mid\\mathbf{x}) = \\arg\\max_{\\mathbf{y}} S_\\theta(\\mathbf{x},\\mathbf{y}).\\]\n\nThe partition function cancels from the last optimization because it is constant for a fixed input. Consequently, an algorithm that finds the best configuration need not also be able to compute its normalized probability efficiently.\n\nFor a linear chain, dynamic programming avoids explicitly enumerating \\(K^T\\) label sequences. Let \\(\\alpha_t(j)\\) be the summed unnormalized weight of all prefixes ending with label \\(j\\) at position \\(t\\). With unary score \\(u_t(j)\\) and transition score \\(a_t(i,j)\\), the forward recursion is\n\n\\[\\alpha_t(j) = \\exp(u_t(j)) \\sum_{i=1}^{K} \\alpha_{t-1}(i)\\exp(a_t(i,j)).\\]\n\nAppropriate start conditions initialize the recurrence, and summing the final forward values, including any end scores, gives \\(Z_\\theta(\\mathbf{x})\\). A backward recursion provides complementary suffix values. Combining the two produces single-position and adjacent-pair marginals; this is the forward–backward algorithm.\n\nDirectly multiplying and adding exponentials can underflow or overflow in floating-point arithmetic. Implementations therefore commonly operate in logarithmic space and use the log-sum-exp identity,\n\n\\[\\operatorname{LSE}(z_1,\\ldots,z_n) = m+\\log\\sum_i\\exp(z_i-m), \\qquad m=\\max_i z_i.\\]\n\nFor best-sequence decoding, the Viterbi algorithm replaces the summation in the forward recurrence with a maximum and stores backpointers. Following those pointers from the best final state reconstructs the optimal sequence.\n\nFor dense transitions, forward–backward and Viterbi both require \\(O(TK^2)\\) time. Computing only the partition function can use \\(O(K)\\) working memory beyond stored scores, whereas ordinary backpointer-based decoding uses \\(O(TK)\\) memory. Sparse transition structures may reduce the arithmetic required. Tree-structured CRFs also permit exact message passing, although the cost depends on label cardinalities and factor sizes. General graph complexity depends strongly on graph structure, particularly treewidth.\n\nMarginal decoding and best-sequence decoding answer different questions. Choosing the most probable label separately at every position minimizes expected tokenwise error under an unconstrained Hamming loss. Viterbi instead finds the most probable complete sequence. Marginal choices need not form the highest-probability sequence and can violate structural constraints. This distinction also matters for uncertainty: the probability of an entire sequence and the marginal confidence in one token are not interchangeable."
    },
    {
      "heading": "Parameter estimation and optimization",
      "text": "The standard learning method is maximum conditional likelihood from labeled examples \\((\\mathbf{x}^{(n)},\\mathbf{y}^{(n)})\\). For a log-linear CRF, a regularized negative log-likelihood is\n\n\\[\\mathcal L(\\theta) = \\sum_{n=1}^{N} \\left[ \\log Z_\\theta(\\mathbf{x}^{(n)}) - \\theta^\\top \\mathbf{F}(\\mathbf{x}^{(n)},\\mathbf{y}^{(n)}) \\right] + \\frac{\\lambda}{2}\\|\\theta\\|_2^2.\\]\n\nThis loss function rewards the observed configurations while accounting for all competing configurations through the partition function. The final term is \\(L_2\\) regularization, which penalizes large parameter values.\n\nThe gradient has an informative structure:\n\n\\[\\nabla_\\theta\\mathcal L = \\sum_n \\left[ \\mathbb E_{p_\\theta(\\mathbf{Y}\\mid\\mathbf{x}^{(n)})} \\mathbf{F}(\\mathbf{x}^{(n)},\\mathbf{Y}) - \\mathbf{F}(\\mathbf{x}^{(n)},\\mathbf{y}^{(n)}) \\right] + \\lambda\\theta.\\]\n\nLearning compares feature counts expected under the model with feature counts observed in the annotations. If a feature occurs more frequently in the data than the model expects, the likelihood gradient favors increasing its weight. For local features, the expectations can be computed from local marginals rather than from explicit enumeration of complete outputs. This makes exact chain inference useful not only during prediction but also at every training iteration.\n\nWith fixed features, scores linear in \\(\\theta\\), and exact normalization, the negative log-likelihood is convex. Its Hessian contains feature-covariance terms and is positive semidefinite; a positive \\(L_2\\) penalty makes it strongly convex in the unconstrained parameter vector. Thus classical CRF learning does not face the same local-minimum problem as general neural-network training. Computational difficulty can nevertheless remain substantial because each objective or gradient evaluation requires inference across the training examples. Limited-memory quasi-Newton methods, especially L-BFGS, have been widely used for this setting.\n\nOther objectives and penalties change the optimization problem. An \\(L_1\\) penalty encourages sparse parameters but introduces nondifferentiability. Online and stochastic methods process examples or minibatches rather than repeatedly evaluating the full dataset. When neural networks generate CRF scores, gradients propagate through the normalization computation and into the encoder using backpropagation. The combined objective is generally nonconvex, even though the CRF layer considered with fixed inputs retains a familiar form.\n\nTraining data may also be incomplete. If only some labels are observed, their conditional likelihood sums over compatible completions. The resulting loss is a difference between two log-partition functions and need not preserve classical convexity. Latent-variable CRFs similarly sum over hidden structures. These models require a clear distinction between labels that are genuinely observed, labels that are missing, and structures introduced only for modeling purposes; they also alter the inference problems performed during training."
    },
    {
      "heading": "Relation to other probabilistic and discriminative models",
      "text": "CRFs are often compared with hidden Markov models (HMMs), which model a joint distribution over observations and hidden states. A conventional HMM factors into state transitions and observation emissions, typically assuming that each observation depends only on its corresponding hidden state. A CRF instead conditions on the complete observation and models only the labels. It can therefore use correlated and overlapping input features without specifying a normalized probability model for those features. This is a difference in modeling commitments, not a universal guarantee of better predictive performance.\n\nA linear-chain HMM conditioned on a fixed observation sequence induces a distribution over state sequences that can be represented as a chain CRF. Its transition and emission probabilities become factors in the conditional model. The reverse comparison is less direct: general CRF scores need not correspond to a conventional HMM with local emissions. Generative modeling can also serve purposes outside a CRF’s scope, such as assigning likelihoods to inputs or generating observations. Conditional modeling trades away those goals to concentrate model capacity on predicting outputs from observations.\n\nAnother important comparison is with the maximum entropy Markov model (MEMM). A common MEMM factorization is\n\n\\[p(\\mathbf{y}\\mid\\mathbf{x}) = \\prod_t p(y_t\\mid y_{t-1},\\mathbf{x}).\\]\n\nEach local conditional distribution normalizes separately over possible next labels. A chain CRF normalizes once over complete sequences. This distinction motivated CRFs partly as a response to the label-bias problem: local normalization can make states with few outgoing alternatives insufficiently responsive to evidence that would distinguish entire paths. A state with only one legal outgoing transition must assign it conditional probability one, whatever the observed evidence.\n\nGlobal normalization allows competing paths to be compared using their complete scores rather than through a product of separately normalized transition decisions. It does not eliminate every source of bias or ensure that a chosen graph represents the task adequately. The benefit is specific: probability allocation is coordinated over complete structures. Both locally and globally normalized models still depend on their features, parameter estimation, annotation conventions, and available training examples.\n\nLogistic regression is a useful limiting case. A single categorical output with log-linear scores yields multinomial logistic regression; multiple outputs without interactions yield a product of independent classifiers. Adding output factors turns this independent model into a structured one. CRFs are also related to structured support-vector machines, which can use similar scores and decoding algorithms but train with margin-based rather than conditional-likelihood objectives. A structured margin model need not define normalized probabilities. The graph, scoring function, training criterion, and decoding procedure are therefore separate design choices, even when several methods share the same best-structure search."
    },
    {
      "heading": "Beyond chains: richer structures and approximate inference",
      "text": "The CRF definition is not restricted to sequences. A pairwise model can assign labels to vertices of a grid, a tree, or an arbitrary graph, with an energy such as\n\n\\[E(\\mathbf{y},\\mathbf{x}) = \\sum_i \\phi_i(y_i,\\mathbf{x}) + \\sum_{(i,j)\\in E} \\phi_{ij}(y_i,y_j,\\mathbf{x}).\\]\n\nUnary terms express local evidence, while pairwise terms express compatibility. In an image, a unary term may favor a category based on appearance, and a pairwise term may penalize assigning different categories to adjacent, visually similar pixels. Such smoothness is a modeling preference, not a requirement that all neighboring pixels share a class.\n\nA semi-Markov CRF assigns labels to segments rather than only to individual positions. A segment can have features describing its complete span, duration, or internal pattern. Sunita Sarawagi and William Cohen introduced semi-Markov CRFs for information extraction in 2004. If segments have a maximum length \\(L\\), a typical first-order dynamic program has complexity on the order of \\(TLK^2\\), assuming constant-time access to segment scores. This lets an extraction model score an entire person name or address component directly rather than reconstructing every span solely through token transitions.\n\nOther extensions include higher-order factors, factorial structures with several interacting label chains, and latent-state CRFs. Their representational power comes with greater computational demands. Even a graph constructed from individually simple factors can make exact summation prohibitively expensive. For broad classes of discrete graphical models, exact normalization and marginal inference are intractable in general. Exact methods remain practical for restricted structures, especially graphs with small treewidth.\n\nApproximate methods include loopy belief propagation, variational approximations, sampling, and specialized optimization procedures. In variational inference, a tractable distribution approximates the desired conditional distribution; mean-field methods often factor this approximation across variables. Loopy belief propagation repeatedly passes local messages around graphs containing cycles, but unlike its tree-structured counterpart it does not generally guarantee exact answers or convergence. Approximate maximum-score decoding and approximate partition-function estimation are distinct tasks and can require different algorithms.\n\nA notable vision development was the fully connected CRF with Gaussian edge potentials described by Philipp Krähenbühl and Vladlen Koltun in 2011. It links many pairs of pixels and uses efficient high-dimensional filtering to accelerate approximate mean-field updates. Spatial and appearance-sensitive interactions can encourage coherent regions without merely smoothing across strong image boundaries. Some binary pairwise energies also admit exact optimization using minimum-cut methods when submodularity conditions hold. That result concerns the best assignment, not necessarily exact probabilities. The suitability of an inference algorithm therefore depends on both the graph and the numerical form of its factors."
    },
    {
      "heading": "Neural CRFs and representation learning",
      "text": "A neural CRF uses an artificial neural network to generate some or all of its potentials. In sequence labeling, an encoder produces a contextual vector \\(h_t\\) for each input position, and a learned map converts it into label scores:\n\n\\[u_t(k)=\\mathbf{w}_k^\\top h_t+b_k.\\]\n\nA transition matrix then adds scores for adjacent labels. The model is trained with the CRF conditional likelihood, so its normalization couples the output decisions while gradients train the encoder. The neural network learns representations; the CRF layer provides an explicit distribution over structured outputs.\n\nNeural sequence models with sentence-level structured objectives appeared in the work of Ronan Collobert and colleagues in 2011. Bidirectional long short-term memory networks combined with CRFs became prominent in the mid-2010s. Zhiheng Huang, Wei Xu, and Kai Yu described a bidirectional LSTM–CRF model in 2015. In 2016, Guillaume Lample and colleagues presented a neural named-entity recognition architecture using character-based and contextual word representations with a CRF output layer. These systems helped replace large collections of manual lexical features with learned representations while retaining explicit sequence-level scoring.\n\nIn a bidirectional recurrent encoder, a token representation incorporates preceding and following input. In a Transformer encoder, attention supplies contextual information through a different architecture. Models based on BERT and related pretrained encoders can place a CRF above token representations. The fact that the encoder already sees the whole sentence does not make the CRF mathematically redundant: contextual input processing and output-label dependence are different operations. Whether the additional structured layer improves performance is an empirical question, not a consequence of the architecture alone.\n\nNeural CRFs require careful alignment between model positions and annotated units. Tokenization into subwords may turn one annotated word into several encoder positions. A system can take one representation per word, pool several subword representations, or expand the annotation scheme to subword positions. Padding and sentence-boundary masks must also be excluded correctly from the dynamic program. A formally correct CRF cannot compensate for a mismatch between the sequence used in its loss and the sequence evaluated as output.\n\nEnd-to-end learning also changes computational priorities. The CRF layer has \\(O(TK^2)\\) cost for a dense first-order chain, but a large encoder can dominate total runtime. Conversely, large label sets or complex output factors can make structured inference expensive. Neural scores do not automatically encode validity constraints: a learned transition can discourage an illegal label pair without forbidding it. Fixed masks can enforce permitted transitions during training, decoding, or both. Those choices define different models and must be kept distinct from merely learning a transition matrix."
    },
    {
      "heading": "Applications, evaluation, and practical limitations",
      "text": "CRFs have been used extensively in linguistic tasks, including named-entity recognition, part-of-speech tagging, word segmentation, chunking, and information extraction. Their appeal is particularly clear when outputs have a small local label vocabulary but must combine into meaningful spans. A document extraction system, for example, can label tokens as parts of names, dates, or addresses while representing boundaries between fields. CRFs have also been applied to biological sequence annotation and temporal activity labeling, where adjacent states or segment durations convey information that separate local classifiers may overlook.\n\nIn computer vision, CRFs have supported image segmentation and semantic segmentation by combining local category evidence with spatial relations. A classifier may produce uncertain or irregular pixel predictions; a CRF can express preferences for compatible categories, coherent regions, and boundaries aligned with observed appearance. Strong pairwise smoothing can also erase small objects or thin structures. As visual encoders became better at incorporating spatial context, explicit CRF postprocessing ceased to be a necessary component of many architectures, although structured graphical models remained useful for particular constraints and interaction patterns.\n\nEvaluation must match the task’s output structure. Token accuracy can conceal severe boundary errors in entity extraction, so such systems commonly use entity-level precision, recall, and F1 with exact span and category matching. Segmentation systems often report intersection over union for predicted and reference regions. Conditional log-likelihood measures the probability assigned to annotated outputs, whereas these task metrics assess decoded decisions. An improvement in likelihood does not necessarily produce a corresponding improvement in span-level F1 or region overlap.\n\nReliable comparison also requires separating the contribution of the CRF from that of its encoder, features, and preprocessing. A meaningful comparison with independent classification keeps the input representation, training examples, and evaluation conventions as similar as possible. Changes in token alignment, transition constraints, or sentence splitting can alter results independently of global normalization. Documents or closely related sequences should not be inadvertently shared across training and evaluation partitions, since such data leakage can inflate measured performance without demonstrating better prediction on new material.\n\nA normalized conditional distribution does not guarantee well-calibrated uncertainty or robustness under distribution shift. CRFs can learn annotation artifacts, rely on domain-specific word patterns, and assign high confidence to outputs on unfamiliar inputs. Long sequences also naturally distribute probability across many complete configurations, so low sequence probability alone is not an interpretable measure of token-level unreliability. Practical implementations must distinguish numerical stability, successful optimization, structural validity, and empirical accuracy: each addresses a different possible failure. Unresolved modeling choices include how to encode long-range constraints without intractable inference, how to learn efficiently from incomplete labels, and how to preserve useful probabilistic behavior when approximate inference replaces exact normalization."
    }
  ],
  "links": [
    "probability-distribution",
    "machine-learning",
    "natural-language-processing",
    "computer-vision",
    "random-variable",
    "partition-function",
    "markov-random-field",
    "factor-graph",
    "feature-engineering",
    "energy-based-model",
    "dynamic-programming",
    "loss-function",
    "regularization",
    "backpropagation",
    "hidden-markov-model",
    "logistic-regression",
    "support-vector-machine",
    "belief-propagation",
    "variational-inference",
    "artificial-neural-network",
    "long-short-term-memory",
    "transformer-architecture",
    "bert",
    "tokenization",
    "image-segmentation",
    "semantic-segmentation",
    "data-leakage",
    "distribution-shift"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}