The IJCAI Computers and Thought Award is presented every two years by the International Joint Conference on Artificial Intelligence (IJCAI), recognizing outstanding young scientists in artificial intelligence. It was originally funded with royalties received from the book Computers and Thought (edited by Edward Feigenbaum and Julian Feldman), and is currently funded by IJCAI. It is considered to be "the premier award for artificial intelligence researchers under the age of 35". == Laureates == Terry Winograd (1971) Patrick Winston (1973) Chuck Rieger (1975) Douglas Lenat (1977) David Marr (1979) Gerald Sussman (1981) Tom Mitchell (1983) Hector Levesque (1985) Johan de Kleer (1987) Henry Kautz (1989) Rodney Brooks (1991) Martha E. Pollack (1991) Hiroaki Kitano (1993) Sarit Kraus (1995) Stuart Russell (1995) Leslie Kaelbling (1997) Nicholas Jennings (1999) Daphne Koller (2001) Tuomas Sandholm (2003) Peter Stone (2007) Carlos Guestrin (2009) Andrew Ng (2009) Vincent Conitzer (2011) Malte Helmert (2011) Kristen Grauman (2013) Ariel D. Procaccia (2015) Percy Liang (2016) for his contributions to both the approach of semantic parsing for natural language understanding and better methods for learning latent-variable models, sometimes with weak supervision, in machine learning. Devi Parikh (2017) Stefano Ermon (2018) Guy Van den Broeck (2019) for his contributions to statistical and relational artificial intelligence, and the study of tractability in learning and reasoning. Piotr Skowron (2020) for his contributions to computational social choice, and to the theory of committee elections. Fei Fang (2021) for her contributions to integrating machine learning with game theory and the use of these novel techniques to tackle societal challenges such as more effective deployment of security resources, enhancing environmental sustainability, and reducing food insecurity. Bo Li (2022) for her contributions to uncovering the underlying connections among robustness, privacy, and generalization in AI, showing how different models are vulnerable to malicious attacks, and how to eliminate these vulnerabilities using mathematical tools that provide robustness guarantees for learning models and privacy protection. Pin-Yu Chen (2023) for his contributions to consolidating properties of trust, robustness and safety into rigorous algorithmic procedures and computable metrics for improving AI systems. Nisarg Shah (2024) for his contributions to AI and society, in particular foundational work on the theory of algorithmic fairness using principles from social choice theory. Aditya Grover (2025) for his foundational contributions uniting deep generative models, representation learning, and reinforcement learning, and for their applications in advancing scientific reasoning.
Dhammin
Dhammin (Arabic: ضمّن) is a political platform that manages candidates' electoral campaigns for the National Assembly, Municipal Council or Cooperative Society councils of Kuwait. The platform was founded by Abdullah Al-Salloum and it is, according to news reports and interviews, the first within the field to apply distributed-systems' methodologies.
Deep learning
In machine learning, deep learning (DL) focuses on utilizing multilayered neural networks to perform tasks such as classification, regression, and representation learning. The field takes inspiration from biological neuroscience and revolves around stacking artificial neurons into layers and "training" them to process data. The adjective "deep" refers to the use of multiple layers (ranging from three to several hundred or thousands) in the network. Methods used can be supervised, semi-supervised or unsupervised. Some common deep learning network architectures include fully connected networks, deep belief networks, recurrent neural networks, convolutional neural networks, generative adversarial networks, transformers, and neural radiance fields. These architectures have been applied to fields including computer vision, speech recognition, natural language processing, machine translation, bioinformatics, drug design, medical image analysis, climate science, material inspection and board game programs, where they have produced results comparable to and in some cases surpassing human expert performance. Early forms of neural networks were inspired by information processing and distributed communication nodes in biological systems, particularly the human brain. However, current neural networks do not intend to model the brain function of organisms, and are generally seen as low-quality models for that purpose. == Overview == Most modern deep learning models are based on multi-layered neural networks such as convolutional neural networks and transformers, although they can also include propositional formulas or latent variables organized layer-wise in deep generative models such as the nodes in deep belief networks and deep Boltzmann machines. Fundamentally, deep learning refers to a class of machine learning algorithms in which a hierarchy of layers is used to transform input data into a progressively more abstract and composite representation. For example, in an image recognition model, the raw input may be an image (represented as a tensor of pixels). The first representational layer may attempt to identify basic shapes such as lines and circles, the second layer may compose and encode arrangements of edges, the third layer may encode a nose and eyes, and the fourth layer may recognize that the image contains a face. Importantly, a deep learning process can learn which features to optimally place at which level on its own. Prior to deep learning, machine learning techniques often involved hand-crafted feature engineering to transform the data into a more suitable representation for a classification algorithm to operate on. In the deep learning approach, features are not hand-crafted and the model discovers useful feature representations from the data automatically. This does not eliminate the need for hand-tuning; for example, varying numbers of layers and layer sizes can provide different degrees of abstraction. The word "deep" in "deep learning" refers to the number of layers through which the data is transformed. More precisely, deep learning systems have a substantial credit assignment path (CAP) depth. The CAP is the chain of transformations from input to output. CAPs describe potentially causal connections between input and output. For a feedforward neural network, the depth of the CAPs is that of the network and is the number of hidden layers plus one (as the output layer is also parameterized). For recurrent neural networks, in which a signal may propagate through a layer more than once, the CAP depth is potentially unlimited. No universally agreed-upon threshold of depth divides shallow learning from deep learning, but most researchers agree that deep learning involves CAP depth higher than two. CAP of depth two has been shown to be a universal approximator in the sense that it can emulate any function. Beyond that, more layers do not add to the function approximator ability of the network. Deep models (CAP > two) are able to extract better features than shallow models and hence, extra layers help in learning the features effectively. Deep learning architectures can be constructed with a greedy layer-by-layer method. Deep learning helps to disentangle these abstractions and pick out which features improve performance. Deep learning algorithms can be applied to unsupervised learning tasks. This is an important benefit because unlabeled data is more abundant than labeled data. Examples of deep structures that can be trained in an unsupervised manner are deep belief networks. The term deep learning was introduced to the machine learning community by Rina Dechter in 1986, and to artificial neural networks by Igor Aizenberg and colleagues in 2000, in the context of Boolean threshold neurons. The etymology of the term is more complicated. == Interpretations == Deep neural networks are generally interpreted in terms of the universal approximation theorem or probabilistic inference. The classic universal approximation theorem concerns the capacity of feedforward neural networks with a single hidden layer of finite size to approximate continuous functions. In 1989, the first proof was published by George Cybenko for sigmoid activation functions and was generalised to feed-forward multi-layer architectures in 1991 by Kurt Hornik. Recent work also showed that universal approximation also holds for non-bounded activation functions such as Kunihiko Fukushima's rectified linear unit. The universal approximation theorem for deep neural networks concerns the capacity of networks with bounded width but the depth is allowed to grow. Lu et al. proved that if the width of a deep neural network with ReLU activation is strictly larger than the input dimension, then the network can approximate any Lebesgue integrable function; if the width is smaller or equal to the input dimension, then a deep neural network is not a universal approximator. The probabilistic interpretation derives from the field of machine learning. It features inference, as well as the optimization concepts of training and testing, related to fitting and generalization, respectively. More specifically, the probabilistic interpretation considers the activation nonlinearity as a cumulative distribution function. The probabilistic interpretation led to the introduction of dropout as regularizer in neural networks. The probabilistic interpretation was introduced by researchers including Hopfield, Widrow and Narendra and popularized in surveys such as the one by Bishop. == History == === Before 1980 === There are two types of artificial neural network (ANN): feedforward neural network (FNN) or multilayer perceptron (MLP) and recurrent neural networks (RNN). RNNs have cycles in their connectivity structure, whereas FNNs do not. In the 1920s, Wilhelm Lenz and Ernst Ising created the Ising model which is essentially a non-learning RNN architecture consisting of neuron-like threshold elements. In 1972, Shun'ichi Amari made this architecture adaptive. His learning RNN was republished by John Hopfield in 1982. Other early recurrent neural networks were published by Kaoru Nakano in 1971. Already in 1948, Alan Turing produced work on "Intelligent Machinery" that was not published in his lifetime, containing "ideas related to artificial evolution and learning RNNs". Frank Rosenblatt (1958) proposed the perceptron, an MLP with 3 layers: an input layer, a hidden layer with randomized weights that did not learn, and an output layer. He later published a 1962 book that also introduced variants and computer experiments, including a version with four-layer perceptrons "with adaptive preterminal networks" where the last two layers have learned weights (here he credits H. D. Block and B. W. Knight). The book cites an earlier network by R. D. Joseph (1960) "functionally equivalent to a variation of" this four-layer system (the book mentions Joseph over 30 times). Should Joseph therefore be considered the originator of proper adaptive multilayer perceptrons with learning hidden units? Unfortunately, the learning algorithm was not a functional one, and fell into oblivion. The first working deep learning algorithm was the Group method of data handling, a method to train arbitrarily deep neural networks, published by Alexey Ivakhnenko and Lapa in 1965. They regarded it as a form of polynomial regression, or a generalization of Rosenblatt's perceptron to handle more complex, nonlinear, and hierarchical relationships. A 1971 paper described a deep network with eight layers trained by this method, which is based on layer by layer training through regression analysis. Superfluous hidden units are pruned using a separate validation set. Since the activation functions of the nodes are Kolmogorov-Gabor polynomials, these were also the first deep networks with multiplicative units or "gates". The first deep learning multilayer perceptron trained by stochastic gradient descent was published in 1967 by Shun'ichi
IT operations analytics
In the fields of information technology (IT) and systems management, IT operations analytics (ITOA) is an approach or method to retrieve, analyze, and report data for IT operations. ITOA may apply big data analytics to large datasets to produce business insights. In 2014, Gartner predicted its use might increase revenue or reduce costs. By 2017, it predicted that 15% of enterprises will use IT operations analytics technologies. == Definition == IT operations analytics (ITOA) (also known as advanced operational analytics, or IT data analytics) technologies are primarily used to discover complex patterns in high volumes of often "noisy" IT system availability and performance data. Forrester Research defined IT analytics as "The use of mathematical algorithms and other innovations to extract meaningful information from the sea of raw data collected by management and monitoring technologies." Note, ITOA is different than AIOps, which focuses on applying artificial intelligence and machine learning to the applications of ITOA. == History == Operations research as a discipline emerged from the Second World War to improve military efficiency and decision-making on the battlefield. However, only with the emergence of machine learning tech in the early 2000s could an artificially intelligent operational analytics platform actually begin to engage in the high-level pattern recognition that could adequately serve business needs. A critical catalyst towards ITOA development was the rise of Google, which pioneered a predictive analytics model that represented the first attempt to read into patterns of human behavior on the Internet. IT specialists then applied predictive analytics to the IT Industry, coming forward with platforms that can sift through data to generate insights without the need for human intervention. Due to the mainstream embrace of cloud computing and the increasing desire for businesses to adopt more big data practices, the ITOA industry has grown significantly since 2010. A 2016 ExtraHop survey of large and mid-size corporations indicates that 65 percent of the businesses surveyed will seek to integrate their data silos either this year or the next. The current goals of ITOA platforms are to improve the accuracy of their APM services, facilitate better integration with the data, and to enhance their predictive analytics capabilities. == Applications == ITOA systems tend to be used by IT operations teams, and Gartner describes seven applications of ITOA systems: Root cause analysis: The models, structures and pattern descriptions of IT infrastructure or application stack being monitored can help users pinpoint fine-grained and previously unknown root causes of overall system behavior pathologies. Proactive control of service performance and availability: Predicts future system states and the impact of those states on performance. Problem assignment: Determines how problems may be resolved or, at least, direct the results of inferences to the most appropriate individuals, or communities in the enterprise for problem resolution. Service impact analysis: When multiple root causes are known, the analytics system's output is used to determine and rank the relative impact, so that resources can be devoted to correcting the fault in the most timely and cost-effective way possible. Complement best-of-breed technology: The models, structures and pattern descriptions of IT infrastructure or application stack being monitored are used to correct or extend the outputs of other discovery-oriented tools to improve the fidelity of information used in operational tasks (e.g., service dependency maps, application runtime architecture topologies, network topologies). Real time application behavior learning: Learns & correlates the behavior of Application based on user pattern and underlying Infrastructure on various application patterns, create metrics of such correlated patterns and store it for further analysis. Dynamically baselines threshold: Learns behavior of Infrastructure on various application user patterns and determines the Optimal behavior of the Infra and technological components, bench marks and baselines the low and high water mark for the specific environments and dynamically changes the bench mark baselines with the changing infra and user patterns without any manual intervention. == Types == In their Data Growth Demands a Single, Architected IT Operations Analytics Platform, Gartner Research describes five types of analytics technologies: Log analysis Unstructured text indexing, search and inference (UTISI) Topological analysis (TA) Multidimensional database search and analysis (MDSA) Complex operations event processing (COEP) Statistical pattern discovery and recognition (SPDR) == Tools and ITOA platforms == A number of vendors operate in the ITOA space:
News analytics
In trading strategy, news analysis refers to the measurement of the various qualitative and quantitative attributes of textual (unstructured data) news stories. Some of these attributes are: sentiment, relevance, and novelty. Expressing news stories as numbers and metadata permits the manipulation of everyday information in a mathematical and statistical way. This data is often used in financial markets as part of a trading strategy or by businesses to judge market sentiment and make better business decisions. News analytics are usually derived through automated text analysis and applied to digital texts using elements from natural language processing and machine learning such as latent semantic analysis, support vector machines, "bag of words" among other techniques. == Applications and strategies == The application of sophisticated linguistic analysis to news and social media has grown from an area of research to mature product solutions since 2007. News analytics and news sentiment calculations are now routinely used by both buy-side and sell-side in alpha generation, trading execution, risk management, and market surveillance and compliance. There is however a good deal of variation in the quality, effectiveness and completeness of currently available solutions. A large number of companies use news analysis to help them make better business decisions. Academic researchers have become interested in news analysis especially with regards to predicting stock price movements, volatility and traded volume. Provided a set of values such as sentiment and relevance as well as the frequency of news arrivals, it is possible to construct news sentiment scores for multiple asset classes such as equities, Forex, fixed income, and commodities. Sentiment scores can be constructed at various horizons to meet the different needs and objectives of high and low frequency trading strategies, whilst characteristics such as direction and volatility of asset returns as well as the traded volume may be addressed more directly via the construction of tailor-made sentiment scores. Scores are generally constructed as a range of values. For instance, values may range between 0 and 100, where values above and below 50 convey positive and negative sentiment, respectively. === Absolute return strategies === The objective of absolute return strategies is absolute (positive) returns regardless of the direction of the financial market. To meet this objective, such strategies typically involve opportunistic long and short positions in selected instruments with zero or limited market exposure. In statistical terms, absolute return strategies should have very low correlation with the market return. Typically, hedge funds tend to employ absolute return strategies. Below, a few examples show how news analysis can be applied in the absolute return strategy space with the purpose to identify alpha opportunities applying a market neutral strategy or based on volatility trading. Example 1 Scenario: The gap between the news sentiment scores for direction, S {\displaystyle S} , of Company X {\displaystyle X} and Market Y {\displaystyle Y} has moved beyond + 20 {\displaystyle +20} . That is, S X − S Y {\displaystyle S_{X}-S_{Y}} ≥ 20 {\displaystyle 20} . Action: Buy the stock on Company X {\displaystyle X} and short the future on Market Y {\displaystyle Y} . Exit Strategy: When the gap in the news sentiment scores for direction of Company X {\displaystyle X} and Market Y {\displaystyle Y} has disappeared, S X − S Y {\displaystyle S_{X}-S_{Y}} = 0 {\displaystyle 0} , sell the stock on Company X {\displaystyle X} and go long the future on Market Y {\displaystyle Y} to close the positions. Example 2 Scenario: The news sentiment score for volatility of Company X {\displaystyle X} goes above 70 {\displaystyle 70} out of 100 {\displaystyle 100} indicating an expected volatility above the option implied volatility. Action: Buy a short-dated straddle (the purchase of both a put and a call) on the stock of Company X {\displaystyle X} . Exit Strategy: Keep the straddle on Company X {\displaystyle X} until expiry or until a certain profit target has been reached. === Relative return strategies === The objective of relative return strategies is to either replicate (passive management) or outperform (active management) a theoretical passive reference portfolio or benchmark. To meet these objectives such strategies typically involve long positions in selected instruments. In statistical terms, relative return strategies often have high correlation with the market return. Typically, mutual funds tend to employ relative return strategies. Below, a few examples show how news analysis can be applied in the relative return strategy space with the purpose to outperform the market applying a stock picking strategy and by making tactical tilts to ones asset allocation model. Example 1 Scenario: The news sentiment score for direction of Company X {\displaystyle X} goes above 70 {\displaystyle 70} out of 100 {\displaystyle 100} . Action: Buy the stock on Company X {\displaystyle X} . Exit Strategy: When the news sentiment score for direction of Company X {\displaystyle X} falls below 60 {\displaystyle 60} , sell the stock on Company X {\displaystyle X} to close the position. Example 2 Scenario: The news sentiment score for direction of Sector Z {\displaystyle Z} goes above 70 {\displaystyle 70} out of 100 {\displaystyle 100} . Action: Include Sector Z {\displaystyle Z} as a tactical bet in the asset allocation model. Exit Strategy: When the news sentiment score for direction of Sector Z {\displaystyle Z} falls below 60 {\displaystyle 60} , remove the tactical bet for Sector Z {\displaystyle Z} from the asset allocation model. === Financial risk management === The objective of financial risk management is to create economic value in a firm or to maintain a certain risk profile of an investment portfolio by using financial instruments to manage risk exposures, particularly credit risk and market risk. Other types include Foreign exchange, Shape, Volatility, Sector, Liquidity, Inflation risks, etc. Below, a few examples show how news analysis can be applied in the financial risk management space with the purpose to either arrive at better risk estimates in terms of Value at Risk (VaR) or to manage the risk of a portfolio to meet ones portfolio mandate. Example 1 Scenario: The bank operates a VaR model to manage the overall market risk of its portfolio. Action: Estimate the portfolio covariance matrix taking into account the development of the news sentiment score for volume. Implement the relevant hedges to bring the VaR of the bank in line with the desired levels. Example 2 Scenario: A portfolio manager operates his portfolio towards a certain desired risk profile. Action: Estimate the portfolio covariance matrix taking into account the development of the news sentiment score for volume. Scale the portfolio exposure according to the targeted risk profile. === Computer algorithms using news analytics === Within 0.33 seconds, computer algorithms using news analytics can notify subscribers which company the news is about, if the news article sentiment is positive or negative, if the news is ranked as high or low relative importance … relative relevance. the stock price reaction and the increase in trade volume is concentrated in the first 5 seconds after an news article is released. === Algorithmic order execution === The objective of algorithmic order execution, which is part of the concept of algorithmic trading, is to reduce trading costs by optimizing on the timing of a given order. It is widely used by hedge funds, pension funds, mutual funds, and other institutional traders to divide up large trades into several smaller trades to manage market impact, opportunity cost, and risk more effectively. The example below shows how news analysis can be applied in the algorithmic order execution space with the purpose to arrive at more efficient algorithmic trading systems. Example 1 Scenario: A large order needs to be placed in the market for the stock on Company X {\displaystyle X} . Action: Scale the daily volume distribution for Company X {\displaystyle X} applied in the algorithmic trading system, thus taking into account the news sentiment score for volume. This is followed by the creation of the desired trading distribution forcing greater market participation during the periods of the day when volume is expected to be heaviest. == Effects == Being able to express news stories as numbers permits the manipulation of everyday information in a statistical way that allows computers not only to make decisions once made only by humans, but to do so more efficiently. Since market participants are always looking for an edge, the speed of computer connections and the delivery of news analysis, measured in milliseconds, have become essential.
Apache OpenNLP
The Apache OpenNLP library is a machine learning based toolkit for the processing of natural language text. It supports the most common NLP tasks, such as language detection, tokenization, sentence segmentation, part-of-speech tagging, named entity extraction, chunking, parsing and coreference resolution. These tasks are usually required to build more advanced text processing services.
BERT (language model)
Bidirectional encoder representations from transformers (BERT) is a language model introduced in October 2018 by researchers at Google. It learns to represent text as a sequence of vectors using self-supervised learning. It uses the encoder-only transformer architecture. BERT dramatically improved the state of the art for large language models. As of 2020, BERT is a ubiquitous baseline in natural language processing (NLP) experiments. BERT is trained by masked token prediction and next sentence prediction. With this training, BERT learns contextual, latent representations of tokens in their context, similar to ELMo and GPT-2. It found applications for many natural language processing tasks, such as coreference resolution and polysemy resolution. It improved on ELMo and spawned the study of "BERTology", which attempts to interpret what is learned by BERT. BERT was originally implemented in the English language at two model sizes, BERTBASE (110 million parameters) and BERTLARGE (340 million parameters). Both were trained on the Toronto BookCorpus (800M words) and English Wikipedia (2,500M words). The weights were released on GitHub. On March 11, 2020, 24 smaller models were released, the smallest being BERTTINY with just 4 million parameters. == Architecture == BERT is an "encoder-only" transformer architecture. At a high level, BERT consists of 4 modules: Tokenizer: This module converts a piece of English text into a sequence of integers ("tokens"). Embedding: This module converts the sequence of tokens into an array of real-valued vectors representing the tokens. It represents the conversion of discrete token types into a lower-dimensional Euclidean space. Encoder: a stack of Transformer blocks with self-attention, but without causal masking. Task head: This module converts the final representation vectors into one-shot encoded tokens again by producing a predicted probability distribution over the token types. It can be viewed as a simple decoder, decoding the latent representation into token types, or as an "un-embedding layer". The task head is necessary for pre-training, but it is often unnecessary for so-called "downstream tasks," such as question answering or sentiment classification. Instead, one removes the task head and replaces it with a newly initialized module suited for the task, and finetune the new module. The latent vector representation of the model is directly fed into this new module, allowing for sample-efficient transfer learning. === Embedding === This section describes the embedding used by BERTBASE. The other one, BERTLARGE, is similar, just larger. The tokenizer of BERT is WordPiece, which is a sub-word strategy like byte-pair encoding. Its vocabulary size is 30,000, and any token not appearing in its vocabulary is replaced by [UNK] ("unknown"). The first layer is the embedding layer, which contains three components: token type embeddings, position embeddings, and segment type embeddings. Token type: The token type is a standard embedding layer, translating a one-hot vector into a dense vector based on its token type. Position: The position embeddings are based on a token's position in the sequence. BERT uses absolute position embeddings, where each position in a sequence is mapped to a real-valued vector. Each dimension of the vector consists of a sinusoidal function that takes the position in the sequence as input. Segment type: Using a vocabulary of just 0 or 1, this embedding layer produces a dense vector based on whether the token belongs to the first or second text segment in that input. In other words, type-1 tokens are all tokens that appear after the [SEP] special token. All prior tokens are type-0. The three embedding vectors are added together representing the initial token representation as a function of these three pieces of information. After embedding, the vector representation is normalized using a LayerNorm operation, outputting a 768-dimensional vector for each input token. After this, the representation vectors are passed forward through 12 Transformer encoder blocks, and are decoded back to 30,000-dimensional vocabulary space using a basic affine transformation layer. === Architectural family === The encoder stack of BERT has 2 free parameters: L {\displaystyle L} , the number of layers, and H {\displaystyle H} , the hidden size. There are always H / 64 {\displaystyle H/64} self-attention heads, and the feed-forward/filter size is always 4 H {\displaystyle 4H} . By varying these two numbers, one obtains an entire family of BERT models. For BERT: the feed-forward size and filter size are synonymous. Both of them denote the number of dimensions in the middle layer of the feed-forward network. the hidden size and embedding size are synonymous. Both of them denote the number of real numbers used to represent a token. The notation for encoder stack is written as L/H. For example, BERTBASE is written as 12L/768H, BERTLARGE as 24L/1024H, and BERTTINY as 2L/128H. == Training == === Pre-training === BERT was pre-trained simultaneously on two tasks: Masked language modeling (MLM): In this task, BERT ingests a sequence of words, where one word may be randomly changed ("masked"), and BERT tries to predict the original words that had been changed. For example, in the sentence "The cat sat on the [MASK]," BERT would need to predict "mat." This helps BERT learn bidirectional context, meaning it understands the relationships between words not just from left to right or right to left but from both directions at the same time. Next sentence prediction (NSP): In this task, BERT is trained to predict whether one sentence logically follows another. For example, given two sentences, "The cat sat on the mat" and "It was a sunny day", BERT has to decide if the second sentence is a valid continuation of the first one. This helps BERT understand relationships between sentences, which is important for tasks like question answering or document classification. ==== Masked language modeling ==== In masked language modeling, 15% of tokens would be randomly selected for masked-prediction task, and the training objective was to predict the masked token given its context. In more detail, the selected token is: replaced with a [MASK] token with probability 80%, replaced with a random word token with probability 10%, not replaced with probability 10%. The reason not all selected tokens are masked is to avoid the dataset shift problem. The dataset shift problem arises when the distribution of inputs seen during training differs significantly from the distribution encountered during inference. A trained BERT model might be applied to word representation (like Word2Vec), where it would be run over sentences not containing any [MASK] tokens. It is later found that more diverse training objectives are generally better. As an illustrative example, consider the sentence "my dog is cute". It would first be divided into tokens like "my1 dog2 is3 cute4". Then a random token in the sentence would be picked. Let it be the 4th one "cute4". Next, there would be three possibilities: with probability 80%, the chosen token is masked, resulting in "my1 dog2 is3 [MASK]4"; with probability 10%, the chosen token is replaced by a uniformly sampled random token, such as "happy", resulting in "my1 dog2 is3 happy4"; with probability 10%, nothing is done, resulting in "my1 dog2 is3 cute4". After processing the input text, the model's 4th output vector is passed to its decoder layer, which outputs a probability distribution over its 30,000-dimensional vocabulary space. ==== Next sentence prediction ==== Given two sentences, the model predicts if they appear sequentially in the training corpus, outputting either [IsNext] or [NotNext]. During training, the algorithm sometimes samples two sentences from a single continuous span in the training corpus, while at other times, it samples two sentences from two discontinuous spans. The first sentence starts with a special token, [CLS] (for "classify"). The two sentences are separated by another special token, [SEP] (for "separate"). After processing the two sentences, the final vector for the [CLS] token is passed to a linear layer for binary classification into [IsNext] and [NotNext]. For example: Given "[CLS] my dog is cute [SEP] he likes playing [SEP]", the model should predict [IsNext]. Given "[CLS] my dog is cute [SEP] how do magnets work [SEP]", the model should predict [NotNext]. === Fine-tuning === BERT is meant as a general pretrained model for various applications in natural language processing. That is, after pre-training, BERT can be fine-tuned with fewer resources on smaller datasets to optimize its performance on specific tasks such as natural language inference and text classification, and sequence-to-sequence-based language generation tasks such as question answering and conversational response generation. The original BERT paper published results demonstrating that a small amount of fine