Subject indexing is the act of describing or classifying a document by index terms, keywords, or other symbols in order to indicate what different documents are about, to summarize their contents or to increase findability. In other words, the objective is to identify and describe the subject of documents. Indexes are constructed, separately, on three distinct levels: terms in a document, such as a book; objects in a collection, such as a library; and documents (such as books and articles) within a field of knowledge. Subject indexing is used in information retrieval especially to create bibliographic indexes to retrieve documents on a particular subject. Examples of academic indexing services are Zentralblatt MATH, Chemical Abstracts, and PubMed. The index terms were mostly assigned by experts but author keywords are also common. The process of indexing begins with any analysis of the subject of the document. The indexer must then identify terms that appropriately identify the subject, either by extracting words directly from the document or assigning words from a controlled vocabulary. The terms in the index are then presented in a systematic order. Indexers must decide how many terms to include and how specific the terms should be. Together this gives a depth of indexing. == Subject analysis == The first step in indexing is to decide on the subject matter of the document. In manual indexing, the indexer would consider the subject matter in terms of answer to a set of questions such as "Does the document deal with a specific product, condition or phenomenon?". As the analysis is influenced by the knowledge and experience of the indexer, it follows that two indexers may analyze the content differently and so come up with different index terms. This will impact on the success of retrieval. === Automatic vs. manual subject analysis === Automatic indexing follows set processes of analyzing frequencies of word patterns and comparing results to other documents in order to assign to subject categories. This requires no understanding of the material being indexed. This leads to more uniform indexing but at the expense of the true meaning being interpreted. A computer program will not understand the meaning of statements and may therefore fail to assign some relevant terms or assign incorrectly. Human indexers focus their attention on certain parts of the document such as the title, abstract, summary and conclusions, as analyzing the full text in depth is costly and time-consuming. An automated system takes away the time limit and allows the entire document to be analyzed, but also has the option to be directed to particular parts of the document. == Term selection == The second stage of indexing involves the translation of the subject analysis into a set of index terms. This can involve extracting from the document or assigning from a controlled vocabulary. With the ability to conduct a full text search widely available, many people have come to rely on their own expertise in conducting information searches and full text search has become very popular. Subject indexing and its experts, professional indexers, catalogers, and librarians, remains crucial to information organization and retrieval. These experts understand controlled vocabularies and are able to find information that cannot be located by full text search. The cost of expert analysis to create subject indexing is not easily compared to the cost of hardware, software and labor to manufacture a comparable set of full-text, fully searchable materials. With new web applications that allow every user to annotate documents, social tagging has gained popularity especially in the Web. One application of indexing, the book index, remains relatively unchanged despite the information revolution. === Extraction/Derived indexing === Extraction indexing involves taking words directly from the document. It uses natural language and lends itself well to automated techniques where word frequencies are calculated and those with a frequency over a pre-determined threshold are used as index terms. A stop-list containing common words (such as "the", "and") would be referred to and such stop words would be excluded as index terms. Automated extraction indexing may lead to loss of meaning of terms by indexing single words as opposed to phrases. Although it is possible to extract commonly occurring phrases, it becomes more difficult if key concepts are inconsistently worded in phrases. Automated extraction indexing also has the problem that, even with use of a stop-list to remove common words, some frequent words may not be useful for allowing discrimination between documents. For example, the term glucose is likely to occur frequently in any document related to diabetes. Therefore, use of this term would likely return most or all the documents in the database. Post-coordinated indexing where terms are combined at the time of searching would reduce this effect but the onus would be on the searcher to link appropriate terms as opposed to the information professional. In addition terms that occur infrequently may be highly significant for example a new drug may be mentioned infrequently but the novelty of the subject makes any reference significant. One method for allowing rarer terms to be included and common words to be excluded by automated techniques would be a relative frequency approach where frequency of a word in a document is compared to frequency in the database as a whole. Therefore, a term that occurs more often in a document than might be expected based on the rest of the database could then be used as an index term, and terms that occur equally frequently throughout will be excluded. Another problem with automated extraction is that it does not recognize when a concept is discussed but is not identified in the text by an indexable keyword. Since this process is based on simple string matching and involves no intellectual analysis, the resulting product is more appropriately known as a concordance than an index. === Assignment indexing === An alternative is assignment indexing where index terms are taken from a controlled vocabulary. This has the advantage of controlling for synonyms as the preferred term is indexed and synonyms or related terms direct the user to the preferred term. This means the user can find articles regardless of the specific term used by the author and saves the user from having to know and check all possible synonyms. It also removes any confusion caused by homographs by inclusion of a qualifying term. A third advantage is that it allows the linking of related terms whether they are linked by hierarchy or association, e.g. an index entry for an oral medication may list other oral medications as related terms on the same level of the hierarchy but would also link to broader terms such as treatment. Assignment indexing is used in manual indexing to improve inter-indexer consistency as different indexers will have a controlled set of terms to choose from. Controlled vocabularies do not completely remove inconsistencies as two indexers may still interpret the subject differently. == Index presentation == The final phase of indexing is to present the entries in a systematic order. This may involve linking entries. In a pre-coordinated index the indexer determines the order in which terms are linked in an entry by considering how a user may formulate their search. In a post-coordinated index, the entries are presented singly and the user can link the entries through searches, most commonly carried out by computer software. Post-coordination results in a loss of precision in comparison to pre-coordination. == Depth of indexing == Indexers must make decisions about what entries should be included and how many entries an index should incorporate. The depth of indexing describes the thoroughness of the indexing process with reference to exhaustivity and specificity. === Exhaustivity === An exhaustive index is one which lists all possible index terms. Greater exhaustivity gives a higher recall, or more likelihood of all the relevant articles being retrieved, however, this occurs at the expense of precision. This means that the user may retrieve a larger number of irrelevant documents or documents which only deal with the subject in little depth. In a manual system a greater level of exhaustivity brings with it a greater cost as more man-hours are required. The additional time taken in an automated system would be much less significant. At the other end of the scale, in a selective index only the most important aspects are covered. Recall is reduced in a selective index as if an indexer does not include enough terms, a highly relevant article may be overlooked. Therefore, indexers should strive for a balance and consider what the document may be used. They may also have to consider the implications of time and expense. === Specificity === The specificity describes how closel
Dropbox Paper
Dropbox Paper, or simply Paper, is a collaborative document-editing service developed by Dropbox. Originating from the company's acquisition of document collaboration company Hackpad in April 2014, Dropbox Paper was officially announced in October 2015, and launched in January 2017. It offers a web application, as well as mobile apps for Android and iOS. Dropbox Paper was described in the official announcement post as "a flexible workspace that brings people and ideas together. With Paper, teams can create, review, revise, manage, and organize — all in shared documents". Reception of Dropbox Paper has been mixed. Critics praised collaboration functionality, including content available immediately, the ability to mention specific collaborators, assign tasks, write comments, as well as editing attribution, and revision history. It received particular praise for its support for rich media from a variety of sources, with one reviewer noting that the Paper's support for rich media exceeds the capabilities of most of its competitors. However, it was criticized for a lack of formatting options and editing features. While the user interface was liked for being minimal, reviewers cited the lack of a fixed formatting bar and missing features present in competitors' products as making Dropbox Paper seem like a "light" tool. == History == Dropbox acquired document collaboration company Hackpad in April 2014. A year later, Dropbox launched a Dropbox Notes note-taking product in beta testing phase. Dropbox Paper was officially announced on October 15, 2015, followed by an open beta and release of mobile Android and iOS apps in August 2016. Dropbox Paper was officially released on January 30, 2017. == Reception == In a comparison between Dropbox Paper and Evernote, PC World's Michael Ansaldo wrote that "With its emphasis on document creation, you might expect formatting to be front and center in Dropbox Paper. That's not the case." Ansaldo noted the lack of a "fixed formatting toolbar as you'd find in Evernote or a word processor like Google Docs or Microsoft Word. Instead, the text editor appears as a floating ribbon only when you highlight selected text." The only formatting options available for emphasis were bolding, strikethrough, bulleted and numbered lists, and H1 and H2 tags. Users can also add links, convert text to checklists, and add comments. Ansaldo wrote that "Both Evernote and Dropbox Paper make it easy to add images to a document", but also noted that "Dropbox Paper doesn't support any image editing". Paper supports rich media, and users can "add rich content to your document just by pasting a link to the file. In addition to Dropbox, Paper supports media from a variety of popular services including YouTube, Spotify, Vimeo, SoundCloud, Facebook, and Google's productivity suite. Once the file appears, you can delete the link for a cleaner display." To start working with other people, Paper "allows you to invite people via email from within a document", with sharing options for who can view the link (anyone with the link or just the invited person), and action permissions (edit or only comment). Regarding collaboration, Ansaldo wrote that "Creative collaboration is Paper’s marquee feature, and it provides a variety of ways to work effectively with others in real time". Users can "make any content immediately visible and accessible to a specific collaborator with "@mentions"", and "You can also use @mentions to create and assign task lists within a document." Paper also "boasts essential collaboration tools including comments, editing attribution, and revision history." Writing for TechRadar, John Brandon wrote that Dropbox Paper "might be a 'light' tool for now without the extensive templates of Microsoft Office or the integration with other apps in the Zoho suite, but it does work well with the Dropbox storage service that's so popular with office workers these days." Kyle Wiggers of Digital Trends wrote that Paper is "all about minimizing distractions. Its interface is quite literally a big, blank canvas on which you tap out your agenda. You can organize notes by title and create to-do lists, but even basic formatting tools are obscured from view", noting Paper's "floating box above words and phrases highlighted by your cursor". Wiggers stated that "Paper is not a to-do organizer", but that it's "well suited to the purpose thanks to a bevy of labor-saving conveniences", highlighting that Paper "supports more media than most of its to-do and note-taking counterparts". He praised the collaboration tools, writing that they "are as extensive as you'd hope, and then some", citing its invitation system with permission controls, lists of changes and revision history, comment and chat support, and "perhaps best of all", the ability to assign tasks with a "@" mention. Business Insider's Alex Heath praised that "Paper's interface is spotless and friendly to write in. You don't feel overwhelmed with formatting options", but criticized the available features, writing that "Google Docs is much more full-featured in the formatting department, so Paper has some catching up to do if it wants to be on par with the competition". Writing for The Verge, Casey Newton praised Paper's handling of rich media, complimenting it for being "great", and added that "I imagine that creative types who work on teams will appreciate having rich media embedded in the documents they're working on rather than in a series of infinite tabs".
Joseph Keshet
Joseph (Yossi) Keshet (Hebrew: יוסף (יוסי) קשת; born: 28 February 1973) is an Israeli professor in the Electrical and Computer Engineering Faculty of the Technion, where he is the director of the Speech, Language, and Deep Learning Lab. His research focuses on human speech processing and machine learning. == Early life and education == Keshet was born in Tel-Aviv. He graduated from the Amal School and began his academic studies at the Department of Electrical Engineering-Systems at Tel-Aviv University in 1991 and received his B.Sc. (Cum Laude) in 1994. Keshet served in the IDF Unit 8200 from 1995 to 2002 as the head of the speech processing research section in the R&D Center. During his service, he received a national award from the Administration for the Development of Weapons and Technological Infrastructure (Maf’at). Keshet was award his M.Sc. from the same department after he completed his Israel Defense Force service in 2002. His Dissertation was titled: Stop consonant spotting in continuous speech and was supervised by Dan Chazan from IBM Research Labs, Haifa. He continued his Ph.D. studies at the Hebrew University of Jerusalem until 2008. Prof. Yoram Singer supervised his thesis on Large Margin Algorithms for Discriminative Continuous Speech. == Career == Keshet was a Research Associate (postdoc) at IDIAP Research Institute, Martigny, Switzerland in 2007, and joined the TTI-Chicago and Department of Computer Science, University of Chicago, Chicago, IL in 2009 as Research Assistant Professor. In 2013, he returned to Israel and joined the Computer Science department at Bar-Ilan University as a senior lecturer and head of the Speech, Language, and Deep Learning Lab. In 2020, Keshet became a Founding Venture Partner at the Disruptive AI Venture Capital. In the same year, he also joined Amazon in Tel-Aviv as an Amazon Scholar. In 2022, Keshet joined the Faculty of Electrical and Computer Engineering at the Technion. == Research == Keshet's research work focuses on both machine learning and computational study of human speech and language. His work on speech and language concentrates on speech processing, speech recognition, acoustic phonetics, and pathological speech. In machine learning, Keshet is focused on deep learning and structured tasks. According to Google Scholar (September 2020), Keshet is one of the 15 most cited researchers in the field of spoken language processing. The algorithms that were developed in the Speech, Language, and Deep Learning Lab can analyze different pathological conditions in the throat and vocal cords based on the subject's voice. Other algorithms showed that the voice can be used to estimate physical and emotional state of the speaker. Another research led by Keshet suggested that it is possible to fool structured AI systems (like Google Voice). == Membership in professional societies == Keshet is the founder and chair of the Machine Learning for Speech and Language Processing Special Interest Group (SIGML) of the International Speech Communication Association (ISCA), from 2011. He is a senior member of the IEEE Signal Processing Society since 2018 and a member of ISCA since 2002. == Publications == Prof. Keshet has authored more than 70 scientific publications and edited one book. === Book === Joseph Keshet and Samy Bengio, Eds., Automatic Speech and Speaker Recognition: Large Margin and Kernel Methods, John Wiley & Sons, March 2009. === Selected articles === Jacob T. Cohen, Alma Cohen, Limor Benyamini, Yossi Adi, Joseph Keshet, Predicting glottal closure insufficiency using fundamental frequency contour analysis, Head & Neck, Journal of the Sciences and Specialities of the Head and Neck, Volume 41, Issue 7, pp. 2324–2331, July 2019. Yehoshua Dissen, Jacob Goldberger, and Joseph Keshet, Formant Estimation and Tracking: A Deep Learning Approach, Journal of the Acoustical Society of America, 145 (2), February 2019. Joseph Keshet, Automatic speech recognition: A primer for speech-language pathology researchers, International Journal of Speech-Language Pathology, Vol. 20 No. 6, pp. 599–609, 2018. Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, Joseph Keshet, Turning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdooring, Usenix, 2018. Tzeviya Fuchs, Joseph Keshet, Spoken Term Detection Automatically Adjusted for a Given Threshold, IEEE Journal of Selected Topics in Signal Processing, Dec 2017, Volume 11, Issue 8, pp. 1–8. Moustapha Cisse, Yossi Adi, Natalia Neverova, Joseph Keshet, Houdini: Fooling Deep Structured Visual and Speech Recognition Models with Adversarial Examples, Neural Information and Processing Systems (NIPS), 2017. Joseph Keshet, Subhransu Maji, Tamir Hazan, and Tommi Jaakkola, Perturbation Models and PAC-Bayesian Generalization Bounds, in Perturbations, Optimization, and Statistics, Tamir Hazan, George Papandreou, and Daniel Tarlow, Eds., The MIT Press, 2016. Matthew Goldrick, Joseph Keshet, Erin Gustafson, Jordana Heller, and Jeremy Needle, Automatic Analysis of Slips of the Tongue: Insights into the Cognitive Architecture of Speech Production, Cognition, 149, 31–39, 2016. Joseph Keshet, Optimizing the Measure of Performance in Structured Prediction, in Advanced Structured Prediction, Sebastian Nowozin, Peter V. Gehler, Jeremy January, and Christoph H. Lampert, Eds., The MIT Press, 2014. Morgan Sonderegger and Joseph Keshet, Automatic Measurement of Voice Onset Time using Discriminative Structured Prediction, Journal of the Acoustical Society of America, Vol. 132, Issue 6, pp. 3965−3979, 2012. David McAllester, Tamir Hazan and Joseph Keshet, Direct Loss Minimization for Structured Prediction, The 24th Annual Conference on Neural Information Processing Systems (NIPS), 2010. Joseph Keshet, David Grangier and Samy Bengio, Discriminative Keyword Spotting, Speech Communication, Volume 51, Issue 4, pp. 317–329, April 2009. == Personal life == Keshet is married to Lital. They have three children.
Maximum-entropy Markov model
In statistics, a maximum-entropy Markov model (MEMM), or conditional Markov model (CMM), is a graphical model for sequence labeling that combines features of hidden Markov models (HMMs) and maximum entropy (MaxEnt) models. An MEMM is a discriminative model that extends a standard maximum entropy classifier by assuming that the unknown values to be learnt are connected in a Markov chain rather than being conditionally independent of each other. MEMMs find applications in natural language processing, specifically in part-of-speech tagging and information extraction. == Model == Suppose we have a sequence of observations O 1 , … , O n {\displaystyle O_{1},\dots ,O_{n}} that we seek to tag with the labels S 1 , … , S n {\displaystyle S_{1},\dots ,S_{n}} that maximize the conditional probability P ( S 1 , … , S n ∣ O 1 , … , O n ) {\displaystyle P(S_{1},\dots ,S_{n}\mid O_{1},\dots ,O_{n})} . In a MEMM, this probability is factored into Markov transition probabilities, where the probability of transitioning to a particular label depends only on the observation at that position and the previous position's label: P ( S 1 , … , S n ∣ O 1 , … , O n ) = ∏ t = 1 n P ( S t ∣ S t − 1 , O t ) . {\displaystyle P(S_{1},\dots ,S_{n}\mid O_{1},\dots ,O_{n})=\prod _{t=1}^{n}P(S_{t}\mid S_{t-1},O_{t}).} Each of these transition probabilities comes from the same general distribution P ( s ∣ s ′ , o ) {\displaystyle P(s\mid s',o)} . For each possible label value of the previous label s ′ {\displaystyle s'} , the probability of a certain label s {\displaystyle s} is modeled in the same way as a maximum entropy classifier: P ( s ∣ s ′ , o ) = P s ′ ( s ∣ o ) = 1 Z ( o , s ′ ) exp ( ∑ a λ a f a ( o , s ) ) . {\displaystyle P(s\mid s',o)=P_{s'}(s\mid o)={\frac {1}{Z(o,s')}}\exp \left(\sum _{a}\lambda _{a}f_{a}(o,s)\right).} Here, the f a ( o , s ) {\displaystyle f_{a}(o,s)} are real-valued or categorical feature-functions, and Z ( o , s ′ ) {\displaystyle Z(o,s')} is a normalization term ensuring that the distribution sums to one. This form for the distribution corresponds to the maximum entropy probability distribution satisfying the constraint that the empirical expectation for the feature is equal to the expectation given the model: E e [ f a ( o , s ) ] = E p [ f a ( o , s ) ] for all a . {\displaystyle \operatorname {E} _{e}\left[f_{a}(o,s)\right]=\operatorname {E} _{p}\left[f_{a}(o,s)\right]\quad {\text{ for all }}a.} The parameters λ a {\displaystyle \lambda _{a}} can be estimated using generalized iterative scaling. Furthermore, a variant of the Baum–Welch algorithm, which is used for training HMMs, can be used to estimate parameters when training data has incomplete or missing labels. The optimal state sequence S 1 , … , S n {\displaystyle S_{1},\dots ,S_{n}} can be found using a very similar Viterbi algorithm to the one used for HMMs. The dynamic program uses the forward probability: α t + 1 ( s ) = ∑ s ′ ∈ S α t ( s ′ ) P s ′ ( s ∣ o t + 1 ) . {\displaystyle \alpha _{t+1}(s)=\sum _{s'\in S}\alpha _{t}(s')P_{s'}(s\mid o_{t+1}).} == Strengths and weaknesses == An advantage of MEMMs rather than HMMs for sequence tagging is that they offer increased freedom in choosing features to represent observations. In sequence tagging situations, it is useful to use domain knowledge to design special-purpose features. In the original paper introducing MEMMs, the authors write that "when trying to extract previously unseen company names from a newswire article, the identity of a word alone is not very predictive; however, knowing that the word is capitalized, that is a noun, that it is used in an appositive, and that it appears near the top of the article would all be quite predictive (in conjunction with the context provided by the state-transition structure)." Useful sequence tagging features, such as these, are often non-independent. Maximum entropy models do not assume independence between features, but generative observation models used in HMMs do. Therefore, MEMMs allow the user to specify many correlated, but informative features. Another advantage of MEMMs versus HMMs and conditional random fields (CRFs) is that training can be considerably more efficient. In HMMs and CRFs, one needs to use some version of the forward–backward algorithm as an inner loop in training. However, in MEMMs, estimating the parameters of the maximum-entropy distributions used for the transition probabilities can be done for each transition distribution in isolation. A drawback of MEMMs is that they potentially suffer from the "label bias problem," where states with low-entropy transition distributions "effectively ignore their observations." Conditional random fields were designed to overcome this weakness, which had already been recognised in the context of neural network-based Markov models in the early 1990s. Another source of label bias is that training is always done with respect to known previous tags, so the model struggles at test time when there is uncertainty in the previous tag.
Best AI Clip Makers in 2026
Trying to pick the best AI clip maker? An AI clip maker is software that uses machine learning to help you get more done — it scales effortlessly from a single task to thousands. The best picks balance beginner-friendly simplicity with the depth power users need, and they ship updates often. Whether you are a beginner or a pro, the right AI clip maker slots into your workflow and pays for itself fast. Read on for hands-on impressions, pricing tiers, and the standout features that matter.
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
Alex James (professor)
Alex James is an Indian scientist who is a professor of AI hardware at School of Electronic Systems and Automation, and Dean at Digital University Kerala (IIITM-K). He is the professor in charge of Maker Village, Kochi, Chief Investigator of the centre for Intelligent IoT Sensors, and India Innovation Centre for Graphene. James features in top 1% scientists list published by Elsevier BV in the world in the field of Electrical and Electronics Engineering. He appeared in the list for the third consecutive time. He specializes in the scientific field of Memristive Systems, AI hardware, Neuromorphic VLSI (very-large-scale integration) system, Intelligent Imaging and Machine learning, and Analogue electronics. == Education and career == James earned his Ph.D. degree from the Queensland Micro and Nanotechnology Centre, Griffith University, Brisbane, Australia. Since 2009, he has been working as a faculty member at different universities in Australia and India. He was a Member of IET Vision and Imaging Network, and is a Member of BCS’ Fellows Technical Advisory Group (F-TAG). He is the founding chair for IEEE Kerala Section Circuits and Systems Society, and is a fellow of British Computer Society (FBCS), and Institution of Engineering and Technology. He was an Editorial Board Member of Information Fusion (2010–2014), Elsevier, and associate editor for HCIS (2015–2020), Springer; and Guest Associate Editor for IEEE Transactions on Emerging Topics in Computational Intelligence (2017). Currently he is serving as an Associate Editor of IEEE Access, Frontiers in Neuroscience, and IEEE Transactions on Circuits and Systems I: Regular Papers journal. == Scientific research == IIITM-K has achieved a breakthrough in developing Analogue Integrated circuit for implementing Generative Adversarial Networks (GAN) in a joint research project with Analogue Circuits and Image Sensors Lab, Siegen university and Fraunhofer, Germany, and Centre for Excellence in Artificial general intelligence and Neuromorphic Systems (neuroAGI). According to A. P. James, professor at the School of Electronics at IIITM-K, this complicated and meticulous AI circuits research can accelerate and operate GAN applications in low power devices. It also can be used to analyze and interpret 2019-nCoV data for a possible solution to the pandemic. An AI Semantic search engine has been created by a research team led by A.P. James to help researchers gain deeper insights into Scientific Investigation, particularly since the COVID-19 issue has necessitated the collection of a significant amount of complex scientific data. The search engine is called "www.vilokana.in, which is Sanskrit for "finding out. == Awards and honors == James is a member of IEEE CASS Technical committee on Nonlinear Circuits and Systems, IEEE CASS Technical committee on Cellular Nanoscale networks and Memristor Array Computing, IEEE Consumer Technology Society Technical Committee on Quantum in Consumer Technology (QCT), Technical Committee on Machine learning, Deep learning and AI in CE (MDA) and Member of BCS’ Fellows Technical Advisory Group (F-TAG). James was awarded best associate editor of IEEE Transactions on Circuits and Systems I: Regular Papers TCAS-I, by the IEEE Circuits and Systems Society (IEEE CASS) for the year 2020–21. He has been an associate editor for the journal since 2017. He is also an editorial board member of PeerJ CS and a Senior Member of IEEE, Life Member of ACM, Senior Fellow of HEA.