AI Companion Zoom Logo

AI Companion Zoom Logo — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

  • Co-occurrence matrix

    Co-occurrence matrix

    A co-occurrence matrix or co-occurrence distribution (also referred to as : gray-level co-occurrence matrices GLCMs) is a matrix that is defined over an image to be the distribution of co-occurring pixel values (grayscale values, or colors) at a given offset. It is used as an approach to texture analysis with various applications especially in medical image analysis. == Method == Given a grey-level image I {\displaystyle I} , co-occurrence matrix computes how often pairs of pixels with a specific value and offset occur in the image. The offset, ( Δ x , Δ y ) {\displaystyle (\Delta x,\Delta y)} , is a position operator that can be applied to any pixel in the image (ignoring edge effects): for instance, ( 1 , 2 ) {\displaystyle (1,2)} could indicate "one down, two right". An image with p {\displaystyle p} different pixel values will produce a p × p {\displaystyle p\times p} co-occurrence matrix, for the given offset. The ( i , j ) th {\displaystyle (i,j)^{\text{th}}} value of the co-occurrence matrix gives the number of times in the image that the i th {\displaystyle i^{\text{th}}} and j th {\displaystyle j^{\text{th}}} pixel values occur in the relation given by the offset. For an image with p {\displaystyle p} different pixel values, the p × p {\displaystyle p\times p} co-occurrence matrix C is defined over an n × m {\displaystyle n\times m} image I {\displaystyle I} , parameterized by an offset ( Δ x , Δ y ) {\displaystyle (\Delta x,\Delta y)} , as: C Δ x , Δ y ( i , j ) = ∑ x = 1 n ∑ y = 1 m { 1 , if I ( x , y ) = i and I ( x + Δ x , y + Δ y ) = j 0 , otherwise {\displaystyle C_{\Delta x,\Delta y}(i,j)=\sum _{x=1}^{n}\sum _{y=1}^{m}{\begin{cases}1,&{\text{if }}I(x,y)=i{\text{ and }}I(x+\Delta x,y+\Delta y)=j\\0,&{\text{otherwise}}\end{cases}}} where: i {\displaystyle i} and j {\displaystyle j} are the pixel values; x {\displaystyle x} and y {\displaystyle y} are the spatial positions in the image I; the offsets ( Δ x , Δ y ) {\displaystyle (\Delta x,\Delta y)} define the spatial relation for which this matrix is calculated; and I ( x , y ) {\displaystyle I(x,y)} indicates the pixel value at pixel ( x , y ) {\displaystyle (x,y)} . The 'value' of the image originally referred to the grayscale value of the specified pixel, but could be anything, from a binary on/off value to 32-bit color and beyond. (Note that 32-bit color will yield a 232 × 232 co-occurrence matrix!) Co-occurrence matrices can also be parameterized in terms of a distance, d {\displaystyle d} , and an angle, θ {\displaystyle \theta } , instead of an offset ( Δ x , Δ y ) {\displaystyle (\Delta x,\Delta y)} . Any matrix or pair of matrices can be used to generate a co-occurrence matrix, though their most common application has been in measuring texture in images, so the typical definition, as above, assumes that the matrix is an image. It is also possible to define the matrix across two different images. Such a matrix can then be used for color mapping. == Aliases == Co-occurrence matrices are also referred to as: GLCMs (gray-level co-occurrence matrices) GLCHs (gray-level co-occurrence histograms) spatial dependence matrices == Application to image analysis == Whether considering the intensity or grayscale values of the image or various dimensions of color, the co-occurrence matrix can measure the texture of the image. Because co-occurrence matrices are typically large and sparse, various metrics of the matrix are often taken to get a more useful set of features. Features generated using this technique are usually called Haralick features, after Robert Haralick. Texture analysis is often concerned with detecting aspects of an image that are rotationally invariant. To approximate this, the co-occurrence matrices corresponding to the same relation, but rotated at various regular angles (e.g. 0, 45, 90, and 135 degrees), are often calculated and summed. Texture measures like the co-occurrence matrix, wavelet transforms, and model fitting have found application in medical image analysis in particular. == Other applications == Co-occurrence matrices are also used for words processing in natural language processing (NLP).

    Read more →
  • Ancient text corpora

    Ancient text corpora

    Ancient text corpora are the entire collection of texts from the period of ancient history, defined in this article as the period from the beginning of writing up to 300 AD. These corpora are important for the study of literature, history, linguistics, and other fields, and are a fundamental component of the world's cultural heritage. Chinese, Latin, and Greek are examples of ancient languages with significant text corpora, although much of these corpora are known to us via transmission (frequently via medieval manuscript copies) rather than in their original form. These texts – both transmitted and original – provide valuable insights into the history and culture of different regions of the world, and have been studied for centuries by scholars and researchers. Other ancient texts – particularly stone inscriptions and papyrus scrolls – have been published following archaeological research, notably the cuneiform corpus of c.10 million words and the c.5 million words in ancient Egyptian. Through advances in technology and digitization, ancient text corpora are more accessible than ever before. Tools such as the Perseus Digital Library and the Digital Corpus of Sanskrit have made it easier for researchers to access and analyze these texts. == Quantifying the corpora == Two types of ancient texts are known to modern scholars – those that have only survived in younger manuscripts, but whose great age is undisputed (this applies to the bulk of the Chinese, Brahmi, Greek, Latin, Hebrew and Avestan tradition), and those known from original inscriptions, papyri and other manuscripts. Counting of the words in each corpus presents significant methodological challenges – in principle, every single occurrence of a word in the text is counted separately, but in the case of parallel transmission of literary texts, only a single transmission is taken into account. Just as the Book of the Dead and the coffin texts are only included once in the number given for the Egyptian, the Greek and Latin literary works should only be counted according to one manuscript. If, on the other hand, tombs, royal inscriptions or economic documents of certain ancient languages often show a more or less identical form, this is not evaluated as a purely "parallel tradition". Attached prepositions are counted as separate words, except in the case of the definite article in Hebrew, Aramaic and Greek since it has no equivalent in most languages, so its frequency would significantly affect the comparability of numbers. === Languages with known size estimates === === South Asian === Sanskrit (Vedic Sanskrit and Classical Sanskrit) Indus script (3,800 items, c.20,000 characters) Brahmi script Old Tamil Early Indian epigraphy and Indian epic poetry Kharosthi Pali literature List of historic Indian texts === Mesoamerican === Olmec hieroglyphs Maya script === East Asian === Old Chinese Chinese classics The pre-Qin corpus: a collection of ancient Chinese texts written before the Qin dynasty (221 BCE). The corpus includes texts from Confucianism, Taoism, Legalism, and other schools of thought. The pre-Han corpus: a collection of ancient Chinese texts written before the Han dynasty (202 BCE). The corpus includes texts from Confucianism, Taoism, Legalism, and other schools of thought. See the Chinese Text Project Chinese bronze inscriptions, Oracle bone script, Seal script, Clerical script === Central Iranian languages === Prior to 300 AD, the Central Iranian languages are mainly in the form of Sassanid stone inscriptions in the two closely related idioms Middle Persian (Pahlavi scripts and Inscriptional Parthian), there are 5000 for the corpus of Middle Persian (mostly 3rd, but also 4th/5th centuries) and for the corpus of Parthian (3rd century) 3000 words. To what extent some of the Manichaean Middle Persian literary texts may date back to the 3rd century is difficult to estimate; Mani is said to have personally written the Shabuhragan totaling about 5000 words. In any case, if we combine Middle Persian and Parthian, we come to over 10,000 words. === Proto-Sinaitic === Proto-Sinaitic script has no more than about 400 letters (number of words is unknown since the script has not been fully interpreted). To a similar extent, there are probably approximately contemporaneous Proto-Canaanite inscriptions (ibid.). === Anatolian === Luwian cuneiform, approx. 3000 words the Palaic language few hundred words. Hieroglyphic Luwian the Lycian alphabet (the best attested Anatolian successor language written in alphabetic script) with about 5000 words The Lydian alphabet 109 inscriptions comprising about 1500 words The Phrygian alphabet the in-tomb inscriptions from the 2nd and 3rd centuries AD (approx. 1000 words) and in the so-called "old Phrygian" inscriptions less than 300 words The Carian alphabets whose texts, mainly from Egypt, contain around 600 words. === Old Italic === the Umbrian language attested essentially by the sacrificial instructions of the Iguvinian Tables with 5000 words the Oscan language (ibid.) with 2000 words the Messapic language with probably a good 1000 words (the estimate is difficult because most texts in this hardly understandable language do not use word separators) the Venetic language a few hundred words the Faliscan language a few hundred words Cisalpine Celtic inscriptions amount to approximately 2000 words, to which are added a number of glosses by classical authors === Iberia === Iberian scripts, more rarely written in Greek or Latin script, approx. 2500 words Celtiberian script, which refers to Celtic language testimonies in Iberian, but also in Latin script from Spain (approx. 1000 words) Southwest Paleohispanic script, 78 inscriptions, a few hundred words Lusitanian language, three monuments in Latin script, approx. 60 words === Germanic Northern Europe === Runic inscriptions dated before the 4th century amount to about 30 pieces, which contain no more than 50 words in total === Africa === Geʽez script: comparatively few inscriptions with a total of around 1,000 words before 300 AD. Following Christianization in the 4th century, more extensive texts are known. Libyco-Berber alphabet: over 1,000 inscriptions from the Maghreb, which are dated to Roman times. Most texts do not use a word separator; Peust estimates that the total number of words could be around 5,000 Meroitic script (Ancient Nubian): about 900 texts are known, which Peust estimates may contain approximately 10,000 words, albeit with uncertainty from the fact that the word separator is not used consistently in the Meroitic script. === Aegean === The Cretan Linear A inscriptions that have not yet been deciphered are available in about 2500 texts, which contain a total of around 20,000 characters. The total number of words can hardly be determined; Peust tentatively put it in the same order of magnitude as in Meroitic. In addition to the Linear A texts, there are also inscriptions Cretan hieroglyphs of a few hundred characters and texts written in the Greek alphabet, but not in Greek, with a few dozen words Cypriot syllabary in the first millennium BC, in which mostly Greek texts were recorded. The relevant texts comprise around 100 to 200 words. === Micro corpora === There are a significant number of ancient micro-corpus languages. Estimating the total number of attested ancient languages may be as difficult as estimating their corpus size. For example, Greek and Latin sources hand down an enormous amount of foreign-language glosses, the seriousness of which is not always certain. == Preservation and curation == Historic preservation and maintaining ancient text corpora presents several challenges, including issues with preservation, translation, and digitization. Many ancient texts have been lost over time, and those that survive may be damaged or fragmented. Translating ancient languages and scripts requires specialized expertise, and digitizing texts can be time-consuming and resource-intensive. == Corpus linguistics == The field of corpus linguistics studies language as expressed in text corpora. This includes the analysis of word frequency, collocations, grammar, and semantics. Ancient text corpora provide a valuable resource for corpus linguistics research, enabling scholars to explore the evolution of language and culture over time.

    Read more →
  • Distributional–relational database

    Distributional–relational database

    A distributional–relational database, or word-vector database, is a database management system (DBMS) that uses distributional word-vector representations to enrich the semantics of structured data. As distributional word-vectors can be built automatically from large-scale corpora, this enrichment supports the construction of databases which can embed large-scale commonsense background knowledge into their operations. Distributional-Relational models can be applied to the construction of schema-agnostic databases (databases in which users can query the data without being aware of its schema), semantic search, schema-integration and inductive and abductive reasoning as well as different applications in which a semantically flexible knowledge representation model is needed. The main advantage of distributional–relational models over purely logical or semantic web models is the fact that the core semantic associations can be automatically captured from corpora, in contrast to the definition of manually curated ontologies and rule knowledge bases. == Distributional–relational models == Distributional–relational models were first formalized as a mechanism to cope with the vocabulary/semantic gap between users and the schema behind the data. In this scenario, distributional semantic relatedness measures, combined with semantic pivoting heuristics can support the approximation between user queries (expressed in their own vocabulary), and data (expressed in the vocabulary of the designer). In this model, the database symbols (entities and relations) are embedded into a distributional semantic space and have a geometric interpretation under a latent or explicit semantic space. The geometric aspect supports the semantic approximation between entities from different databases, or between a query term and a database entity. The distributional relational model then becomes a double layered model where the semantics of the structured data provides the fine-grained semantics intended by the database designer, which is extended by the distributional semantic model which contains the semantic associations expressed at a broader use. These models support the generalization from a closed communication scenario (in which database designers and users live in the same context, e.g. the same organization) to an open communication scenario (e.g. different organizations, the Web), creating an abstraction layer between users and the specific representation of the conceptual model.

    Read more →
  • Xu Li (computer scientist)

    Xu Li (computer scientist)

    Xu Li is a Chinese computer scientist and co-founder and current CEO of SenseTime, an artificial intelligence (AI) company. Xu has led SenseTime since the company's incorporation and helped it independently develop its proprietary deep learning platform. == Education and research == Xu obtained both his bachelor's and master's degrees in computer science from Shanghai Jiao Tong University. He received his doctorate in computer science from the Chinese University of Hong Kong. Xu has published more than 50 papers at international conferences and in journals in the field of computer vision and won the Best Paper Award at the international conference on Non-Photorealistic Rendering and Animation (NPAR) 2012 and the Best Reviewer Award at the international conferences Asian Conference on Computer Vision ACCV 2012 and International Conference on Computer Vision (ICCV) 2015. He has three algorithms that have been included into the visual open-source platform OpenCV, and his "L0 Smoothing" algorithm garnered the most citations in research papers over a span of five years (2011–2015) within the ACM Transactions on Graphics (TOG), a scientific journal that Thomson Reuters InCites has placed first among software engineering journals. == Career == Previously, Xu worked at Lenovo Corporate Research & Development. He was also a visiting researcher at Motorola China R&D Institute, Omron Research Institute, and Microsoft Research. == Selected publications == Jimmy Ren, Xiaohao Chen, Jianbo Liu, Wenxiu Sun, Li Xu, Jiahao Pang, Qiong Yan, Yu-wing Tai, "Accurate Single Stage Detector Using Recurrent Rolling Convolution", (CVPR), 2017. Jimmy SJ. Ren, Yongtao Hu, Yu-Wing Tai, Chuan Wang, Li Xu, Wenxiu Sun, Qiong Yan, "Look, Listen and Learn – A Multimodal LSTM for Speaker Identification", The 30th AAAI Conference on Artificial Intelligence (AAAI), 2016 Jimmy SJ. Ren, Li Xu, Qiong Yan, Wenxiu Sun, "Shepard Convolutional Neural Networks" Advances in Neural Information Processing Systems (NIPS), 2015. Xiaoyong Shen, Chao Zhou, Li Xu, Jiaya Jia, "Mutual-Structure for Joint Filtering" International Conference on Computer Vision (ICCV), (oral presentation), 2015. Jianping Shi, Qiong Yan, Li Xu, Jiaya Jia, "Hierarchical Image Saliency Detection on Extended CSSD" IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2015. Jianping Shi, Xin Tao, Li Xu, Jiaya Jia, "Break Ames Room Illusion: Depth from General Single Images" ACM Transactions on Graphics (TOG), (Proc. ACM SIGGRAPH ASIA2015). Yongtao Hu, Jimmy SJ. Ren, Jingwen Dai, Chang Yuan, Li Xu, Wenping Wang, "Deep Multimodal Speaker Naming" ACM International Conference on Multimedia (MM), 2015. Li Xu, Jimmy SJ. Ren, Qiong Yan, Renjie Liao, Jiaya Jia "Deep Edge-Aware Filters" International Conference on Machine Learning (ICML), 2015. Jianping Shi, Li Xu, Jiaya Jia "Just Noticeable Defocus Blur Detection and Estimation" IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. Ziyang Ma, Renjie Liao, Xin Tao, Li Xu, Jiaya Jia, Enhua Wu "Handling Motion Blur in Multi-Frame Super-Resolution" IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. Xiaoyong Shen, Qiong Yan, Li Xu, Lizhuang Ma, Jiaya Jia"Multispectral Joint Image Restoration via Optimizing a Scale Map" IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2015. Jimmy SJ. Ren, Li Xu, "On Vectorization of Deep Convolutional Neural Networks for Vision Tasks" AAAI Conference on Artificial Intelligence (AAAI), 2015. == Awards and honors == Xu was ranked 7th in Fortune magazine's 2018 edition of its 40 Under 40. He was also named "China's Outstanding AI Industry Leader" by The Economic Observer, received the "Innovative Business Leader" Award under NetEase's "Future Technology Talent Awards", and was honored as Sina's "2017 Top Ten Economic Figures". In 2018, Xu was named EY's "Entrepreneur of the Year China" in the Technology category.

    Read more →
  • Tom's Planner

    Tom's Planner

    Tom's Planner is a web-based tool and application service provider for project planning, management and collaboration. == History == Tom's Planner is based on Curaçao. In November 2009, it announced its public beta launch on TechCrunch and moved out of beta in August 2010. In 2013 Tom's Planner acquired its competitor Gantto. == Software == Tom's Planner is project management software that enables the creation of project schedules (Gantt charts) using a visual perspective. Tom's Planner uses the Freemium Business Model. Users can register for a free account or choose a paid version. Tom's Planner is available in five languages and is used by thousands of users on a daily basis in more than 100 countries worldwide. Customers range from fortune 500 companies to small mom-and-pop shops. == Reviews == Tom's Planner has been reviewed by PC World, TechCrunch, Lifehacker, and several other periodicals.

    Read more →
  • Raymond J. Mooney

    Raymond J. Mooney

    Raymond J. Mooney is an American computer scientist, professor of computer science, and director of the Artificial Intelligence laboratory at the University of Texas at Austin. His research focuses on machine learning and natural language processing. He was educated at O'Fallon Township High School in O'Fallon, Illinois and earned a BS, MS, and Ph.D. in computer science at the University of Illinois at Urbana-Champaign, where he was advised by Gerald DeJong. He is a fellow of the Association for Computing Machinery (ACM), Association for Computational Linguistics (ACL), and Association for the Advancement of Artificial Intelligence (AAAI).

    Read more →
  • Liz Liddy

    Liz Liddy

    Elizabeth DuRoss Liddy (May 12, 1944 – August 21, 2025) was an American computer scientist and academic who was professor of information science and dean of the Syracuse University School of Information Studies. She was a pioneer in the field of natural language processing. == Early life and education == Liddy was born in Dayton, Ohio, on May 14, 1944, and grew up in Utica, New York. She was one of five children, all of whom worked in her father's family business. Liddy attended St. Francis DeSalle High School, where she was awarded a Regent's Scholarship, and eventually attended Daemen College. She was literary editor of her high school year book and edited a literary magazine during her time at college. At Daemen College Liddy studied English language and literature. After graduating Liddy remained in New York, where she volunteered in an elementary school library. She joined the Syracuse University School of Information Studies in 1983, where she started a graduate program in library science. She worked as a faculty librarian at Onondaga Community College whilst earning her degree. Here Liddy worked as a Visiting assistant professor, whilst completing her doctorate part-time in information transfer. Her dissertation research involved natural language processing, a computerized approach to analyzing text. She was hired to the faculty at Syracuse University whilst completing her PhD. == Research and career == In 1994 Liddy was the founding President of TextWise, a semantics-based search engine. The first product she developed was called Document Retrieval Using Linguistic Knowledge (DR-LINK). She left TextWise in 1999, after growing the number of employees to over 50. She started the Syracuse University Center for Natural Language Processing in 1999, and was honored with the university's Outstanding Alumni Award the following year. Liddy was appointed Dean of the School of Information Studies (iSchool) in 2008, and held the position for over ten years. She temporarily left the role in 2015. The school was transformed under her leadership, increasing the enrollment of students by over 70% and launching a graduate certificate in data science. She raised over $20 million to support research and development at Syracuse University. She chaired the iSchool Organization, which connects information science schools all over the world, from 2012 to 2014. Liddy worked to increase the representation of women at the iSchool, through initiatives such as the IT Girls Overnight Retreat – an annual weekend to introduce high school girls to Information Technology. She improved the career development programs of students at Syracuse University, increasing student employment to almost 100% post graduation. Liddy retired as Dean of the iSchool in 2019. === Selected innovations === US 6026388, Liddy, Elizabeth D., "User interface and other enhancements for natural language information retrieval system and method", published August 16, 1995, issued February 15, 2000 US 5963940, Liddy, Elizabeth D., "Natural language information retrieval system and method", published August 16, 1995, issued October 5, 1999 US 6006221, Liddy, Elizabeth D., "Multilingual document retrieval system and method using semantic vector matching", published August 16, 1995, issued December 21, 1999 == Personal life and death == Liddy was married shortly after graduating Daemen College in 1966. She had three children. Liddy died in Charlotte, North Carolina, on August 21, 2025, at the age of 81.

    Read more →
  • Hidden Markov model

    Hidden Markov model

    A hidden Markov model (HMM) is a Markov model in which the observations are dependent on a latent (or hidden) Markov process (referred to as X {\displaystyle X} ). An HMM requires that there be an observable process Y {\displaystyle Y} whose outcomes depend on the outcomes of X {\displaystyle X} in a known way. Since X {\displaystyle X} cannot be observed directly, the goal is to learn about state of X {\displaystyle X} by observing Y {\displaystyle Y} . By definition of being a Markov model, an HMM has an additional requirement that the outcome of Y {\displaystyle Y} at time t = t 0 {\displaystyle t=t_{0}} must be "influenced" exclusively by the outcome of X {\displaystyle X} at t = t 0 {\displaystyle t=t_{0}} and that the outcomes of X {\displaystyle X} and Y {\displaystyle Y} at t < t 0 {\displaystyle t

  • Israeli cybersecurity industry

    Israeli cybersecurity industry

    The Israeli cybersecurity industry is a rapidly growing sector within Israel's technology and innovation ecosystem. Israel is internationally recognized as a powerhouse in the cybersecurity domain, with numerous cybersecurity startups, established companies, research institutions, and government initiatives. Tel Aviv itself is being ranked 7th in annual list of best global tech ecosystems, as reported by the Jerusalem Post. == History == The roots of Israel's cybersecurity industry can be traced back to the country's strong focus on national security and intelligence. The establishment of elite military units such as Unit 8200, the Israeli Intelligence Corps unit responsible for signals intelligence and code decryption, played a significant role in the development of cybersecurity expertise in the country. Many former members of Unit 8200 have gone on to establish successful cybersecurity companies or join existing organizations, bringing their unique skill sets and experience to the private sector. == Market overview == As of 2024, Israel housed more than 450 cybersecurity startups and companies. In 2023, the value of exits by Israeli tech companies reached $7.5 billion. Israel's cybersecurity industry is characterized by a high concentration of startups develop new technologies in areas such as network security, endpoint protection, data security, cloud security, and threat intelligence. In recent years, the sector has attracted significant investment from both local and international venture capital firms, as well as major technology companies such as Microsoft, Google, and IBM. Several Israeli cybersecurity companies have gained global recognition and success, with some being acquired by major corporations or conducting successful initial public offerings (IPOs). === Key Israeli cybersecurity companies === Some key Israeli cybersecurity companies include: Check Point Software Technologies CyberArk Cato Networks Radware Wiz === Financial activity === Israel’s cybersecurity sector has seen significant financial activity. As of 2023, mergers and acquisitions in the cybersecurity sector totaled $2.8 billion. In the first quarter of 2024, the sector secured $846 million in private funding. == Background == The military experience helped much. Israel's mandatory military service, combined with the expertise developed within elite units such as Unit 8200, has fostered a strong talent pool with practical experience in cybersecurity. Israel's thriving startup ecosystem, often referred to as the "Startup Nation," has fostered an environment of innovation and collaboration that has contributed to the growth of the cybersecurity industry. Israeli cybersecurity companies often collaborate with international partners, both in the private and public sectors, to share knowledge and develop joint solutions. === Government Initiatives and Support === The government also supported well through various initiatives, such as the Israel National Cyber Directorate (INCD), which works to strengthen cybersecurity defenses and promote the development of the sector. === Academic institutions === Israeli universities and research centers are involved in cybersecurity research and education, contributing to the development of new technologies and training the next generation of cybersecurity professionals. Academic Tech transfer offices in Israel also facilitate the commercialization of cybersecurity technologies. Some academic institutions with cybersecurity laboratories include: Tel Aviv University Technion Ben-Gurion University

    Read more →
  • Regina Barzilay

    Regina Barzilay

    Regina Barzilay (Hebrew: רגינה ברזילי; born 1970) is an Israeli-American computer scientist. She is a professor at the Massachusetts Institute of Technology and a faculty lead for artificial intelligence at the MIT Jameel Clinic. Her research interests are in natural language processing and applications of deep learning to chemistry and oncology. == Early life and education == Barzilay was born in Chișinău, Moldova and emigrated to Israel with her parents at the age of 20. She received bachelor's and master's degrees from Ben-Gurion University of the Negev in 1993 and 1998, respectively. She obtained a PhD in computer science from Columbia University in 2003 for research supervised by Kathleen McKeown. == Career and research == After her PhD, she spent a year as a postdoctoral researcher at Cornell University. She was appointed as Delta Electronics Professor of Electrical Engineering and Computer Science at MIT in 2016. She was diagnosed with breast cancer in 2014, which prompted her to conduct research in oncology. Barzilay won the MacArthur Fellowship in 2017. For her doctoral dissertation at Columbia University, she led the development of Newsblaster, which recognized stories from different news sources as being about the same basic subject, and then paraphrased elements from the stories to create a summary. In computational linguistics, Barzilay created algorithms that learned annotations from common languages (i.e. English) to analyze less understood languages. Prompted by her experience with breast cancer, Barzilay is applying machine learning to oncology. She is collaborating with physicians and students to devise deep learning models that utilize images, text, and structured data to identify trends that affect early diagnosis, treatment, and disease prevention. Frontline Documentary Following her battle with breast cancer in 2014, and her researching into applying artificial intelligence to improve early detection methods, she collaborated with Dr. Connie Lehman at Massachusetts General Hospital. While there Barzilay developed an AI-based system capable of predicting the likelihood of breast cancer up to five years in advance. The system leverages deep learning techniques to analyze mammograms and diagnostic notes, surpassing traditional pattern recognition by human radiologists. This breakthrough, while still in development, has the potential to significantly enhance early diagnosis and treatment outcomes. [1] Barzilay's work in this area was featured in the FRONTLINE documentary In the Age of AI, which explores the broader impact of artificial intelligence on society. === MIT Jameel Clinic === In 2018, Barzilay was appointed faculty lead for AI at the new MIT Jameel Clinic, a research center in the field of AI health sciences, including disease detection, drug discovery, and the development of medical devices. In 2020, she was part of the team—with fellow MIT Jameel Clinic faculty lead Professor James J. Collins—that announced the discovery through deep learning of halicin, the first new antibiotic compound for 30 years, which kills over 35 powerful bacteria, including antimicrobial-resistant tuberculosis, the superbug C. difficile, and two of the World Health Organization's top-three most deadly bacteria. In 2020, Collins, Barzilay and the MIT Jameel Clinic were also awarded funding through The Audacious Project to expand on the discovery of halicin in using AI to respond to the antibiotic resistance crisis through the development of new classes of antibiotics. == Awards and recognition == In 2017, Barzilay won the MacArthur Fellowship, known as the "Genius Grant", for "developing machine learning methods that enable computers to process and analyze vast amounts of human language data." She is also a recipient of various awards including the NSF Career Award, the MIT Technology Review TR-35 Award, Microsoft Faculty Fellowship and several Best Paper Awards at NAACL and ACL. Her teaching has also been recognized by MIT as she won the Jamieson Teaching Award in 2016. She was nominated an AAAI Fellow in 2018 by the Association for the Advancement of Artificial Intelligence. In 2020, she became the first recipient of the $1 million AAAI Squirrel AI Award for Artificial Intelligence for the Benefit of Humanity. In 2023, she was elected to the National Academy of Medicine and the National Academy of Engineering.

    Read more →
  • How to Choose an AI Text-to-image Tool

    How to Choose an AI Text-to-image Tool

    Curious about the best AI text-to-image tool? An AI text-to-image tool is software that uses machine learning to help you get more done — it combines speed, accuracy, and an interface that just works. Hands-on testing shows real-world results vary, so a short free trial is the smartest way to decide. Whether you are a beginner or a pro, the right AI text-to-image tool slots into your workflow and pays for itself fast. Read on for hands-on impressions, pricing tiers, and the standout features that matter.

    Read more →
  • Mirella Lapata

    Mirella Lapata

    Mirella Lapata is a computer scientist and Professor in the School of Informatics at the University of Edinburgh. Working on the general problem of extracting semantic information from large bodies of text, Lapata develops computer algorithms and models in the field of natural language processing (NLP). == Education == Lapata obtained a Master of Arts (MA) degree from Carnegie Mellon University and subsequently earned a doctorate from the University of Edinburgh. Lapata's doctoral research investigated the acquisition of information from polysemous linguistic units using probabilistic methods supervised by Alex Lascarides, Chris Brew and Steve Finch. == Career and research == After her doctorate, Lapata assumed academic positions at Saarland University and at the Department of Computer Science at the University of Sheffield. At the University of Edinburgh she became a reader in the School of Informatics where she is a full Professor and holds a personal chair in natural language processing. Lapata is a member of the Human Communication Research Center and Institute for Language, Cognition and Computation, both in Edinburgh. Between 2015 and 2017, Lapata served as a member of the Royal Society Machine Learning Working Group. Recently Lapata was granted a European Research Council (ERC) Consolidator Grant worth €1.9M to fund five years of her project, TransModal: Translating from Multiple Modalities into Text. === Awards and honours === In 2009 Lapata became the first recipient of the Microsoft British Computer Society (BCS)/BCS IRSG Karen Spärck Jones Award. The award recognises achievement in furthering the progress in information retrieval and natural language processing; the award commemorates the life and work of Karen Spärck Jones. In 2012 Lapata won an Empirical Methods in Natural Language Processing (EMNLP)-CoNLL 2012 Best Reviewer Award. In 2018 Lapata was awarded, alongside Li Dong, an Association for Computational Linguistics (ACL) Best Paper Honorable Mention. In 2019 Lapata was elected a Fellow of the Royal Society of Edinburgh In 2020 Lapata was elected to the Academia Europaea. In 2025 Lapata was awarded the BCS Lovelace Medal for Computing Research.

    Read more →
  • Motor theory of speech perception

    Motor theory of speech perception

    The motor theory of speech perception is the hypothesis that people perceive spoken words by identifying the vocal tract gestures with which they are pronounced rather than by identifying the sound patterns that speech generates. It originally claimed that speech perception is done through a specialized module that is innate and human-specific. Though the idea of a module has been qualified in more recent versions of the theory, the idea remains that the role of the speech motor system is not only to produce speech articulations but also to detect them. The hypothesis has gained more interest outside the field of speech perception than inside. This has increased particularly since the discovery of mirror neurons that link the production and perception of motor movements, including those made by the vocal tract. The theory was initially proposed in the Haskins Laboratories in the 1950s by Alvin Liberman and Franklin S. Cooper, and developed further by Donald Shankweiler, Michael Studdert-Kennedy, Ignatius Mattingly, Carol Fowler and Douglas Whalen. == Origins and development == The hypothesis has its origins in research using pattern playback to create reading machines for the blind that would substitute sounds for orthographic letters. This led to a close examination of how spoken sounds correspond to the acoustic spectrogram of them as a sequence of auditory sounds. This found that successive consonants and vowels overlap in time with one another (a phenomenon known as coarticulation). This suggested that speech is not heard like an acoustic "alphabet" or "cipher," but as a "code" of overlapping speech gestures. === Associationist approach === Initially, the theory was associationist: infants mimic the speech they hear and that this leads to behavioristic associations between articulation and its sensory consequences. Later, this overt mimicry would be short-circuited and become speech perception. This aspect of the theory was dropped, however, with the discovery that prelinguistic infants could already detect most of the phonetic contrasts used to separate different speech sounds. === Cognitivist approach === The behavioristic approach was replaced by a cognitivist one in which there was a speech module. The module detected speech in terms of hidden distal objects rather than at the proximal or immediate level of their input. The evidence for this was the research finding that speech processing was special such as duplex perception. === Changing distal objects === Initially, speech perception was assumed to link to speech objects that were both the invariant movements of speech articulators the invariant motor commands sent to muscles to move the vocal tract articulators This was later revised to include the phonetic gestures rather than motor commands, and then the gestures intended by the speaker at a prevocal, linguistic level, rather than actual movements. === Modern revision === The "speech is special" claim has been dropped, as it was found that speech perception could occur for nonspeech sounds (for example, slamming doors for duplex perception). === Mirror neurons === The discovery of mirror neurons has led to renewed interest in the motor theory of speech perception, and the theory still has its advocates, although there are also critics. == Support == === Nonauditory gesture information === If speech is identified in terms of how it is physically made, then nonauditory information should be incorporated into speech percepts even if it is still subjectively heard as "sounds". This is, in fact, the case. The McGurk effect shows that seeing the production of a spoken syllable that differs from an auditory cue synchronized with it affects the perception of the auditory one. In other words, if someone hears "ba" but sees a video of someone pronouncing "ga", what they hear is different—some people believe they hear "da". People find it easier to hear speech in noise if they can see the speaker. People can hear syllables better when their production can be felt haptically. === Categorical perception === Using a speech synthesizer, speech sounds can be varied in place of articulation along a continuum from /bɑ/ to /dɑ/ to /ɡɑ/, or in voice onset time on a continuum from /dɑ/ to /tɑ/ (for example). When listeners are asked to discriminate between two different sounds, they perceive sounds as belonging to discrete categories, even though the sounds vary continuously. In other words, 10 sounds (with the sound on one extreme being /dɑ/ and the sound on the other extreme being /tɑ/, and the ones in the middle varying on a scale) may all be acoustically different from one another, but the listener will hear all of them as either /dɑ/ or /tɑ/. Likewise, the English consonant /d/ may vary in its acoustic details across different phonetic contexts (the /d/ in /du/ does not technically sound the same as the one in /di/, for example), but all /d/'s as perceived by a listener fall within one category (voiced alveolar plosive) and that is because "linguistic representations are abstract, canonical, phonetic segments or the gestures that underlie these segments." This suggests that humans identify speech using categorical perception, and thus that a specialized module, such as that proposed by the motor theory of speech perception, may be on the right track. === Speech imitation === If people can hear the gestures in speech, then the imitation of speech should be very fast, as in when words are repeated that are heard in headphones as in speech shadowing. People can repeat heard syllables more quickly than they would be able to produce them normally. === Speech production === Hearing speech activates vocal tract muscles, and the motor cortex and premotor cortex. The integration of auditory and visual input in speech perception also involves such areas. Disrupting the premotor cortex disrupts the perception of speech units such as plosives. The activation of the motor areas occurs in terms of the phonemic features which link with the vocal track articulators that create speech gestures. The perception of a speech sound is aided by pre-emptively stimulating the motor representation of the articulators responsible for its pronunciation . Auditory and motor cortical coupling is restricted to a specific range of neuronal firing frequency. === Perception-action meshing === Evidence exists that perception and production are generally coupled in the motor system. This is supported by the existence of mirror neurons that are activated both by seeing (or hearing) an action and when that action is carried out. Another source of evidence is that for common coding theory between the representations used for perception and action. == Criticisms == The motor theory of speech perception is not widely held in the field of speech perception, though it is more popular in other fields, such as theoretical linguistics. As three of its advocates have noted, "it has few proponents within the field of speech perception, and many authors cite it primarily to offer critical commentary".p. 361 Several critiques of it exist. === Multiple sources === Speech perception is affected by nonproduction sources of information, such as context. Individual words are hard to understand in isolation but easy when heard in sentence context. It therefore seems that speech perception uses multiple sources that are integrated together in an optimal way. === Production === The motor theory of speech perception would predict that speech motor abilities in infants predict their speech perception abilities, but in actuality it is the other way around. It would also predict that defects in speech production would impair speech perception, but they do not. However, this only affects the first and already superseded behaviorist version of the theory, where infants were supposed to learn all production-perception patterns by imitation early in childhood. This is no longer the mainstream view of motor-speech theorists. === Speech module === Several sources of evidence for a specialized speech module have failed to be supported. Duplex perception can be observed with door slams. The McGurk effect can also be achieved with nonlinguistic stimuli, such as showing someone a video of a basketball bouncing but playing the sound of a ping-pong ball bouncing. As for categorical perception, listeners can be sensitive to acoustic differences within single phonetic categories. As a result, this part of the theory has been dropped by some researchers. === Sublexical tasks === The evidence provided for the motor theory of speech perception is limited to tasks such as syllable discrimination that use speech units not full spoken words or spoken sentences. As a result, "speech perception is sometimes interpreted as referring to the perception of speech at the sublexical level. However, the ultimate goal of these studies is presumably to understand the neural processes supporting the ability to process spee

    Read more →
  • Hapax legomenon

    Hapax legomenon

    In corpus linguistics, a hapax legomenon ( also or ; pl. hapax legomena; sometimes abbreviated to hapax, plural hapaxes) is a word or an expression that occurs only once within a context: either in the written record of an entire language, in the works of an author, or in a single text. The term is also sometimes used to describe a word that occurs in just one of an author's works but more than once in that particular work. Hapax legomenon is a transliteration of Greek ἅπαξ λεγόμενον, meaning "said once". The related terms dis legomenon, tris legomenon, and tetrakis legomenon respectively (, , ) refer to double, triple, or quadruple occurrences, but are far less commonly used. Hapax legomena are quite common, as predicted by Zipf's law, which states that the frequency of any word in a corpus is inversely proportional to its rank in the frequency table. For large corpora, about 40% to 60% of the words are hapax legomena, and another 10% to 15% are dis legomena. Thus, in the Brown Corpus of American English, about half of the 50,000 distinct words are hapax legomena within that corpus. Hapax legomenon refers to the appearance of a word or an expression in a body of text, not to either its origin or its prevalence in speech. It thus differs from a nonce word, which may never be recorded, may find currency and may be widely recorded, or may appear several times in the work which coins it, and so on. == Significance == Hapax legomena in ancient texts are usually difficult to decipher, since it is easier to infer meaning from multiple contexts than from just one. For example, many of the remaining undeciphered Mayan glyphs are hapax legomena, and Biblical (particularly Hebrew; see § Hebrew) hapax legomena sometimes pose problems in translation. Hapax legomena also pose challenges in natural language processing. Some scholars consider Hapax legomena useful in determining the authorship of written works. P. N. Harrison, in The Problem of the Pastoral Epistles (1921) made hapax legomena popular among Bible scholars, when he argued that there are considerably more of them in the three Pastoral Epistles than in other Pauline Epistles. He argued that the number of hapax legomena in a putative author's corpus indicates his or her vocabulary and is characteristic of the author as an individual. Harrison's theory has faded in significance due to a number of problems raised by other scholars. For example, in 1896, W. P. Workman found the following numbers of hapax legomena in each Pauline Epistle: At first glance, the last three totals (for the Pastoral Epistles) are not out of line with the others. To take account of the varying length of the epistles, Workman also calculated the average number of hapax legomena per page of the Greek text, which ranged from 3.6 to 13, as summarized in the diagram on the right. Although the Pastoral Epistles have more hapax legomena per page, Workman found the differences to be moderate in comparison to the variation among other Epistles. This was reinforced when Workman looked at several plays by Shakespeare, which showed similar variations (from 3.4 to 10.4 per page of Irving's one-volume edition), as summarized in the second diagram on the right. Apart from author identity, there are several other factors that can explain the number of hapax legomena in a work: text length: this directly affects the expected number and percentage of hapax legomena; the brevity of the Pastoral Epistles also makes any statistical analysis problematic. text topic: if the author writes on different subjects, of course many subject-specific words will occur only in limited contexts. text audience: if the author is writing to a peer rather than a student, or their spouse rather than their employer, again quite different vocabulary will appear. time: over the course of years, both the language and an author's knowledge and use of language will change. In the particular case of the Pastoral Epistles, all of these variables are quite different from those in the rest of the Pauline corpus, and hapax legomena are no longer widely accepted as strong indicators of authorship; those who reject Pauline authorship of the Pastorals rely on other arguments. There are also subjective questions over whether two forms amount to "the same word": dog vs. dogs, clue vs. clueless, sign vs. signature; many other gray cases also arise. The Jewish Encyclopedia points out that, although there are 1,500 hapaxes in the Hebrew Bible, only about 400 are not obviously related to other attested word forms. A final difficulty with the use of hapax legomena for authorship determination is that there is considerable variation among works known to be by a single author, and disparate authors often show similar values. In other words, hapax legomena are not a reliable indicator. Authorship studies now usually use a wide range of measures to look for patterns rather than relying upon single measurements. == Computer science == In the fields of computational linguistics and natural language processing (NLP), esp. corpus linguistics and machine-learned NLP, it is common to disregard hapax legomena (and sometimes other infrequent words), as they are likely to have little value for computational techniques. This disregard has the added benefit of significantly reducing the memory use of an application, since, by Zipf's law, many words are hapax legomena. == Examples == The following are some examples of hapax legomena in languages or corpora. === Arabic === In the Qurʾān: The proper nouns Iram (Q 89:7, Iram of the Pillars), Bābil (Q 2:102, Babylon), Bakka(t) (Q 3:96, Bakkah), Jibt (Q 4:51), Ramaḍān (Q 2:185, Ramadan), ar-Rūm (Q 30:2, Byzantine Empire), Tasnīm (Q 83:27), Qurayš (Q 106:1, Quraysh), Majūs (Q 22:17, Magian/Zoroastrian), Mārūt (Q 2:102, Harut and Marut), Makka(t) (Q 48:24, Mecca), Nasr (Q 71:23), (Ḏū) an-Nūn (Q 21:87) and Hārūt (Q 2:102, Harut and Marut) occur only once. zanjabīl (زَنْجَبِيل – ginger) is a Qurʾānic hapax (Q 76:17). zamharīr (زَمْهَرِيرًۭ) is a Qurʾānic hapax (Q 76:13), usually glossed as referring to extreme cold. The epitheton ornans aṣ-ṣamad (الصَّمَد – the One besought) is a Qurʾānic hapax (Q 112:2). ṭūd (طُودْ - mountain) is a Qurʾānic hapax (Q 26:63). === Chinese and Japanese === Classical Chinese and Japanese literature contains many Chinese characters that feature only once in the corpus, and their meaning and pronunciation has often been lost. Known in Japanese as kogo (孤語), literally "lonely characters", these can be considered a type of hapax legomenon. For example, the Classic of Poetry (c. 1000 BC) uses the character 篪 exactly once in the verse 「伯氏吹塤, 仲氏吹篪」, and it was only through the discovery of a description by Guo Pu (276–324 AD) that the character could be associated with a specific type of ancient flute. === English === It is fairly common for authors to "coin" new words to convey a particular meaning or for the sake of entertainment, without any suggestion that they are "proper" words. For example, P.G. Wodehouse and Lewis Carroll frequently coined novel words. Indexy, below, appears to be an example of this. Flother, as a synonym for snowflake, is a hapax legomenon of written English found in a manuscript entitled The XI Pains of Hell (c. 1275). Honorificabilitudinitatibus is a hapax legomenon of Shakespeare's works, coming from Erasmus' Adagia Indexy, in Bram Stoker's Dracula, used as an adjective to describe a situational state with no other further use in the language: "If that man had been an ordinary lunatic I would have taken my chance of trusting him; but he seems so mixed up with the Count in an indexy kind of way that I am afraid of doing anything wrong by helping his fads." Manticratic, meaning "of the rule by the Prophet's family or clan", was apparently invented by T. E. Lawrence and appears once in Seven Pillars of Wisdom. Nortelrye, a word for "education", occurs only once in Chaucer's The Reeve's Tale. Sassigassity, perhaps with the meaning of "audacity", occurs only once in Dickens's short story "A Christmas Tree". Slæpwerigne, "sleep-weary", occurs exactly once in the Old English corpus, in the Exeter Book. There is debate over whether it means "weary with sleep" or "weary for sleep". === German === The name of the 9th-century poem Muspilli is a back-formation from "muspille", Old High German hapax legomenon of unclear meaning only found in this text (see Muspilli § Etymology for discussion). === Ancient Greek === According to classical scholar Clyde Pharr, "the Iliad has 1,097 hapax legomena, while the Odyssey has 868". Others have defined the term differently, however, and count as few as 303 in the Iliad and 191 in the Odyssey. panaōrios (παναώριος), ancient Greek for "very untimely", is one of many words that occur only once in the Iliad. The Greek New Testament contains 686 local hapax legomena, which are sometimes called "New Testament hapaxes". 62 of these occur in 1 Peter and 54 occur in 2 Peter

    Read more →
  • Unsupervised learning

    Unsupervised learning

    Unsupervised learning is a framework in machine learning where, in contrast to supervised learning, algorithms learn patterns exclusively from unlabeled data. Other frameworks in the spectrum of supervisions include weak- or semi-supervision, where a small portion of the data is tagged, and self-supervision. Some researchers consider self-supervised learning a form of unsupervised learning. Conceptually, unsupervised learning divides into the aspects of data, training, algorithm, and downstream applications. Typically, the dataset is harvested cheaply "in the wild", such as massive text corpus obtained by web crawling, with only minor filtering (such as Common Crawl). This compares favorably to supervised learning, where the dataset (such as the ImageNet1000) is typically constructed manually, which is much more expensive. There are algorithms designed specifically for unsupervised learning, such as clustering algorithms like k-means, dimensionality reduction techniques like principal component analysis (PCA), Boltzmann machine learning, and autoencoders. After the rise of deep learning, most large-scale unsupervised learning has been done by training general-purpose neural network architectures by gradient descent, adapted to performing unsupervised learning by designing an appropriate training procedure. Sometimes a trained model can be used as-is, but more often they are modified for downstream applications. For example, the generative pretraining method trains a model to generate a textual dataset, before finetuning it for other applications, such as text classification. As another example, autoencoders are trained to produce good features, which can then be used as a module for other models, such as in a latent diffusion model. == Tasks == Tasks are often categorized as discriminative (recognition) or generative (imagination). Often but not always, discriminative tasks use supervised methods and generative tasks use unsupervised (see Venn diagram); however, the separation is very hazy. For example, object recognition favors supervised learning but unsupervised learning can also cluster objects into groups. Furthermore, as progress marches onward, some tasks employ both methods, and some tasks swing from one to another. For example, image recognition started off as heavily supervised, but became hybrid by employing unsupervised pre-training, and then moved towards supervision again with the advent of dropout, ReLU, and adaptive learning rates. A typical generative task is as follows. At each step, a datapoint is sampled from the dataset, and part of the data is removed, and the model must infer the removed part. This is particularly clear for the denoising autoencoders and BERT. == Neural network architectures == === Training === During the learning phase, an unsupervised network tries to mimic the data it is given and uses the error in its mimicked output to correct itself (i.e. correct its weights and biases). Sometimes the error is expressed as a low probability that the erroneous output occurs, or it might be expressed as an unstable high energy state in the network. In contrast to supervised methods' dominant use of backpropagation, unsupervised learning also employs other methods including: Hopfield learning rule, Boltzmann learning rule, Contrastive Divergence, Wake Sleep, Variational Inference, Maximum Likelihood, Maximum A Posteriori, Gibbs Sampling, and backpropagating reconstruction errors or hidden state reparameterizations. See the table below for more details. === Energy === An energy function is a macroscopic measure of a network's activation state. In Boltzmann machines, it plays the role of the Cost function. This analogy with physics is inspired by Ludwig Boltzmann's analysis of a gas' macroscopic energy from the microscopic probabilities of particle motion p ∝ e − E / k T {\displaystyle p\propto e^{-E/kT}} , where k is the Boltzmann constant and T is temperature. In the RBM network the relation is p = e − E / Z {\displaystyle p=e^{-E}/Z} , where p {\displaystyle p} and E {\displaystyle E} vary over every possible activation pattern and Z = ∑ All Patterns e − E ( pattern ) {\displaystyle \textstyle {Z=\sum _{\scriptscriptstyle {\text{All Patterns}}}e^{-E({\text{pattern}})}}} . To be more precise, p ( a ) = e − E ( a ) / Z {\displaystyle p(a)=e^{-E(a)}/Z} , where a {\displaystyle a} is an activation pattern of all neurons (visible and hidden). Hence, some early neural networks bear the name Boltzmann Machine. Paul Smolensky calls − E {\displaystyle -E\,} the Harmony. A network seeks low energy which is high Harmony. === Networks === This table shows connection diagrams of various unsupervised networks, the details of which will be given in the section Comparison of Networks. Circles are neurons and edges between them are connection weights. As network design changes, features are added on to enable new capabilities or removed to make learning faster. For instance, neurons change between deterministic (Hopfield) and stochastic (Boltzmann) to allow robust output, weights are removed within a layer (RBM) to hasten learning, or connections are allowed to become asymmetric (Helmholtz). Of the networks bearing people's names, only Hopfield worked directly with neural networks. Boltzmann and Helmholtz came before artificial neural networks, but their work in physics and physiology inspired the analytical methods that were used. === History === === Specific Networks === Here, we highlight some characteristics of select networks. The details of each are given in the comparison table below. Hopfield Network Ferromagnetism inspired Hopfield networks. A neuron corresponds to an iron domain with binary magnetic moments Up and Down, and neural connections correspond to the domain's influence on each other. Symmetric connections enable a global energy formulation. During inference the network updates each state using the standard activation step function. Symmetric weights and the right energy functions guarantees convergence to a stable activation pattern. Asymmetric weights are difficult to analyze. Hopfield nets are used as Content Addressable Memories (CAM). Boltzmann Machine These are stochastic Hopfield nets. Their state value is sampled from this pdf as follows: suppose a binary neuron fires with the Bernoulli probability p(1) = 1/3 and rests with p(0) = 2/3. One samples from it by taking a uniformly distributed random number y, and plugging it into the inverted cumulative distribution function, which in this case is the step function thresholded at 2/3. The inverse function = { 0 if x <= 2/3, 1 if x > 2/3 }. Sigmoid Belief Net Introduced by Radford Neal in 1992, this network applies ideas from probabilistic graphical models to neural networks. A key difference is that nodes in graphical models have pre-assigned meanings, whereas Belief Net neurons' features are determined after training. The network is a sparsely connected directed acyclic graph composed of binary stochastic neurons. The learning rule comes from Maximum Likelihood on p(X): Δwij ∝ {\displaystyle \propto } sj (si - pi), where pi = 1 / ( 1 + eweighted inputs into neuron i ). sj's are activations from an unbiased sample of the posterior distribution and this is problematic due to the Explaining Away problem raised by Judea Perl. Variational Bayesian methods uses a surrogate posterior and blatantly disregard this complexity. Deep Belief Network Introduced by Hinton, this network is a hybrid of RBM and Sigmoid Belief Network. The top 2 layers is an RBM and the second layer downwards form a sigmoid belief network. One trains it by the stacked RBM method and then throw away the recognition weights below the top RBM. As of 2009, 3-4 layers seems to be the optimal depth. Helmholtz machine These are early inspirations for the Variational Auto Encoders. Its 2 networks combined into one—forward weights operates recognition and backward weights implements imagination. It is perhaps the first network to do both. Helmholtz did not work in machine learning but he inspired the view of "statistical inference engine whose function is to infer probable causes of sensory input". the stochastic binary neuron outputs a probability that its state is 0 or 1. The data input is normally not considered a layer, but in the Helmholtz machine generation mode, the data layer receives input from the middle layer and has separate weights for this purpose, so it is considered a layer. Hence this network has 3 layers. Variational autoencoder These are inspired by Helmholtz machines and combines probability network with neural networks. An Autoencoder is a 3-layer CAM network, where the middle layer is supposed to be some internal representation of input patterns. The encoder neural network is a probability distribution qφ(z given x) and the decoder network is pθ(x given z). The weights are named phi & theta rather than W and V as in Helmholtz—a cosmetic difference. These 2 networks h

    Read more →