Text-to-image personalization

Text-to-Image personalization is a task in deep learning for computer graphics that augments pre-trained text-to-image generative models. In this task, a generative model that was trained on large-scale data (usually a foundation model), is adapted such that it can generate images of novel, user-provided concepts. These concepts are typically unseen during training, and may represent specific objects (such as the user's pet) or more abstract categories (new artistic style or object relations). Text-to-Image personalization methods typically bind the novel (personal) concept to new words in the vocabulary of the model. These words can then be used in future prompts to invoke the concept for subject-driven generation, inpainting, style transfer and even to correct biases in the model. To do so, models either optimize word-embeddings, fine-tune the generative model itself, or employ a mixture of both approaches. == Technology == Text-to-Image personalization was first proposed during August 2022 by two concurrent works, Textual Inversion and DreamBooth. In both cases, a user provides a few images (typically 3–5) of a concept, like their own dog, together with a coarse descriptor of the concept class (like the word "dog"). The model then learns to represent the subject through a reconstruction based objective, where prompts referring to the subject are expected to reconstruct images from the training set. In Textual Inversion, the personalized concepts are introduced into the text-to-image model by adding new words to the vocabulary of the model. Typical text-to-image models represent words (and sometimes parts-of-words) as tokens, or indices in a predefined dictionary. During generation, an input prompt is converted into such tokens, each of which is converted into a ‘word-embedding’: a continuous vector representation which is learned for each token as part of the model's training. Textual Inversion proposes to optimize a new word-embedding vector for representing the novel concept. This new embedding vector can then be assigned to a user-chosen string, and invoked whenever the user's prompt contains this string. In DreamBooth, rather than optimizing a new word vector, the full generative model itself is fine-tuned. The user first selects an existing token, typically one which rarely appears in prompts. The subject itself is then represented by a string containing this token, followed by a coarse descriptor of the subject's class. A prompt describing the subject will then take the form: "A photo of " (e.g. "a photo of sks cat" when learning to represent a specific cat). The text-to-image model is then tuned so that prompts of this form will generate images of the subject. == Textual Inversion == The key idea in Textual Inversion is to add a new term to the vocabulary of the diffusion model that corresponds to the new (personalized) concept. Textual Inversion operates by inverting the concepts into new pseudo-words within the textual embedding space of a pre-trained text-to-image model. These pseudo-words can be injected into new scenes using simple natural language descriptions, allowing for simple and intuitive modifications. The method allows a user to leverage multi-modal information — using a text-driven interface for ease of editing, but providing visual cues when approaching the limits of natural language. The resulting model is extremely light-weight per concept: only 1K long, but succeeds to encode detailed visual properties of the concept. == Extensions == Several approaches were proposed to refine and improve over the original methods. These include the following. Low-rank Adaptation (LoRA) - an adapter-based technique for efficient finetuning of models. In the case of text-to-image models, LoRA is typically used to modify the cross-attention layers of a diffusion model. Perfusion - a low rank update method that also locks the activations of the key matrix in the diffusion model's cross attention layers to the concept's coarse class. Extended Textual Inversion - a technique that learns an individual word embedding for each layer in the diffusion model's denoising network. Encoder-based methods that use another neural network to quickly personalize a model == Challenges and limitations == Text-to-image personalization methods must contend with several challenges. At their core is the goal of achieving high-fidelity to the personal concept while maintaining high alignment between novel prompts containing the subject, and the generated images (typically referred to as ‘editability’). Another challenge that personalization methods must contend with is memory requirements. Initial implementations of personalization methods required more than 20 Gigabytes of GPU memory, and more recent approaches have reported requirements of more than 40 Gigabytes. However, optimizations such as Flash Attention have since reduced this requirement considerably. Approaches that tune the entire generative model may also create checkpoints that are several gigabytes in size, making it difficult to share or store many models. Embedding based approaches require only a few kilobytes, but typically struggle to preserve identity while maintaining editability. More recent approaches have proposed hybrid tuning goals which optimize both an embedding and a subset of network weights. These can reduce storage requirements to as little as 100 Kilobytes while achieving quality comparable to full tuning methods. Finally, optimization processes can be lengthy, requiring several minutes of tuning for each novel concept. Encoder and quick-tuning methods aim to reduce this to seconds or less.

Normalization (image processing)

In image processing, normalization is a process that changes the range of pixel intensity values, a kind of intensity mapping. Applications include photographs with poor contrast due to glare, for example. A typical case is contrast stretching. In more general fields of data processing, such as digital signal processing, it is referred to as dynamic range expansion. The purpose of dynamic range expansion in the various applications is usually to bring the image, or other type of signal, into a range that is more familiar or normal to the senses, hence the term normalization. Often, the motivation is to achieve consistency in dynamic range for a set of data, signals, or images to avoid mental distraction or fatigue. For example, a newspaper will strive to make all of the images in an issue share a similar range of grayscale. Auto-normalization in image processing software typically normalizes to the full dynamic range of the number system specified in the image file format. == Definition == Normalization transforms an n-dimensional grayscale image I : { X ⊆ R n } → { Min , . . , Max } {\displaystyle I:\{\mathbb {X} \subseteq \mathbb {R} ^{n}\}\rightarrow \{{\text{Min}},..,{\text{Max}}\}} with intensity values in the range ( Min , Max ) {\displaystyle ({\text{Min}},{\text{Max}})} , into a new image I N : { X ⊆ R n } → { newMin , . . , newMax } {\displaystyle I_{N}:\{\mathbb {X} \subseteq \mathbb {R} ^{n}\}\rightarrow \{{\text{newMin}},..,{\text{newMax}}\}} with intensity values in the range ( newMin , newMax ) {\displaystyle ({\text{newMin}},{\text{newMax}})} . The linear normalization of a grayscale digital image is performed according to the formula I N = ( I − Min ) newMax − newMin Max − Min + newMin {\displaystyle I_{N}=(I-{\text{Min}}){\frac {{\text{newMax}}-{\text{newMin}}}{{\text{Max}}-{\text{Min}}}}+{\text{newMin}}} For example, if the intensity range of the image is 50 to 180 and the desired range is 0 to 255 the process entails subtracting 50 from each of pixel intensity, making the range 0 to 130. Then each pixel intensity is multiplied by 255/130, making the range 0 to 255. Normalization might also be non-linear, as the relationship between I {\displaystyle I} and I N {\displaystyle I_{N}} may not be linear. An example of non-linear normalization is when the normalization follows a sigmoid function, in which case the normalized image is computed according to the formula I N = ( newMax − newMin ) 1 1 + e − I − β α + newMin {\displaystyle I_{N}=({\text{newMax}}-{\text{newMin}}){\frac {1}{1+e^{-{\frac {I-\beta }{\alpha }}}}}+{\text{newMin}}} Where α {\displaystyle \alpha } defines the width of the input intensity range, and β {\displaystyle \beta } defines the intensity around which the range is centered. Gamma correction (log/inverse log) is also a common transformation function. === Colorspace === Intensity operations generally operate on a colorspace that maps to the human perception of lightness without intentionally changing the other properties. This can be done, for example, by operating on the L component of the CIELAB color space, or approximately by operating on the Y component of YCbCr. It is also possible to operate on each of the RGB color channels, though the result will not always make sense. == Contrast stretching == This is the most significant and essential technique of spatial-based image enhancement. The basic intent of this contrast enhancement technique is to adjust the local contrast in the image so as to bring out the clear regions or objects in the image. Low-contrast images often result from poor or non-uniform lighting conditions, a limited dynamic range of the imaging sensor, or improper settings of the lens aperture. This operation tries to change the intensity of the pixel in the image, particularly in the input image, to obtain an enhanced image. It is based on the number of techniques, namely local, global, dark and bright levels of contrast. The contrast enhancement is considered as the amount of color or gray differentiation that lies among the different features in an image. The contrast enhancement improves the quality of image by increasing the luminance difference between the foreground and background. A contrast stretching transformation can be achieved by: Stretching the dark range of input values into a wider range of output values: This involves increasing the brightness of the darker areas in the image to enhance details and improve visibility. Shifting the mid-range of input values: This involves adjusting the brightness levels of the mid-tones in the image to improve overall contrast and clarity. Compressing the bright range of input values: This process involves reducing the brightness of the brighter areas in the image to prevent overexposure resulting in a more balanced and visually appealing image. It can be described as the following piecewise funciton: I N = { s 1 r 1 I if I < r 1 s 2 − s 1 r 1 − r 2 ( I − r 1 ) if r 1 ≤ I ≤ r 2 1 − s 2 1 − r 2 ( I − r 2 ) if I > r 2 {\displaystyle I_{N}={\begin{cases}{\frac {s_{1}}{r_{1}}}I&{\text{if }}Ir_{2}\end{cases}}} Where: ( r 1 , s 1 ) {\displaystyle (r_{1},s_{1})} defines the transition point between the "dark" range to the "main" range. ( r 2 , s 2 ) {\displaystyle (r_{2},s_{2})} defines the transition point between the "main" range to the "bright" range. A typical linear stretch is obtained when ( r 1 , s 1 ) = ( r min , 0 ) {\displaystyle (r_{1},s_{1})=(r_{\text{min}},0)} and ( r 2 , s 2 ) = ( r max , 1 ) {\displaystyle (r_{2},s_{2})=(r_{\text{max}},1)} , where r min {\displaystyle r_{\text{min}}} and r max {\displaystyle r_{\text{max}}} denote the minimum and maximum levels in the source image. === Global contrast stretching === Global Contrast Stretching considers all color palate ranges at once to determine the maximum and minimum values for the entire RGB color image. This approach utilizes the combination of RGB colors to derive a single maximum and minimum value for contrast stretching across the entire image. === Local contrast stretching === Local contrast stretching (LCS) is an image enhancement method that focuses on locally adjusting each pixel's value to improve the visualization of structures within an image, particularly in both the darkest and lightest portions. It operates by utilizing sliding windows, known as kernels, which traverse the image. The central pixel within each kernel is adjusted using the following formula: I p ( x , y ) = 255 × [ I 0 ( x , y ) − m i n ] ( m a x − m i n ) {\displaystyle I_{p}(x,y)=255\times {\frac {[I_{0}(x,y)-min]}{(max-min)}}} Where: Ip(x,y) is the color level for the output pixel (x,y) after the contrast stretching process. I0(x,y) is the color level input for data pixel (x, y). max is the maximum value for color level in the input image within the selected kernel. min is the minimum value for color level in the input image within the selected kernel. A piecewise form (see above) may also be used. LCS can be applied to the three color channels of an image separately.

Nicolò Cesa-Bianchi

Nicolò Cesa-Bianchi (Italian pronunciation: [nikoˈlɔ tˈtʃɛːza ˈbjaŋki]) is an Italian computer scientist and Professor of Computer Science at the Department of Computer Science of the University of Milan. He is a researcher in the field of machine learning, and co-author of the books "Prediction, Learning, and Games" with Gabor Lugosi and "Regret analysis of stochastic and nonstochastic multi-armed bandit problems" with Sébastien Bubeck == Education and career == Cesa-Bianchi graduated in Computer Science from the University of Milan in 1988 where he received a PhD in Computer Science in 1993 supervised by Alberto Bertoni. During his PhD, he visited UC Santa Cruz where he worked with Manfred Warmuth and David Haussler. He did his postdoctoral studies at Graz University of Technology under the supervision of Wolfgang Maass. == Research == His research contributions focus on the following areas: design and analysis of machine learning algorithms, especially in online machine learning algorithms for multi-armed bandit problems, with applications to recommender systems and online auctions graph analytics, with applications to social networks and bioinformatics == Awards and honors == Cesa-Bianchi received a Google Research Award in 2010, a Xerox University Affairs Committee Award in 2011, a Criteo Faculty Award in 2017, a Google Faculty Award in 2018, and a IBM Academic Award in 2021. Since 2023 he is corresponding member of the Accademia dei Lincei.

Robert Wilensky

Robert Wilensky (26 March 1951 – 15 March 2013) was an American computer scientist and professor at the UC Berkeley School of Information, with his main focus of research in artificial intelligence. == Academic career == In 1971, Wilensky received his bachelor's degree in mathematics from Yale University, and in 1978, a Ph.D. in computer science from the same institution. After finishing his thesis, "Understanding Goal-Based Stories", Wilensky joined the faculty from the EECS Department of UC Berkeley. In 1986, he worked as the doctoral advisor of Peter Norvig, who then later published the standard textbook of the field: Artificial Intelligence: A Modern Approach. From 1993 to 1997, Wilensky was the Berkeley Computer Science Division Chair. During this time, he also served as director of the Berkeley Cognitive Science Program, director of the Berkeley Artificial Intelligence Research Project, and board member of the International Computer Science Institute. In 1997, he became a fellow of the Association for Computing Machinery "for research contributions to the areas of natural language processing and digital libraries as well as outstanding leadership in Computer Science." Furthermore, he also was a Fellow of the Association for the Advancement of Artificial Intelligence. He retired from faculty in 2007 and died on Friday, March 15, 2013, of a bacterial infection at the Alta Bates Summit Medical Center. Wilensky was married to Ann Danforth and he is survived by her and their two children, Avi and Eli Wilensky == Research == Throughout his career, Wilensky authored and co-authored over 60 scholarly articles and technical reports on AI, natural language processing, and information dissemination. In addition to his numerous technical publications, Wilensky also published two books on the programming language LISP, LISPcraft and Common LISPcraft, and had almost completed another book manuscript when he suffered a cardiac arrest and stopped writing. Among his publications are: R. Wilensky, (1986-09-17). Common LISPcraft. W. W. Norton & Company. ISBN 9780393955446. T. A. Phelps and R. Wilensky, "Toward active, extensible, networked documents: Multivalent architecture and applications," in Proc. 1st ACM Intl. Conf. on Digital Libraries, E. A. Fox and G. Marchionini, Eds., New York, NY: ACM Press, 1996, pp. 100–108. J. Traupman and R. Wilensky, "Experiments in Improving Unsupervised Word Sense Disambiguation," University of California, Berkeley, Department of EECS, Computer Science Division, Tech. Rep. 03–1227, Feb. 2003. R. Wilensky, Planning and Understanding: A Computational Approach to Human Reasoning, Advanced Book Program, Reading, MA: Addison-Wesley Publishing Co., 1983. R. Wilensky, "Understanding Goal-Based Stories," Yale University, Sep. 1978. B. Kahn and R. Wilensky, "A Framework for Distributed Digital Object Services", May 1995.

Larry Heck

Larry Paul Heck is the Rhesa Screven Farmer, Jr., Advanced Computing Concepts Chair, Georgia Research Alliance Eminent Scholar, Co-Executive Director of the Machine Learning Center and Professor at the Georgia Institute of Technology. His career spans many of the sub-disciplines of artificial intelligence, including conversational AI, speech recognition and speaker recognition, natural language processing, web search, online advertising and acoustics. He is best known for his role as a co-founder of the Microsoft Cortana Personal Assistant and his early work in deep learning for speech processing. == Education and career == Larry Heck was born in Havre, Montana. After receiving the Bachelor of Science in electrical engineering at Texas Tech University, he was admitted to graduate school at the Georgia Institute of Technology in 1986. Heck received the MSEE in 1989 and the PhD in 1991 under advisor Prof. James H. McClellan. From 1992 to 1998, he was a senior research engineer at SRI International with the Acoustics and Radar Technology Lab (ARTL) and Speech Technology and Research (STAR) Lab, and in 1998 joined Nuance Communications, serving as vice president of R&D. Funded by the US government's NSA and DARPA from 1995-1998, Heck led the SRI team that was the first to successfully create large-scale deep neural network (DNN) deep learning technology in the field of speech processing. The deep learning technology was used to win the 1998 National Institute of Standards and Technology Speaker Recognition evaluation. The approach trained a 5-layer deep neural network, with the first two layers used as a (learned) feature extractor. To stabilize the training of the DNN, a weight normalization method was used (later rediscovered in 2010 by Xavier, et.al). Heck deployed this DNN in 1999 with Nuance Communications at the Home Shopping Network, representing the first major industrial application of deep learning with over 100K Nuance Verifier voiceprints. From 2005 to 2008, he was vice president of search & advertising quality at Yahoo!. In 2008, Heck and Ron Brachman combined search & advertising quality with Yahoo! Research to form Yahoo! Labs. Beginning in 2009, he was the chief scientist of speech products at Microsoft. In this role, he established the vision, mission and long-range plan and hired the initial team to create Microsoft’s digital-personal-assistant Cortana. Heck was named a Microsoft Distinguished Engineer in 2012 and joined Microsoft Research that same year. In 2014, he joined Google as a principal research scientist, where he founded the deep learning-based conversational AI team "Deep Dialogue". The team works on advanced research for the Google Assistant. In 2017, Heck joined Samsung as SVP and co-head of global AI Research. In 2019, he became head of Bixby (virtual assistant) North America and the CEO of Viv Labs, an independent subsidiary of Samsung. In that same year, Heck led one of the first large scale deployments of Transformer-Based LLMs as part of the Bixby Categories launch at the 2019 Samsung Developer Conference. In 2021, Heck returned to the Georgia Institute of Technology as a Professor. == Awards and honors == Larry Heck was named Fellow of the Institute of Electrical and Electronics Engineers (IEEE) in 2016 for leadership in application of machine learning to spoken and text language processing. Heck was inducted as a Fellow of the National Academy of Inventors (NAI) in 2024. Heck received the 2017 Academy of Distinguished Engineering Alumni Award from the Georgia Institute of Technology. In the same year, he also received the Texas Tech University Whitacre College of Engineering Distinguished Engineer Award. Larry Heck has several best papers including the 2020 IEEE Signal Processing Society (SPS) Best Paper Award: “Using Recurrent Neural Networks for Slot Filling in Spoken Language Understanding” published in the IEEE/ACM Transactions on Audio, Speech, and Language Processing in March 2015, and the 2020 ACM Conference on Information and Knowledge Management (CIKM) Test of Time Award for the paper "Learning Deep Structured Semantic Models for Web Search using Clickthrough Data".

Waveform graphics

Waveform graphics is a simple vector graphics system introduced by Digital Equipment Corporation (DEC) on the VT55 and VT105 terminals in the mid-1970s. It was used to produce graphics output from mainframes and minicomputers. DEC used the term "waveform graphics" to refer specifically to the hardware, but it was used more generally to describe the whole system. The system was designed to use as little computer memory as possible. At any given X location it could draw two dots at given Y locations, making it suitable for producing two superimposed waveforms, line charts or histograms. Text and graphics could be mixed, and there were additional tools for drawing axes and markers. The waveform graphics system was used only for a short period of time before it was replaced by the more sophisticated ReGIS system, first introduced on the VT125 in 1981. ReGIS allowed the construction of arbitrary vectors and other shapes. Whereas DEC normally provided a backward compatible solution in newer terminal models, they did not choose to do this when ReGIS was introduced, and waveform graphics disappeared from later terminals. == Description == Waveform graphics was introduced on the VT55 terminal in October 1975, an era when memory was extremely expensive. Although it was technically possible to produce a bitmap display using a framebuffer using technology of the era, the memory needed to do so at a reasonable resolution was typically beyond the price point that made it practical. All sorts of systems were used to replace computer memory with other concepts, like the storage tubes used in the Tektronix 4010 terminals, or the zero memory racing-the-beam system used in the Atari 2600. DEC chose to attack this problem through a clever use of a small buffer representing only the vertical positions on the screen. Such a system could not draw arbitrary shapes, but would allow the display of graph data. The system was based on a 512 by 236 pixel display, producing 512 vertical columns along the X-axis, and 236 horizontal rows on the Y-axis. Y locations were counted up from the bottom, so the coordinate 0,0 was in the lower left, and 511, 235 in the upper right. Had this been implemented using a framebuffer with each location represented by a single bit, 512 × 236 × 1 = 120,832 bits, or 15,104 bytes, would have been required. At the time, memory cost about $50 per kilobyte, so the buffer alone would cost over $700 (equivalent to $4,570 in 2025). Instead, the waveform graphic system used one byte of memory for each X axis location, with the byte's value representing the Y location. This required only 512 bytes for each graph, a total of 1024 bytes for the two graphs. Drawing a line required the programmer to construct a series of Y locations and send them as individual points, the terminal could not connect the dots itself. To make this easier, the terminal automatically incremented the X location every time an Y coordinate was received, so a graph line could be sent as a long string of numbers for subsequent Y locations instead of having to repeatedly send the X location every time. Drawing normally started by sending a single instruction to set the initial X location, often 0 on the left, and then sending in data for the entire curve. The system also included storage for up to 512 markers on both lines. These were always drawn centered on the Y value of the line they were associated with, meaning that a simple on/off indication for X locations was all that was needed, requiring only 1024 bits, or 128 bytes, in total. The markers extended 16 pixels vertically, and could only be aligned on 16-pixel boundaries, so they were not necessarily centered across the underlying graph. Markers were used to indicate important points on the graph, where a symbol of some sort would normally be used. The system also allowed a vertical line to be drawn for every horizontal location and a horizontal one at every vertical location. These were also stored as simple on/off bits, requiring another 128 bytes of memory. These lines were used to draw axes and scale lines, or could be used for a screen-spanning crosshair cursor. A separate set of two 7-bit registers held additional information about the drawing style and other settings. Although complex from the user's perspective, this system was easy to implement in hardware. A cathode ray tube produces a display by scanning the screen in a series of horizontal motions, moving down one vertical line after each horizontal scan. At any given instant during this process, the display hardware examines a few memory locations to see if anything needs to be displayed. For instance, it can determine whether to draw a marker on graph 0 by examining register 1 to see if markers are turned on, looking in the marker buffer to see if there is a 1 at the current X location, and then examining the Y location of graph 0 to see if it is within 16 pixels of the current scan line. If all of these are true, a spot is drawn to present that portion of the marker. As this will be true for 16 vertical locations during the scanning process, a 16-pixel high marker will be drawn. Sold alone, the VT55 was priced at $2,496 (equivalent to $16,295 in 2025),. Like other models of the VT50 series, the terminal could be equipped with an optional wet-paper printer in a panel on the right of the screen. This added $800 (equivalent to $5,223 in 2025) to the price. DEC also offered VT55 in a package with a small model of the PDP-11 to create one model of the DEClab 11/03 system. The DEClab normally sold for $14,000 (equivalent to $91,397 in 2025) with a DECwriter II (LA36) hard-copy terminal for $15,000 (equivalent to $97,925 in 2025), with the VT55. The system had I/O channels for up to 15 lab devices, and included libraries for FORTRAN and BASIC for reading the data and creating graphs. The fairly extensive VT55 Programmers Manual covered the latter in depth. == Commands and data == Data was sent to the terminal using an extended set of codes similar to those introduced on the VT52. VT52 codes generally started with the ESC character (octal 33, decimal 27) and was then followed by a single letter instruction. For instance, the string of four characters ESC H ESC J would reposition the cursor in the upper left (home) and then clear the screen from that point down. These codes were basically modeless; triggered by the ESC the resulting escape mode automatically exited again when the command was complete. Escape codes could be interspersed with display text anywhere in the stream of data. In contrast, the graphics system was entirely modal, with escape sequences being sent to cause the terminal to enter or exit graph drawing mode. Data sent between these two codes were interpreted by the graphics hardware, so text and graphics could not be mixed in a single stream of instructions. Graphics mode was entered by sending the string ESC 1, and exited again with the string ESC 2. Even the commands within the graphics mode were modal; characters were interpreted as being additional data for the previous load character (command) until another load character is seen. Ten load characters were available: @ - no operation, used to tell the terminal the last command is no longer active A - load data into register 0, selecting the drawing mode for the two graphs I - load data into register 1, selecting other drawing options H - load the starting X position (Horizontal) for the following commands B - load data for Y locations for graph 0 starting at the H position selected earlier J - load data for Y locations for graph 1 starting at the H position selected earlier C - store a marker on graph 0 at the following X location K - store a marker on graph 1 at the following X location D - draw a horizontal line at the given Y location L - draw a vertical line at the given X location X and Y locations were sent as 10-bit decimal numbers, encoded as ASCII characters, with 5 bits per character. This means that any number within the 1024 number space (210) can be stored as a string of two characters. To ensure the characters can be transmitted over 7-bit links, the pattern 01 is placed in front of both 5-bit numbers, producing 7-bit ASCII values that are always within the printable range. This results in a somewhat complex encoding algorithm. For instance, if one wanted to encode the decimal value 102, first you convert that to the 10-bit decimal pattern 0010010010. That is then split that into upper and lower 5-bit parts, 00100 and 10010. Then append 01 binary to produce 7-bit numbers 0100100 and 0110010. Individually convert back to decimal 40 and 50, and then look up those characters in an ASCII chart, finding ( and 2. These have to be sent to the terminal least significant character first. If these were being used to set the X coordinate, the complete string would be H2(. When used as X and Y locations for the graphs, extra digits were ignored. For instance, the 512 pixel X axis r

Danqi Chen

Danqi Chen (Chinese: 陈丹琦; pinyin: Chén Dānqí, IPA: [ʈ͡ʂʰə̌n tan t͡ɕʰǐ]; born in Changsha, China) is a Chinese computer scientist and assistant professor at Princeton University specializing in the AI field of natural language processing (NLP). In 2019, she joined the Princeton NLP group, alongside Sanjeev Arora, Christiane Fellbaum, and Karthik Narasimhan. She was previously a visiting scientist at Facebook AI Research (FAIR). She earned her Ph.D. at Stanford University and her BS from Tsinghua University. Chen is the author of Neural Reading Comprehension and Beyond, a dissertation on using artificial intelligence to access knowledge in ordinary and structured documents. She is the author or co-author of a number of journal articles, including Reading Wikipedia to Answer Open-Domain Questions. Google's SyntaxNet is based on algorithms developed by Danqi Chen and Christopher Manning at Stanford. Her primary research interests are in text understanding and knowledge representation and reasoning. She won a gold medal at the 2008 International Informatics Olympiad. She is known among friends as CDQ. A well known algorithm in competitive programming, CDQ Divide and Conquer, is named after this acronym. She is married to Huacheng Yu, an assistant professor in theoretical computer science at Princeton University.