Yejin Choi (Korean: 최예진; born 1977) is the Dieter Schwarz Foundation Professor and Senior Fellow at the Department of Computer Science at Stanford University and the Stanford Institute for Human-Centered Artificial Intelligence (HAI) respectively. Her research considers natural language processing and computer vision. == Early life and education == Choi is from South Korea. She attended Seoul National University. After earning a bachelor's degree in Computer Science, Choi moved to the United States, where she joined Cornell University as a graduate student. There she worked with Claire Cardie on natural language processing. After earning her doctorate, Choi joined Stony Brook University as an Assistant Professor of Computer Science. At Stony Brook University Choi developed a statistical technique to identify fake hotel reviews. == Research and career == In 2018 Choi joined the Allen Institute for AI. Her research looks to endow computers with a statistical understanding of written language. She became interested in neural networks and their application in artificial intelligence. She started to assemble a knowledge base that became known as the atlas of machine commonsense (ATOMIC). By the time she had finished the creation of ATOMIC, the language model generative Pre-trained Transformer 2 (GPT-2) had been released. ATOMIC does not make use of linguistic rules, but combines the representations of different languages within a neural network. In 2020, Choi was endowed with the Brett Helsel Professorship, which she held until she became Chair of Computer Science in 2023. She has since made use of Commonsense Transformers (COMET) with Good old fashioned artificial intelligence (GOFAI). The approach combines symbolic reasoning and neural networks. She has developed computational models that can detect biases in language that work against people from underrepresented groups. For example, one study demonstrated that female film characters are portrayed as less powerful than their male counterparts. In 2023, Choi became The Wissner-Slivka Chair of Computer Science. Choi is also a scientific advisor to French research group Kyutai which is being funded by Xavier Niel, Rodolphe Saadé, Eric Schmidt, and others. In 2025, Stanford HAI announced the appointment of Choi as senior fellow and the Dieter Schwarz Foundation HAI Professor and Professor of Computer Science at Stanford University. == Awards and honours == 2013 International Conference on Computer Vision Marr Prize 2016 Institute of Electrical and Electronics Engineers AI One to Watch 2017 Facebook ParlAI Research Award 2018 Anita Borg Early Career Award 2020 Association for the Advancement of Artificial Intelligence Outstanding Paper Award 2021 Conference on Neural Information Processing Systems Outstanding Paper Award 2021 Association for Computational Linguistics Test-of-time Paper Award 2021 Conference on Computer Vision and Pattern Recognition Longuet-Higgins Prize 2022 North American Chapter of the Association for Computational Linguistics Best Paper Award 2022 International Conference on Machine Learning Outstanding Paper Award 2022 MacArthur Fellowship 2023 Association for Computational Linguistics Best Paper Award 2023 TIME100 Archived 2024-12-27 at the Wayback Machine AI 2023 2023 Empirical Methods in Natural Language Processing Outstanding Paper Award 2025 Association for Computational Linguistics Outstanding Paper Award 2025 Association for Computational Linguistics Best Demo Paper Award 2025 TIME100 AI 2025 == Select publications == Ott, Myle; Choi, Yejin; Cardie, Claire; Hancock, Jeffrey T. (2011). "Finding Deceptive Opinion Spam by Any Stretch of the Imagination". Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies. Portland, Oregon, USA: Association for Computational Linguistics: 309–319. arXiv:1107.4557. Bibcode:2011arXiv1107.4557O. ISBN 9781932432879. S2CID 2510724. Kulkarni, Girish; Premraj, Visruth; Ordonez, Vicente; Dhar, Sagnik; Li, Siming; Choi, Yejin; Berg, Alexander C.; Berg, Tamara L. (2013). "BabyTalk: Understanding and Generating Simple Image Descriptions". IEEE Transactions on Pattern Analysis and Machine Intelligence. 35 (12): 2891–2903. Bibcode:2013ITPAM..35.2891K. CiteSeerX 10.1.1.225.5228. doi:10.1109/TPAMI.2012.162. ISSN 1939-3539. PMID 22848128. Choi, Yejin; Cardie, Claire; Riloff, Ellen; Patwardhan, Siddharth (2005). "Identifying sources of opinions with conditional random fields and extraction patterns". Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing - HLT '05. Morristown, NJ, USA: Association for Computational Linguistics. pp. 355–362. doi:10.3115/1220575.1220620.
EfficientNet
EfficientNet is a family of convolutional neural networks (CNNs) for computer vision published by researchers at Google AI in 2019. Its key innovation is compound scaling, which uniformly scales all dimensions of depth, width, and resolution using a single parameter. EfficientNet models have been adopted in various computer vision tasks, including image classification, object detection, and segmentation. == Compound scaling == EfficientNet introduces compound scaling, which, instead of scaling one dimension of the network at a time, such as depth (number of layers), width (number of channels), or resolution (input image size), uses a compound coefficient ϕ {\displaystyle \phi } to scale all three dimensions simultaneously. Specifically, given a baseline network, the depth, width, and resolution are scaled according to the following equations: depth multiplier: d = α ϕ width multiplier: w = β ϕ resolution multiplier: r = γ ϕ {\displaystyle {\begin{aligned}{\text{depth multiplier: }}d&=\alpha ^{\phi }\\{\text{width multiplier: }}w&=\beta ^{\phi }\\{\text{resolution multiplier: }}r&=\gamma ^{\phi }\end{aligned}}} subject to α ⋅ β 2 ⋅ γ 2 ≈ 2 {\displaystyle \alpha \cdot \beta ^{2}\cdot \gamma ^{2}\approx 2} and α ≥ 1 , β ≥ 1 , γ ≥ 1 {\displaystyle \alpha \geq 1,\beta \geq 1,\gamma \geq 1} . The α ⋅ β 2 ⋅ γ 2 ≈ 2 {\displaystyle \alpha \cdot \beta ^{2}\cdot \gamma ^{2}\approx 2} condition is such that increasing ϕ {\displaystyle \phi } by a factor of ϕ 0 {\displaystyle \phi _{0}} would increase the total FLOPs of running the network on an image approximately 2 ϕ 0 {\displaystyle 2^{\phi _{0}}} times. The hyperparameters α {\displaystyle \alpha } , β {\displaystyle \beta } , and γ {\displaystyle \gamma } are determined by a small grid search. The original paper suggested 1.2, 1.1, and 1.15, respectively. Architecturally, they optimized the choice of modules by neural architecture search (NAS), and found that the inverted bottleneck convolution (which they called MBConv) used in MobileNet worked well. The EfficientNet family is a stack of MBConv layers, with shapes determined by the compound scaling. The original publication consisted of 8 models, from EfficientNet-B0 to EfficientNet-B7, with increasing model size and accuracy. EfficientNet-B0 is the baseline network, and subsequent models are obtained by scaling the baseline network by increasing ϕ {\displaystyle \phi } . == Variants == EfficientNet has been adapted for fast inference on edge TPUs and centralized TPU or GPU clusters by NAS. EfficientNet V2 was published in June 2021. The architecture was improved by further NAS search with more types of convolutional layers. It also introduced a training method, which progressively increases image size during training, and uses regularization techniques like dropout, RandAugment, and Mixup. The authors claim this approach mitigates accuracy drops often associated with progressive resizing.
Nicholas Carlini
Nicholas Carlini is an American researcher affiliated with Anthropic and previously with Google DeepMind who has published research in the fields of computer security and machine learning. He is known for his work on adversarial machine learning, particularly his work on the Carlini & Wagner attack in 2016. This attack was particularly useful in defeating defensive distillation, a method used to increase model robustness, and has since been effective against other defenses against adversarial input. In 2018, Carlini demonstrated an attack on Mozilla's DeepSpeech model, showing that hidden commands could be embedded in speech inputs, which the model would execute even if they were inaudible to humans. He also led a team at UC Berkeley that successfully broke seven out of nine defenses against adversarial attacks presented at the 2018 International Conference on Learning Representations. In addition to his work on adversarial attacks, Carlini has made significant contributions to understanding the privacy risks of machine learning models. In 2020, he revealed that large language models, like GPT-2, could memorize and output personally identifiable information. His research demonstrated that this issue worsened with larger models, and he later showed similar vulnerabilities in generative image models, such as Stable Diffusion. == Life and career == Nicholas Carlini obtained his Bachelor of Arts in Computer Science and Mathematics from the University of California, Berkeley, in 2013. He then continued his studies at the same university, where he pursued a PhD under the supervision of David Wagner, completing it in 2018. Carlini became known for his work on adversarial machine learning. In 2016, he worked alongside Wagner to develop the Carlini & Wagner attack, a method of generating adversarial examples against machine learning models. The attack was proved to be useful against defensive distillation, a popular mechanism where a student model is trained based on the features of a parent model to increase the robustness and generalizability of student models. The attack gained popularity when it was shown that the methodology was also effective against most other defenses, rendering them ineffective. In 2018, Carlini demonstrated an attack against Mozilla Foundation's DeepSpeech model where he showed that by hiding malicious commands inside normal speech input the speech model would respond to the hidden commands even when the commands were not discernible by humans. In the same year, Carlini and his team at UC Berkeley showed that out of the 11 papers presenting defenses to adversarial attacks accepted in that year's ICLR conference, seven of the defenses could be broken. Since 2021, he and his team have been working on large language models, creating a questionnaire where humans typically scored 35% whereas AI models scored in the 40%, with GPT-3 getting 38% which could be improved to 40% through few shot prompting. The best performer in the test was UnifiedQA, a model developed by Google specifically for answer questions and answer sets. Carlini has also developed methods to cause large language models like ChatGPT to answer harmful questions like how to construct bombs. He is also known for his work studying the privacy of machine learning models. In 2020, he showed for the first time that large language models would memorize some of the text data that they were trained on. For example, he found that GPT-2 could output personally identifiable information. He then led an analysis of larger models and studied how memorization increased with model size. Then, in 2022 he showed the same vulnerability in generative image models, and specifically diffusion models, by showing that Stable Diffusion could output images of people's faces that it was trained on. Following on this, Carlini then showed that ChatGPT would also sometimes output exact copies of webpages it was trained on, including personally identifiable information. Some of these studies have since been referenced by the courts in debating the copyright status of AI models. == Other work == Carlini received the Best of Show award at the 2020 IOCCC for implementing a tic-tac-toe game entirely with calls to printf, expanding on work from a research paper of his from 2015. The judges commented on his submission "This year's Best of Show (carlini) is such a novel way of obfuscation that it would be worth of a special mention in the (future) Best of IOCCC list!". [sic] == Awards == Best Student Paper Award, IEEE S&P 2017 ("Towards Evaluating the Robustness of Neural Networks") Best Paper Award, ICML 2018 ("Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples") Distinguished Paper Award, USENIX 2021 ("Poisoning the Unlabeled Dataset of Semi-Supervised Learning") Distinguished Paper Award, USENIX 2023 ("Tight Auditing of Differentially Private Machine Learning") Best Paper Award, ICML 2024 ("Stealing Part of a Production Language Model") Best Paper Award, ICML 2024 ("Considerations for Differentially Private Learning with Large-Scale Public Pretraining")
Top 10 AI Logo Makers Compared (2026)
In search of the best AI logo maker? An AI logo maker is software that uses machine learning to help you get more done — it turns a rough idea into a polished result in seconds. When choosing one, weigh output quality, pricing, export formats, and how well it fits the tools you already use. Whether you are a beginner or a pro, the right AI logo maker slots into your workflow and pays for itself fast. We tested the leading options and ranked them by quality, value, and ease of use.
AI Text-to-image Tools: Free vs Paid (2026)
Shopping for 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 keeps getting smarter as the underlying models improve. Pricing, accuracy, and the size of the model behind the tool are the three factors that most affect daily usefulness. Whether you are a beginner or a pro, the right AI text-to-image tool slots into your workflow and pays for itself fast. We tested the leading options and ranked them by quality, value, and ease of use.
Ordered dithering
Ordered dithering is any image dithering algorithm which uses a pre-set threshold map tiled across an image. It is commonly used to display a continuous image on a display of smaller color depth. For example, Microsoft Windows uses it in 16-color graphics modes. With the most common "Bayer" threshold map, the algorithm is characterized by noticeable crosshatch patterns in the result. == Threshold map == The algorithm reduces the number of colors by applying a threshold map M to the pixels displayed, causing some pixels to change color, depending on the distance of the original color from the available color entries in the reduced palette. The first threshold maps were designed by hand to minimise the perceptual difference between a grayscale image and its two-bit quantisation for up to a 4x4 matrix. An optimal threshold matrix is one that for any possible quantisation of color has the minimum possible texture so that the greatest impression of the underlying feature comes from the image being quantised. It can be proven that for matrices whose side length is a power of two there is an optimal threshold matrix. The map may be rotated or mirrored without affecting the effectiveness of the algorithm. This threshold map (for sides with length as power of two) is also known as a Bayer matrix or, when unscaled, an index matrix. For threshold maps whose dimensions are a power of two, the map can be generated recursively via: M 2 n = 1 ( 2 n ) 2 [ 4 M n 4 M n + 2 J n 4 M n + 3 J n 4 M n + J n ] = J 2 ⊗ M n + 1 n 2 M 2 ⊗ J n , {\displaystyle \mathbf {M} _{2n}={\frac {1}{(2n)^{2}}}{\begin{bmatrix}4\mathbf {M} _{n}&4\mathbf {M} _{n}+2\mathbf {J} _{n}\\4\mathbf {M} _{n}+3\mathbf {J} _{n}&4\mathbf {M} _{n}+\mathbf {J} _{n}\end{bmatrix}}=\mathbf {J} _{2}\otimes \mathbf {M} _{n}+{\frac {1}{n^{2}}}\mathbf {M} _{2}\otimes \mathbf {J} _{n},} where J n {\displaystyle \mathbf {J} _{n}} are n × n {\displaystyle n\times n} matrices of ones and ⊗ {\displaystyle \otimes } is the Kronecker product. While the metric for texture that Bayer proposed could be used to find optimal matrices for sizes that are not a power of two, such matrices are uncommon as no simple formula for finding them exists, and relatively small matrix sizes frequently give excellent practical results (especially when combined with other modifications to the dithering algorithm). This function can also be expressed using only bit arithmetic: M(i, j) = bit_reverse(bit_interleave(bitwise_xor(i, j), i)) / n ^ 2 == Pre-calculated threshold maps == Rather than storing the threshold map as a matrix of n {\displaystyle n} × n {\displaystyle n} integers from 0 to n 2 {\displaystyle n^{2}} , depending on the exact hardware used to perform the dithering, it may be beneficial to pre-calculate the thresholds of the map into a floating point format, rather than the traditional integer matrix format shown above. For this, the following formula can be used: Mpre(i,j) = Mint(i,j) / n^2 This generates a standard threshold matrix. for the 2×2 map: this creates the pre-calculated map: Additionally, normalizing the values to average out their sum to 0 (as done in the dithering algorithm shown below) can be done during pre-processing as well by subtracting 1⁄2 of the largest value from every value: Mpre(i,j) = Mint(i,j) / n^2 – 0.5 maxValue creating the pre-calculated map: == Algorithm == The ordered dithering algorithm renders the image normally, but for each pixel, it offsets its color value with a corresponding value from the threshold map according to its location, causing the pixel's value to be quantized to a different color if it exceeds the threshold. For most dithering purposes, it is sufficient to simply add the threshold value to every pixel (without performing normalization by subtracting 1⁄2), or equivalently, to compare the pixel's value to the threshold: if the brightness value of a pixel is less than the number in the corresponding cell of the matrix, plot that pixel black, otherwise, plot it white. This lack of normalization slightly increases the average brightness of the image, and causes almost-white pixels to not be dithered. This is not a problem when using a gray scale palette (or any palette where the relative color distances are (nearly) constant), and it is often even desired, since the human eye perceives differences in darker colors more accurately than lighter ones, however, it produces incorrect results especially when using a small or arbitrary palette, so proper normalization should be preferred. In other words, the algorithm performs the following transformation on each color c of every pixel: c ′ = n e a r e s t _ p a l e t t e _ c o l o r ( c + r × ( M ( x mod n , y mod n ) − 1 / 2 ) ) {\displaystyle c'=\mathrm {nearest\_palette\_color} {\mathopen {}}\left(c+r\times \left(M(x{\bmod {n}},y{\bmod {n}})-1/2\right){\mathclose {}}\right)} where M(i, j) is the threshold map on the i-th row and j-th column, c′ is the transformed color, and r is the amount of spread in color space. Assuming an RGB palette with 23N evenly distanced colors where each color (a triple of red, green and blue values) is represented by an octet from 0 to 255, one would typically choose r ≈ 255 N {\textstyle r\approx {\frac {255}{N}}} . (1⁄2 is again the normalizing term.) Because the algorithm operates on single pixels and has no conditional statements, it is very fast and suitable for real-time transformations. Additionally, because the location of the dithering patterns always stays the same relative to the display frame, it is less prone to jitter than error-diffusion methods, making it suitable for animations. Because the patterns are more repetitive than error-diffusion method, an image with ordered dithering compresses better. Ordered dithering is more suitable for line-art graphics as it will result in straighter lines and fewer anomalies. The values read from the threshold map should preferably scale into the same range as the minimal difference between distinct colors in the target palette. Equivalently, the size of the map selected should be equal to or larger than the ratio of source colors to target colors. For example, when quantizing a 24 bpp image to 15 bpp (256 colors per channel to 32 colors per channel), the smallest map one would choose would be 4×2, for the ratio of 8 (256:32). This allows expressing each distinct tone of the input with different dithering patterns. === A variable palette: pattern dithering === == Non-Bayer approaches == The above thresholding matrix approach describes the Bayer family of ordered dithering algorithms. A number of other algorithms are also known; they generally involve changes in the threshold matrix, which changes the distribution of the "noise" introduced by all kinds of dithering (the difference between the original image and the dithered image). === Halftone === Halftone dithering performs a form of clustered dithering, creating a look similar to halftone patterns, using a specially crafted matrix. === Void and cluster === The Void and cluster algorithm uses a pre-generated blue noise as the matrix for the dithering process. The blue noise matrix keeps the Bayer's good high frequency content, but with a more uniform coverage of all the frequencies involved shows a much lower amount of patterning. The "voids-and-cluster" method gets its name from the matrix generation procedure, where a black image with randomly initialized white pixels is gaussian-blurred to find the brightest and darkest parts, corresponding to voids and clusters. After a few swaps have evenly distributed the bright and dark parts, the pixels are numbered by importance. It takes significant computational resources to generate the blue noise matrix: on a modern computer a 64×64 matrix requires a couple seconds using the original algorithm. This algorithm can be extended to make animated dither masks which also consider the axis of time. This is done by running the algorithm in three dimensions and using a kernel which is a product of a two-dimensional gaussian kernel on the XY plane, and a one-dimensional Gaussian kernel on the Z axis. === Simulated Annealing === Simulated annealing can generate dither masks by starting with a flat histogram and swapping values to optimize a loss function. The loss function controls the spectral properties of the mask, allowing it to make blue noise or noise patterns meant to be filtered by specific filters. The algorithm can also be extended over time for animated dither masks with chosen temporal properties.
Permutation automaton
In automata theory, a permutation automaton, or pure-group automaton, is a deterministic finite automaton such that each input symbol permutes the set of states. Formally, a deterministic finite automaton A may be defined by the tuple (Q, Σ, δ, q0, F), where Q is the set of states of the automaton, Σ is the set of input symbols, δ is the transition function that takes a state q and an input symbol x to a new state δ(q,x), q0 is the initial state of the automaton, and F is the set of accepting states (also: final states) of the automaton. A is a permutation automaton if and only if, for every two distinct states qi and qj in Q and every input symbol x in Σ, δ(qi,x) ≠ δ(qj,x). A formal language is p-regular (also: a pure-group language) if it is accepted by a permutation automaton. For example, the set of strings of even length forms a p-regular language: it may be accepted by a permutation automaton with two states in which every transition replaces one state by the other. == Applications == The pure-group languages were the first interesting family of regular languages for which the star height problem was proved to be computable. Another mathematical problem on regular languages is the separating words problem, which asks for the size of a smallest deterministic finite automaton that distinguishes between two given words of length at most n – by accepting one word and rejecting the other. The known upper bound in the general case is O ( n 2 / 5 ( log n ) 3 / 5 ) {\displaystyle O(n^{2/5}(\log n)^{3/5})} . The problem was later studied for the restriction to permutation automata. In this case, the known upper bound changes to O ( n 1 / 2 ) {\displaystyle O(n^{1/2})} .