Cyber-Duck is a digital transformation agency founded in 2005 and based in Elstree, United Kingdom. The company specialises in user experience (UX), software development and digital optimisation. The company employs over 90 staff in the UK and Europe. It works with clients from the financial, pharmaceutical, sport, motoring and security sectors, among others. These include the Bank of England, Cancer Research UK, GOV.UK Verify partner CitizenSafe, The Commonwealth of Nations and Sport England. == History == Cyber-Duck was founded in 2005 by Danny Bluestone in his flat in Mill Hill, United Kingdom. After a few months, the firm moved into its first office in Borehamwood. Projects with Ogilvy, London Creative and Wisteria followed before Cyber-Duck moved to offices in Devonshire House, Borehamwood. In 2010, the firm was commissioned to develop a website for the European Commission in the UK. In 2011, the company moved to a self-contained premises in Elstree, Hertfordshire. Shortly afterward, Cyber-Duck was listed on the Deloitte Technology Fast 500 EMEA in recognition of its substantial revenue growth over the previous five years. As the company grew, its expertise also broadened. This resulted in guest spots on several television shows. Cyber-Duck was featured in an episode of the Gadget Show in 2011, and Chief Production Officer Matt Gibson appeared on BBC Watchdog in 2013 to assist in researching websites and their checkout processes. The firm continued to attract business from companies in London, so the decision was made to open a new office in central London. The Farringdon office opened in 2015, and was followed by a rebrand. In 2016, Cyber-Duck went on to work with the Bank of England. Ahead of the launch of the new polymer £5 note, featuring Winston Churchill, the company was tasked with creating a user-friendly website to showcase the new banknote and promote public awareness. The success of the campaign led to further commissions, including 2017's website the New Ten and a redesign of the Bank of England's main website. The firm underwent significant growth in 2020, beginning working partnerships with Sport England and the College of Policing. During this time they also launched DevOps as a new service. In 2022, the Farringdon office closed and was relocated to a new office space in Holborn. The Laravel, Drupal and DevOps teams expanded, and Cyber-Duck became the lead Digital Agency for Worcester, Bosch Group. Several members of the team appeared on The Digital Society on Sky UK. == Awards and accreditations == Cyber-Duck is known for its focus on process accreditation as a driver of creativity. In 2011, the company obtained its first ISO 9241 accreditation in Human Centred Design for interactive systems. Two years later, Cyber-Duck obtained a further certification, the ISO 9001 for Quality Management Systems. It acquired another certification in 2016 with the ISO 27001 – the focus of this accreditation was Information Security Management. In 2022, Cyber-Duck gained the ISO 14001 certification in Environmental Management. Cyber-Duck's digital products have won numerous Wirehive 100, BIMA and Webby awards. Notably, the company's UX Companion, a free iOS and Android app that is a glossary of UX theories, featured in Usability Geek and Smashing Magazine. In 2021 they were awarded as one of the UK's 100 Best Small Companies to work for, and BIMA10 shortlisted for their work with Sport England and This Girl Can.
Grokking (machine learning)
In machine learning, grokking, or delayed generalization, is a phenomenon observed in some settings where a model abruptly transitions from overfitting (performing well only on training data) to generalizing (performing well on both training and test data), after many training iterations with little or no improvement on the held-out data. This contrasts with what is typically observed in machine learning, where generalization occurs gradually alongside improved performance on training data. == Origin == Grokking was introduced by OpenAI researcher Alethea Power and colleagues in the January 2022 paper "Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets". It is derived from the word grok coined by Robert Heinlein in his novel Stranger in a Strange Land. In ML research, "grokking" is not used as a synonym for "generalization"; rather, it names a sometimes-observed delayed‑generalization training phenomenon in which training and held‑out performance do not improve in tandem, and in which held‑out performance rises abruptly later. Authors also analyze the "grokking time", the epoch or step at which this transition occurs in those scenarios. == Interpretations == Grokking can be understood as a phase transition during the training process. In particular, recent work has shown that grokking may be due to a complexity phase transition in the model during training. While grokking has been thought of as largely a phenomenon of relatively shallow models, grokking has been observed in deep neural networks and non-neural models and is the subject of active research. One potential explanation is that the weight decay (a component of the loss function that penalizes higher values of the neural network parameters, also called regularization) slightly favors the general solution that involves lower weight values, but that is also harder to find. According to Neel Nanda, the process of learning the general solution may be gradual, even though the transition to the general solution occurs more suddenly later. Recent theories have hypothesized that grokking occurs when neural networks transition from a "lazy training" regime where the weights do not deviate far from initialization, to a "rich" regime where weights abruptly begin to move in task-relevant directions. Follow-up empirical and theoretical work has accumulated evidence in support of this perspective, and it offers a unifying view of earlier work as the transition from lazy to rich training dynamics is known to arise from properties of adaptive optimizers, weight decay, initial parameter weight norm, and more. This perspective is complementary to a unifying "pattern learning speeds" framework that links grokking and double descent; within this view, delayed generalization can arise across training time ("epoch‑wise") or across model size ("model‑wise"), and the authors report "model‑wise grokking".
Frequent pattern discovery
Frequent pattern discovery (or FP discovery, FP mining, or Frequent itemset mining) is part of knowledge discovery in databases, Massive Online Analysis, and data mining; it describes the task of finding the most frequent and relevant patterns in large datasets. The concept was first introduced for mining transaction databases. Frequent patterns are defined as subsets (itemsets, subsequences, or substructures) that appear in a data set with frequency no less than a user-specified or auto-determined threshold. == Techniques == Techniques for FP mining include: market basket analysis cross-marketing catalog design clustering classification recommendation systems For the most part, FP discovery can be done using association rule learning with particular algorithms Eclat, FP-growth and the Apriori algorithm. Other strategies include: Frequent subtree mining Structure mining Sequential pattern mining and respective specific techniques. Implementations exist for various machine learning systems or modules like MLlib for Apache Spark.
Quadratic classifier
In statistics, a quadratic classifier is a statistical classifier that uses a quadratic decision surface to separate measurements of two or more classes of objects or events. It is a more general version of the linear classifier. == The classification problem == Statistical classification considers a set of vectors of observations x of an object or event, each of which has a known type y. This set is referred to as the training set. The problem is then to determine, for a given new observation vector, what the best class should be. For a quadratic classifier, the correct solution is assumed to be quadratic in the measurements, so y will be decided based on x T A x + b T x + c {\displaystyle \mathbf {x^{T}Ax} +\mathbf {b^{T}x} +c} In the special case where each observation consists of two measurements, this means that the surfaces separating the classes will be conic sections (i.e., either a line, a circle or ellipse, a parabola or a hyperbola). In this sense, we can state that a quadratic model is a generalization of the linear model, and its use is justified by the desire to extend the classifier's ability to represent more complex separating surfaces. == Quadratic discriminant analysis == Quadratic discriminant analysis (QDA) is closely related to linear discriminant analysis (LDA), where it is assumed that the measurements from each class are normally distributed. Unlike LDA however, in QDA there is no assumption that the covariance of each of the classes is identical. When the normality assumption is true, the best possible test for the hypothesis that a given measurement is from a given class is the likelihood ratio test. Suppose there are only two groups, with means μ 0 , μ 1 {\displaystyle \mu _{0},\mu _{1}} and covariance matrices Σ 0 , Σ 1 {\displaystyle \Sigma _{0},\Sigma _{1}} corresponding to y = 0 {\displaystyle y=0} and y = 1 {\displaystyle y=1} respectively. Then the likelihood ratio is given by Likelihood ratio = | 2 π Σ 1 | − 1 exp ( − 1 2 ( x − μ 1 ) T Σ 1 − 1 ( x − μ 1 ) ) | 2 π Σ 0 | − 1 exp ( − 1 2 ( x − μ 0 ) T Σ 0 − 1 ( x − μ 0 ) ) < t {\displaystyle {\text{Likelihood ratio}}={\frac {{\sqrt {|2\pi \Sigma _{1}|}}^{-1}\exp \left(-{\frac {1}{2}}(\mathbf {x} -{\boldsymbol {\mu }}_{1})^{T}\Sigma _{1}^{-1}(\mathbf {x} -{\boldsymbol {\mu }}_{1})\right)}{{\sqrt {|2\pi \Sigma _{0}|}}^{-1}\exp \left(-{\frac {1}{2}}(\mathbf {x} -{\boldsymbol {\mu }}_{0})^{T}\Sigma _{0}^{-1}(\mathbf {x} -{\boldsymbol {\mu }}_{0})\right)}} Linear discriminant analysis (LDA), normal discriminant analysis (NDA), canonical variates analysis (CVA), or discriminant function analysis is a generalization of Fisher's linear discriminant, a method used in statistics and other fields, to find a linear combination of features that characterizes or separates two or more classes of objects or events. The resulting combination may be used as a linear classifier, or, more commonly, for dimensionality reduction before later classification. LDA is closely related to analysis of variance (ANOVA) and regression analysis, which also attempt to express one dependent variable as a linear combination of other features or measurements. However, ANOVA uses categorical independent variables and a continuous dependent variable, whereas discriminant analysis has continuous independent variables and a categorical dependent variable (i.e. the class label). Logistic regression and probit regression are more similar to LDA than ANOVA is, as they also explain a categorical variable by the values of continuous independent variables. These other methods are preferable in applications where it is not reasonable to assume that the independent variables have a normal distribution, which is a fundamental assumption of the LDA method. LDA is also closely related to principal component analysis (PCA) and factor analysis in that they both look for linear combinations of variables which best explain the data. LDA explicitly attempts to model the difference between the classes of data. PCA, in contrast, does not take into account any difference in class, and factor analysis builds the feature combinations based on similarities rather than differences. Discriminant analysis is also different from factor analysis in that it is not an interdependence technique: a distinction between independent variables and dependent variables (also called criterion variables) must be made. LDA works when the measurements made on independent variables for each observation are continuous quantities. When dealing with categorical independent variables, the equivalent technique is discriminant correspondence analysis. Discriminant analysis is used when groups are known a priori (unlike in cluster analysis). Each case must have a score on one or more quantitative predictor measures, and a score on a group measure. In simple terms, discriminant function analysis is classification - the act of distributing things into groups, classes or categories of the same type. == History == The original dichotomous discriminant analysis was developed by Sir Ronald Fisher in 1936. It is different from an ANOVA or MANOVA, which is used to predict one (ANOVA) or multiple (MANOVA) continuous dependent variables by one or more independent categorical variables. Discriminant function analysis is useful in determining whether a set of variables is effective in predicting category membership. == LDA for two classes == Consider a set of observations x → {\displaystyle {\vec {x}}} (also called features, attributes, variables or measurements) for each sample of an object or event with known class y {\displaystyle y} . This set of samples is called the training set in a supervised learning context. The classification problem is then to find a good predictor for the class y {\displaystyle y} of any sample of the same distribution (not necessarily from the training set) given only an observation x → {\displaystyle {\vec {x}}} . LDA approaches the problem by assuming that the conditional probability density functions p ( x → | y = 0 ) {\displaystyle p({\vec {x}}|y=0)} and p ( x → | y = 1 ) {\displaystyle p({\vec {x}}|y=1)} are both the normal distribution with mean and covariance parameters ( μ → 0 , Σ 0 ) {\displaystyle \left({\vec {\mu }}_{0},\Sigma _{0}\right)} and ( μ → 1 , Σ 1 ) {\displaystyle \left({\vec {\mu }}_{1},\Sigma _{1}\right)} , respectively. Under this assumption, the Bayes-optimal solution is to predict points as being from the second class if the log of the likelihood ratios is bigger than some threshold T, so that: 1 2 ( x → − μ → 0 ) T Σ 0 − 1 ( x → − μ → 0 ) + 1 2 ln | Σ 0 | − 1 2 ( x → − μ → 1 ) T Σ 1 − 1 ( x → − μ → 1 ) − 1 2 ln | Σ 1 | > T {\displaystyle {\frac {1}{2}}({\vec {x}}-{\vec {\mu }}_{0})^{\mathrm {T} }\Sigma _{0}^{-1}({\vec {x}}-{\vec {\mu }}_{0})+{\frac {1}{2}}\ln |\Sigma _{0}|-{\frac {1}{2}}({\vec {x}}-{\vec {\mu }}_{1})^{\mathrm {T} }\Sigma _{1}^{-1}({\vec {x}}-{\vec {\mu }}_{1})-{\frac {1}{2}}\ln |\Sigma _{1}|\ >\ T} Without any further assumptions, the resulting classifier is referred to as quadratic discriminant analysis (QDA). LDA instead makes the additional simplifying homoscedasticity assumption (i.e. that the class covariances are identical, so Σ 0 = Σ 1 = Σ {\displaystyle \Sigma _{0}=\Sigma _{1}=\Sigma } ) and that the covariances have full rank. In this case, several terms cancel: x → T Σ 0 − 1 x → = x → T Σ 1 − 1 x → {\displaystyle {\vec {x}}^{\mathrm {T} }\Sigma _{0}^{-1}{\vec {x}}={\vec {x}}^{\mathrm {T} }\Sigma _{1}^{-1}{\vec {x}}} x → T Σ i − 1 μ → i = μ → i T Σ i − 1 x → {\displaystyle {\vec {x}}^{\mathrm {T} }{\Sigma _{i}}^{-1}{\vec {\mu }}_{i}={{\vec {\mu }}_{i}}^{\mathrm {T} }{\Sigma _{i}}^{-1}{\vec {x}}} because both sides are scalar and transpose to each other ( Σ i {\displaystyle \Sigma _{i}} is Hermitian) and the above decision criterion becomes a threshold on the dot product w → T x → > c {\displaystyle {\vec {w}}^{\mathrm {T} }{\vec {x}}>c} for some threshold constant c, where w → = Σ − 1 ( μ → 1 − μ → 0 ) {\displaystyle {\vec {w}}=\Sigma ^{-1}({\vec {\mu }}_{1}-{\vec {\mu }}_{0})} c = 1 2 w → T ( μ → 1 + μ → 0 ) {\displaystyle c={\frac {1}{2}}\,{\vec {w}}^{\mathrm {T} }({\vec {\mu }}_{1}+{\vec {\mu }}_{0})} This means that the criterion of an input x → {\displaystyle {\vec {x}}} being in a class y {\displaystyle y} is purely a function of this linear combination of the known observations. It is often useful to see this conclusion in geometrical terms: the criterion of an input x → {\displaystyle {\vec {x}}} being in a class y {\displaystyle y} is purely a function of projection of multidimensional-space point x → {\displaystyle {\vec {x}}} onto vector w → {\displaystyle {\vec {w}}} (thus, we only consider its direction). In other words, the observation belongs to y {\displaystyle y} if corresponding x → {\displaystyle {\vec {x}}} is located on a certain side of a hyperplane perpendicular to w → {\displaystyle {\vec {w}}} . The location of the plane is defined by the threshold c {\displaystyle c} . == Assumptions == The assumptions of discriminant analysis are the same as those for MANOVA. The analysis is quite sensitive to outliers and the size of the smallest group must be larger than the number of predictor variables. Multivariate normality: Independent variables are normal for each level of the grouping variable. Homogeneity of variance/covariance (homoscedasticity): Variances among group variables are the same across levels of predictors. Can be tested with Box's M statistic. It has been suggested, however, that linear discriminant analysis be used when covariances are equal, and that quadratic discriminant analysis may be used when covariances are not equal. Independence: Participants are assumed to be randomly sampled, and a participant's score on one variable is assumed to be independent of scores on that variable for all other participants. It has been suggested that discriminant analysis is relatively robust to slight violations of these assumptions, and it has also been shown that discriminant analysis may still be reliable when using dichotomous variables (where multivariate normality is often violated). == Discriminant functions == Discriminant analysis works by creating one or more linear combinations of predictors, creating a new latent variable for each function. These functions are called discriminant functions. The number of functions possible is either N g − 1 {\displaystyle N_{g}-1} where N g {\displaystyle N_{g}} = number of groups, or p {\displaystyle p} (the number of predictors), whichever is smaller. The first function created maximizes the differences between groups on that function. The second function maximizes differences on that function, but also must not be correlated with the previous function. This continues with subsequent functions with the requirement that the new function not be correlated with any of the previous functions. Given group j {\displaystyle j} , with R j {\displaystyle \mathbb {R} _{j}} sets of sample space, there is a discriminant rule such that if x ∈ R j {\displaystyle x\in \mathbb {R} _{j}} , then x ∈ j {\displaystyle x\in j} . Discriminant analysis then, finds “good” regions of R j {\displaystyle \mathbb {R} _{j}} to minimize classification error, therefore leading to a high percent correct classified in the classification table. Each function is given a discriminant score to determine how well it predicts group placement. Structure Corr Macromedia FreeHand (formerly Aldus FreeHand) is a discontinued computer application for creating two-dimensional vector graphics oriented primarily to professional illustration, desktop publishing and content creation for the Web. FreeHand was similar in scope, intended market, and functionality to Adobe Illustrator, CorelDRAW and Xara Designer Pro. Because of FreeHand's dedicated page layout and text control features, it also compares to Adobe InDesign and QuarkXPress. Professions using FreeHand include graphic design, illustration, cartography, fashion and textile design, product design, architects, scientific research, and multimedia production. FreeHand was created by Altsys Corporation in 1988 and licensed to Aldus Corporation, which released versions 1 through 4. In 1994, Aldus merged with Adobe Systems and because of the overlapping market with Adobe Illustrator, FreeHand was returned to Altsys by order of the Federal Trade Commission. Altsys was later bought by Macromedia, which released FreeHand versions 5 through 11 (FreeHand MX). In 2005, Adobe Systems acquired Macromedia and its product line which included FreeHand MX, under whose ownership it presently resides. Since 2003, FreeHand development has been discontinued; in the Adobe Systems catalog, FreeHand has been replaced by Adobe Illustrator. FreeHand MX continues to run under Windows 11 and under Mac OS X 10.6 (Snow Leopard) within Rosetta, a PowerPC code emulator, and requires a registration patch supplied by Adobe. FreeHand 10 runs without problems on Mac OS X Snow Leopard with Rosetta enabled, and does not require a registration patch. Later versions of macOS can use a Mac OS X Snow Leopard Server virtual machine to emulate the required PowerPC support. == History == === Altsys and Aldus FreeHand === In 1984, James R. Von Ehr founded Altsys Corporation to develop graphics applications for personal computers. Based in Plano, Texas, the company initially produced font editing and conversion software; Fontastic Plus, Metamorphosis, and the Art Importer. Their premier PostScript font-design package, Fontographer, was released in 1986 and was the first such program on the market. With the PostScript background having been established by Fontographer, Altsys also developed FreeHand (originally called Masterpiece) as a Macintosh Postscript-based illustration program that used Bézier curves for drawing and was similar to Adobe Illustrator. FreeHand was announced as "... a Macintosh graphics program described as having all the features of Adobe's Illustrator plus drawing tools such as those in Mac Paint and Mac Draft and special effects similar to those in Cricket Draw." Seattle's Aldus Corporation acquired a licensing agreement with Altsys Corporation to release FreeHand along with their flagship product, Pagemaker, and Aldus FreeHand 1.0 was released in 1988. FreeHand's product name used intercaps; the F and H were capitalized. The partnership between the two companies continued with Altsys developing FreeHand and with Aldus controlling marketing and sales. After 1988, a competitive exchange between Aldus FreeHand and Adobe Illustrator ensued on the Macintosh platform with each software advancing new tools, achieving better speed, and matching significant features. Windows PC development also allowed Illustrator 2 (aka, Illustrator 88 on the Mac) and FreeHand 3 to release Windows versions to the graphics market. FreeHand 1.0 sold for $495 in 1988. It included the standard drawing tools and features as other draw programs including special effects in fills and screens, text manipulation tools, and full support for CMYK color printing. It was also possible to create and insert PostScript routines anywhere within the program. FreeHand performed in preview mode instead of keyline mode but performance was slower. FreeHand 2.0 sold for $495 in 1989. Besides improving on the features of FreeHand 1.0, FreeHand 2 added faster operation, Pantone colors, stroked text, flexible fill patterns and automatically import graphic assets from other programs. It added accurate control over a color monitor screen display, limited only by its resolution. FreeHand 3.0 sold for $595 in 1991. New features included resizable color, style, and layer panels including an Attributes menu. Also tighter precision of both the existing tools and aligning of objects. FH3 created compound Paths. Text could be converted to paths, applied to an ellipse, or made vertical. Carried over from version 1.0, FreeHand 3 suffered by having text entered into a dialog box instead of directly to the page. In October 1991, a 3.1 upgrade made FreeHand work with System 7 but additionally, it supported pressure-sensitive drawing which offered varying line widths with a users stroke. It improved element manipulation and added more import/export options. FreeHand 4.0 sold for $595 in 1994. Altsys ported FreeHand 3.0 to the NeXT system creating a new program named Virtuoso. Virtuoso continued its development at Altsys and version 2.0 of Virtuoso was feature-equivalent to FreeHand 4 (with the addition of NeXT-specific features such as Services and Display PostScript) and file compatible, with Virtuoso 2 able to open FreeHand 4 files and vice versa. A prominent feature of this version was the ability to type directly into the page and wrap inside or outside any shape. It also included drag-and-drop color imaging, a larger pasteboard, and a user interface that featured floating, rollup panels. The colors palette included a color mixer for adding new colors to the swatch list. Speed increases were made. In the same year of FreeHand 4 release, Adobe Systems announced merger plans with Aldus Corporation for $525 million. Fear about the end of competition between these two leading applications was reported in the media and expressed by customers (Illustrator versus FreeHand and Adobe Photoshop versus Aldus PhotoStyler.) Because of this overlapping of the market, Altsys stepped in by suing Aldus, saying that the merger deal was "a prima facie violation of a non-compete clause within the FreeHand licensing agreement." Altsys CEO Jim Von Ehr explained, "No one loves FreeHand more than we do. We will do whatever it takes to see it survive." The Federal Trade Commission issued a complaint against Adobe Systems on October 18, 1994, ordering a divestiture of FreeHand to "remedy the lessening of competition resulting from the acquisition as alleged in the Commission's complaint," and further, the FTC ordering, "That for a period of ten (10) years from the date on which this order becomes final, respondents shall not, without the prior approval of the Commission, directly or indirectly, through subsidiaries, partnerships, or otherwise .. Acquire any Professional Illustration Software or acquire or enter into any exclusive license to Professional Illustration Software;" (referring to FreeHand.) FreeHand was returned to Altsys with all licensing and marketing rights as well as Aldus FreeHand's customer list. === Macromedia Freehand === By late 1994, Altsys still retained all rights to FreeHand. Despite brief plans to keep it in-house to sell it along with Fontographer and Virtuoso, Altsys reached an agreement with the multimedia software company, Macromedia, to be acquired. This mutual agreement provided FreeHand and Fontographer a new home with ample resources for marketing, sales, and competition against the newly merged Adobe-Aldus company. Altsys would remain in Richardson, Texas, but would be renamed as the Digital Arts Group of Macromedia and was responsible for the continued development of FreeHand. Macromedia received FreeHand's 200,000 customers and expanded its traditional product line of multimedia graphics software to illustration and design graphics software. CEO James Von Ehr became a Macromedia vice-president until 1997 when he left to start another venture. FreeHand 5.0 sold for $595 in 1995. This version featured a more customizable and expanded workspace, multiple views, stronger design and editing tools, a report generator, spell check, paragraph styles, multicolor gradient fills up to 64 colors, speed improvements, and it accepted Illustrator plugins. In September 1995, a 5.5 upgrade added Photoshop plug-in support, PDF import capabilities, the Extract feature, inline graphics to text, improved auto-expanding text containers, the Crop feature, and the Create PICT Image feature. A FreeHand 5.5 upgrade was part of the FreeHand Graphics Studio (a suite that included Fontographer, Macromedia xRes image editing application, and Extreme 3D animation and modeling application). FreeHand 6.0 in 1996. This version only existed in beta. Some Freehand 7 prerelease versions were released under the Freehand 6 tag. FreeHand 7.0 sold for $399 in 1996, or $449 as part of the FreeHand Graphics Studio (see above.) Features included a redesigned user interface that allowed recombining Inspectors, Panel Tabs, Dockable Panels, Smart Cursors, Sharpness Aware Minimization (SAM) is an optimization algorithm used in machine learning that aims to improve model generalization. The method seeks to find model parameters that are located in regions of the loss landscape with uniformly low loss values, rather than parameters that only achieve a minimal loss value at a single point. This approach is described as finding "flat" minima instead of "sharp" ones. The rationale is that models trained this way are less sensitive to variations between training and test data, which can lead to better performance on unseen data. The algorithm was introduced in a 2020 paper by a team of researchers including Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. == Underlying Principle == SAM modifies the standard training objective by minimizing a "sharpness-aware" loss. This is formulated as a minimax problem where the inner objective seeks to find the highest loss value in the immediate neighborhood of the current model weights, and the outer objective minimizes this value: min w max ‖ ϵ ‖ p ≤ ρ L train ( w + ϵ ) + λ ‖ w ‖ 2 2 {\displaystyle \min _{w}\max _{\|\epsilon \|_{p}\leq \rho }L_{\text{train}}(w+\epsilon )+\lambda \|w\|_{2}^{2}} In this formulation: w {\displaystyle w} represents the model's parameters (weights). L train {\displaystyle L_{\text{train}}} is the loss calculated on the training data. ϵ {\displaystyle \epsilon } is a perturbation applied to the weights. ρ {\displaystyle \rho } is a hyperparameter that defines the radius of the neighborhood (an L p {\displaystyle L_{p}} ball) to search for the highest loss. An optional L2 regularization term, scaled by λ {\displaystyle \lambda } , can be included. A direct solution to the inner maximization problem is computationally expensive. SAM approximates it by taking a single gradient ascent step to find the perturbation ϵ {\displaystyle \epsilon } . This is calculated as: ϵ ( w ) = ρ ∇ L train ( w ) ‖ ∇ L train ( w ) ‖ 2 {\displaystyle \epsilon (w)=\rho {\frac {\nabla L_{\text{train}}(w)}{\|\nabla L_{\text{train}}(w)\|_{2}}}} The optimization process for each training step involves two stages. First, an "ascent step" computes a perturbed set of weights, w adv = w + ϵ ( w ) {\displaystyle w_{\text{adv}}=w+\epsilon (w)} , by moving towards the direction of the highest local loss. Second, a "descent step" updates the original weights w {\displaystyle w} using the gradient calculated at these perturbed weights, ∇ L train ( w adv ) {\displaystyle \nabla L_{\text{train}}(w_{\text{adv}})} . This update is typically performed using a standard optimizer like SGD or Adam. == Application and Performance == SAM has been applied in various machine learning contexts, primarily in computer vision. Research has shown it can improve generalization performance in models such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) on image datasets including ImageNet, CIFAR-10, and CIFAR-100. The algorithm has also been found to be effective in training models with noisy labels, where it performs comparably to methods designed specifically for this problem. Some studies indicate that SAM and its variants can improve out-of-distribution (OOD) generalization, which is a model's ability to perform well on data from distributions not seen during training. Other areas where it has been applied include gradual domain adaptation and mitigating overfitting in scenarios with repeated exposure to training examples. == Limitations == A primary limitation of SAM is its computational cost. By requiring two gradient computations (one for the ascent and one for the descent) per optimization step, it approximately doubles the training time compared to standard optimizers. The theoretical convergence properties of SAM are still under investigation. Some research suggests that with a constant step size, SAM may not converge to a stationary point. The accuracy of the single gradient step approximation for finding the worst-case perturbation may also decrease during the training process. The effectiveness of SAM can also be domain-dependent. While it has shown benefits for computer vision tasks, its impact on other areas, such as GPT-style language models where each training example is seen only once, has been reported as limited in some studies. Furthermore, while SAM seeks flat minima, some research suggests that not all flat minima necessarily lead to good generalization. The algorithm also introduces the neighborhood size ρ {\displaystyle \rho } as a new hyperparameter, which requires tuning. == Research, Variants, and Enhancements == Active research on SAM focuses on reducing its computational overhead and improving its performance. Several variants have been proposed to make the algorithm more efficient. These include methods that attempt to parallelize the two gradient computations, apply the perturbation to only a subset of parameters, or reduce the number of computation steps required. Other approaches use historical gradient information or apply SAM steps intermittently to lower the computational burden. To improve performance and robustness, variants have been developed that adapt the neighborhood size based on model parameter scales (Adaptive SAM or ASAM) or incorporate information about the curvature of the loss landscape (Curvature Regularized SAM or CR-SAM). Other research explores refining the perturbation step by focusing on specific components of the gradient or combining SAM with techniques like random smoothing. Theoretical work continues to analyze the algorithm's behavior, including its implicit bias towards flatter minima and the development of broader frameworks for sharpness-aware optimization that use different measures of sharpness.Linear discriminant analysis
Macromedia FreeHand
Sharpness aware minimization