(Please visit the site to view this video)
This week Bruno sits down with Bernie Brennan, co-author of Branded! How Retailers Engage Consumers with Social Media and Mobility , and discusses how the customer has been put in the driver’s seat when it comes to retail shopping because of their unprecedented access to vast amounts of data. Bernie’s mission when he wrote the book was to explain to retailers how they can use this customer access in a positive way instead of worrying about…(read more)

  Question (paraphrased):  “Conor, does changing the max server memory settings always empty the plan cache on SQL 2005 and 2008?” Answer: yes, this is what happens and what is supposed to happen.  One of the inputs to the optimizer is the max server memory, and it is used in the costing functions to determine which plan gets chosen during optimization.  Changing this setting could imply that other plans are now better.  As such, the plan cache is emptied and plans are compiled…(read more)

SVG Open 2011, the 9th International Conference on Scalable Vector Graphics, is just over two weeks away. This year’s theme, “Where SVG Meets the Web,” underscores SVG’s inclusion in HTML5 and its growing importance on the Web. SVG is also the declarative vector graphics format for Windows 8 Metro style apps built using HTML.

The conference is an annual showcase of SVG innovation bringing together the leaders who drive SVG forward. Microsoft is honored to host this year’s conference at our New England Research & Development Center in the heart of Cambridge, Massachusetts. Joining us as sponsors are Google, Adobe, IBM, Slippery Rock University, Rochester Institute of Technology, Harvard’s Institute for Applied Computational Science, and the W3C.

SVG Open presenters are recognized for widening the use of SVG on the Web. This year’s tracks feature SVG Authoring Tools, SVG for Charting and Visualization, SVG for Games and Interactive Applications, SVG for Web Applications and HMI, Publishing and Printing with SVG, SVG Implementations and Implementation Issues, SVG for Webmapping and GIS, and a set of Case Studies of SVG usage.

Also attending are members of the W3C SVG Working Group who will share information, solicit input, and respond to questions about features proposed for SVG 2.0, as well as the integration of certain CSS and SVG features as contemplated by the CSS-SVG Effects Task Force. Two areas of particular interest are applying SVG Filter Effects to HTML elements and CSS Animations and Transitions to SVG elements.

To those who are attending, I hope to meet you. Jennifer Yu, also of the IE team, and I are both attending all three days of the conference.

—Patrick Dengler, Senior Program Manager, Internet Explorer, and member, W3C SVG Working Group and CSS-SVG Effects Task Force

Data Gathering – SQL Server Setup Problems

Cindy Gross, Dedicated Support Engineer

Gathering the right data during a problem can get you a long ways towards resolving the problem. When you ask the right questions and clearly define the problem it changes the way you approach the remaining steps. Sometimes the answer to a seemingly simple question leads you right to the solution. In this series of data gathering blogs I am NOT going to tell you how to solve the problem, but I am…(read more)

http://msdn.microsoft.com/es-es/library/ee405484.aspx

Windows® Azure™ Tools para Microsoft® Visual Studio® extiende Visual Studio para permitir la creación, compilación, depuración, ejecución y empaquetado de servicios y aplicaciones web escalables en Windows Azure.

Windows Azure Tools está disponible para estos productos de Visual Studio:

• Visual Studio 2010
• Visual Web Developer 2010

Para mantenerse al día sobre todas las actualizaciones del kit para Windows Azure, vea http://msdn.microsoft.com/es-es/library/ff683673.aspx

?????????????????Android ??????????????????????

10/5 CEATEC JAPAN <http://www.ceatec.com/2011/ja/index.html>

10/17 ?????????

10/22-23 Microsoft ????????

10/28 ??????? ????

10/29 ???????????????

11/4 ?MOSA Software Meeting ?<http://www.mosa.gr.jp/?p=3718>

?????????????????????????????????????????????????CEATEC?

Events are available around the United States for developers to learn about Windows Phone. MSDN Events is offering one-day and two-day sessions around the US.

Register at MSDN Events.

…(read more)

Die CeBIT ist die weltweit größte Messe zur Darstellung digitaler Lösungen aus der Informations- und Kommunikationstechnik für die Arbeits- und Lebenswelt. Besucherzielgruppen sind Anwender aus Industrie, Handel, Handwerk, Banken, dem Dienstleistungsgewerbe, der öffentlichen Verwaltung und der Wissenschaft sowie alle technikbegeisterten Nutzer. Für Microsoft hat die CeBIT einen sehr hohen Stellenwert.

Auch das Konzept rund um unseren Messeauftritt bei der CeBIT 2011 hat sich bewährt. Während der fünf Messetage haben wir mehr als 238.000 Besucher auf dem Microsoft-Messestand begrüßen dürfen. Ein Anstieg von über 18 % gegenüber dem Vorjahr. Bei mehr als 60 % unserer Besucher steht der Besuch zumindest mittelbar im Zusammenhang mit Investitionsplänen.

Nach der sehr erfolgreichen CeBIT 2011 wird Microsoft auch im Frühjahr 2012 wieder auf der weltgrößten ITK – Messe vertreten sein. Mit dem Schwerpunkt „Managing Trust“ rückt die CeBIT 2012 als weltweit wichtigste Veranstaltung der digitalen Wirtschaft das Thema Vertrauen und Sicherheit in der digitalen Welt in den Mittelpunkt.

Ihnen als Microsoft-Partner möchten wir gern die exklusive Gelegenheit bieten, Ihre Lösungen auf unserem Stand zu präsentieren.

Mit dem innovativen Standkonzept für die CeBIT 2012 orientieren wir uns noch mehr an der Sprache des Marktes und damit an den Bedürfnissen unserer gemeinsamen Kunden. Dafür haben wir die folgenden übergeordneten Fokusthemen identifiziert, die wir eng mit unseren Produkt- und Lösungsangeboten verknüpfen.

• Desktoplösungen mit Microsoft Office und Windows 7 
• Infrastrukturlösungen mit Servern 
• Lösungen mit Microsoft-Geschäftsanwendungen 

Diesen Fokusthemen entspricht die Platzierung unserer Produktstände und Ihrer Partner-Lösungsstände. Unsere Kunden haben weiterhin direkten Zugriff auf die Microsoft-Kollegen und Ihnen als Partner. Gemeinsam präsentieren wir uns als ein Team. So kommt der Besucher schnell vom Produkt zur Partnerlösung.

Ein enormes Potenzial, das Ihnen als Messeteilnehmer erfolgreiche Projekte sichert. Für Ihre Teilnahme als Aussteller auf der Messe haben wir für Sie attraktive Angebotspakete geschnürt. Detailinformationen finden Sie in der Datei unten.

Lassen Sie uns frühzeitig Ihre Bewerbung zur CeBIT 2012 zukommen und sichern Sie sich so die Möglichkeit für Ihren Partnerplatz auf dem Microsoft Stand.
Der Einsendeschluss für Ihre Bewerbung ist Freitag, 25. November 2011.

Bitte senden Sie Ihre schriftliche Bewerbung an:

Microsoft Deutschland GmbH
Microsoft CeBIT Partner Service
Frau Yvonne Albers
Tel.: +49 (0)89 3176-5351
Fax: +49 (0)89 3176-3032
E-Mail: v-yvalbe@microsoft.com

Leider werden uns nicht alle Bewerber auf die CeBIT begleiten können. Auch werden wir nicht alle Wünsche nach einem bestimmten Standbereich oder einer besonderen Platzierung berücksichtigen können. Bei der Auswahl und der Standvergabe an die Teilnehmer sind Anforderungen der jeweiligen Microsoft-Geschäftsbereiche zu bedenken, zum Beispiel die strategische Ausrichtung, die Abdeckung entscheidender Branchen oder die Verfügbarkeit der jeweiligen Kontingente. Bitte haben Sie daher Verständnis, dass wir Ihnen vorab keine Zusage für die Teilnahme oder die Platzierung Ihres Unternehmens geben können. Nach Prüfung aller Anmeldungen und der verfügbaren Kontingente lassen wir Ihnen eine verbindliche Entscheidung zukommen.

This blog has been migrated from community.research.microsoft.com

Original date of blog: November 24, 2010

Original blog author: David Knowles

The new release of Infer.NET supports a number of new factors you may find useful when designing your models. Some of these features have been well tested, while others are still experimental and should be used with caution. How do you know which features are experimental I hear you ask? To answer just that question we have introduced quality bands, which are assigned to every piece of functionality. By default you’ll get a warning if your model is using any experimental features. For more information and guidance on quality bands take a look at the user guide and the list of factors. 

In this blog post I’ll give one example of a new feature, the softmax factor, which lets you perform efficient multinomial logistic regression using Variational Message Passing (VMP). The inputs to the regression are continuous variables, and the output is categorical, i.e. belongs to one of C classes. One example would be predicting what political party someone will vote for based on some known features such as age and income. Another would be classifying cancer type based on gene expression data.

Multinomial regression involves constructing a separate linear regression for each class, the output of which is known as an auxiliary variable. The vector of auxiliary variables for all classes is put through the softmax function to give a probability vector, corresponding to the probability of being in each class. The factor graph for this model is shown below (see John G.’s blog post on Bayesian PCA for more on factor graphs).

Multinomial softmax regression is very similar in spirit to the multiclass Bayes Point Machine (BPM) which we also have as an example. The key differences are that where the BPM uses a max operation constructed from multiple “greater than” factors, multinomial regression uses the softmax factor, and that where Expectation Propagation is the recommended algorithm  for the BPM (although VMP is now supported), VMP is currently the only option when using the softmax factor. Two consequence of these differences is that multinomial regression scales computationally much better with the number of classes than the BPM (linear rather quadratic complexity) and it is possible to have multiple counts for a single sample. An example of where these differences would make multinomial regression appropriate would be predicting word counts in documents, based on some known features such as date of publication. The number of classes C would be the size of the vocabulary, which could be very large, making the quadratic O(C²) scaling of the BPM prohibitive. The outputs are counts rather than single class assignments, which are straightforward to model with multinomial regression using the multinomial factor.

Now let’s set up our model in Infer.NET. First we’ll need ranges over the C classes and N samples.

var c = new Range(C);
var n = new Range(N);

Here I’ve used lower case to denote the Infer.NET ranges, and upper case to denote the (fixed) size of the range. We need a separate vector of coefficients B and mean m for each class c . The coefficient vectors B have length K, which is the number of features (inputs) to the model.

var B = Variable.Array<Vector>(c);
B[c] = Variable.VectorGaussianFromMeanAndPrecision(Vector.Zero(K),
PositiveDefiniteMatrix.Identity(K)).ForEach(c);
var m = Variable.Array<double>(c);
m[c] = Variable.GaussianFromMeanAndPrecision(0, 1).ForEach(c);

var x = Variable.Array<Vector>(n);
x.ObservedValue = xObs;
var yData = Variable.Array(Variable.Array<int>(c), n);
yData.ObservedValue = yObs;
var trialsCount = Variable.Array<int>(n);
trialsCount.ObservedValue = yObs.Select(o => o.Sum()).ToArray();

The variable trialsCount is simply the total count for each sample, which I’ve calculated on the fly here using a little Linq magic. We are now ready to write down the core of the model: we create the auxiliary variables g, pass them through the softmax factor, and use the resulting probability vectors in the multinomial factor. If you have an individual class assignment for each sample rather than counts you could use the Discrete rather than Multinomial factor.

var g = Variable.Array(Variable.Array<double>(c), n);
g[n][c] = Variable.InnerProduct(B[c], x[n]) + m[c];
var p = Variable.Array<Vector>(n);
p[n] = Variable.Softmax(g[n]);
using (Variable.ForEach(n))
     yData[n] = Variable.Multinomial(trialsCount[n], p[n]);

A problem we need to address here is that the model is not identifiable as we currently have it set up. You can see this because adding a constant value to all the auxiliary variables g will not change the output of the prediction. To make the model identifiable we enforce that the last auxiliary variable (corresponding to the last class) will be zero, by constraining the last coefficient vector and mean to be zero:

Variable.ConstrainEqual(B[C - 1], Vector.Zero(K));
Variable.ConstrainEqual(m[C - 1], 0);

Finally we run inference over the model:

var ie = new InferenceEngine(new VariationalMessagePassing());
ie.Compiler.GivePriorityTo(typeof(SaulJordanSoftmaxOp_NCVMP));
var bPost = ie.Infer<VectorGaussian[]>(B);
var meanPost = ie.Infer<Gaussian[]>(m);

A specific handling of the softmax factor is required because its input is deterministic: this is specified by the GivePriorityTo function. If you don’t do this you’ll get an error saying the input to the softmax function cannot be deterministic. For some factors multiple implementations are available with different characteristics: GivePriorityTo allows you to choose which version you would like to use. The default implementation for the softmax factor assumes the input is stochastic which allows a faster converging algorithm to be used. 

Let’s test this model using synthetic data drawn from the same model with six features, four classes and a total count of ten per sample. For a sample size of 1000, the following plot shows the accuracy of the inferred coefficients and mean. We have recovered the coefficients really well: but of course we had quite a lot of data! Running 50 iterations of VMP took only three seconds.

A good sanity check is to see how the inference performance improves with increasing sample size. The following plot shows the improvement in the RMSE (root mean squared error) between the true and mean inferred coefficients. Of course, we would expect this to decrease as the number of samples increases because each sample gives us some more information to use. Indeed we see that this is the case: error rapidly decreases with sample size. Note the decreasing returns, which is typical of statistical estimation of unknown quantities: initially additional samples help us a lot, but once we have a reasonable amount of data, new samples don’t tell us much we didn’t already know. 

The code including generating synthetic data, setting up the model, running inference and computing errors is available here. Feel free to download it, try it out on your own data or modify/extend the model however you please. If you don’t have it yet, be sure to install the latest Infer.NET release.

I hope this blog has shown you the potential of just one of the new factors in the new release, the softmax factor. The possibilities extend far beyond simple multinomial regression. The great thing about the softmax factor is that any time you want to model probability vectors with complex dependencies, you can model those dependencies in continuous space and map to probability vectors using the softmax factor. A model of this type in the machine learning literature is the correlated topic model, a version of Latent Dirichlet Alocation where the document specific distributions over topics are represented as continuous vectors drawn from a learnt multivariate Gaussian, and are put through a softmax function to give valid probability distiributions.

I hope you’ve enjoyed reading about just one new piece of functionality available in the new release of Infer.NET, and that you feel inspired to get stuck in designing your own probabilistic models! 

Happy modelling! 

David K and the Infer.NET team.