Deep Neural Networks and Restricted Boltzmann Machines


The new version of the Accord.NET brings a nice addition for those working with machine learning and pattern recognition: Deep Neural Networks and Restricted Boltzmann Machines.

Class diagram for Deep Neural Networks in the Accord.Neuro namespace.

Deep neural networks have been listed as a recent breakthrough in signal and image processing applications, such as in speech recognition and visual object detection. However, is not the neural networks which are the new things here; but rather, the learning algorithms. Neural Networks have existed for decades, but previous learning algorithms were unsuitable to learn networks with more than one or two hidden layers.

But why more layers?

The Universal Approximation Theorem (Cybenko 1989; Hornik 1991) states that a standard multi-layer activation neural network with a single hidden layer is already capable of approximating any arbitrary real function with arbitrary precision. Why then create networks with more than one layer?
To reduce complexity. Networks with a single hidden layer may arbitrarily approximate any function, but they may require an exponential number of neurons to do so. We can borrow a more tactile example from the electronics field. Any boolean function can be expressed using only a single layer of AND, OR and NOT gates (or even only NAND gates). However, one would hardly use only this to fully design, let’s say, a computer processor. Rather, specific behaviors would be modeled in logic blocks, and those blocks would then be combined to form more complex blocks until we create a all-compassing block implementing the entire CPU.
The use of several hidden layers is no different. By allowing more layers we allow the network to model more complex behavior with less activation neurons; futhermore the first layers of the network may specialize on detecting more specific structures to help in the later classification. Dimensionality reduction and feature extraction could have been performed directly inside the network on its first layers rather than using specific separate algorithms. 

Do computers dream of electric sheep?

The key insight in learning deep networks was to apply a pre-training algorithm which could be used to tune individual hidden layers separately. Each layer is learned separately without supervision. This means the layers are able to learn features without knowing their corresponding output label. This is known as a pre-training algorithm because, after all layers have been learned unsupervised, a final supervised algorithm is used to fine-tune the network to perform the specific classification task at hand.

As shown in the class diagram on top of this post, Deep Networks are simply cascades of Restricted Boltzmann Machines (RBMs). Each layer of the final network is created by connecting the hidden layers of each RBM as if they were hidden layers of a single activation neural network.

Now, the most interesting part about this approach will given now. It is about one specific detail on how the RBMs are learned, which in turn allows a very interesting use of the final networks. As each layer is a RBM learned using an unsupervised algorithm, they can be seen as standard generative models. And if they are generative, they can be used to reconstruct what they have learned. And by sequentially alternating computation and reconstruction steps initialized with a random observation vector, the networks may produce patterns which have been created using solely they inner knowledge about the concepts it has learned. This may be seen fantastically close to the concept of a dream.

At this point I would also like to invite you to watch the video linked above. And if you like what you see, I also invite you to download the latest version of the Accord.NET Framework and experiment with those newly added features.

The new release also includes k-dimensional trees, also known as kd-trees, which can be use to speed up nearest neighbor lookups in algorithms which need it. They are particularly useful in algorithms such as the mean shift algorithm for data clustering, which has been included as well; and in instance classification algorithms such as the k-nearest neighbors.

Accord.NET and .NET 3.5


Quite some time ago, Accord.NET had been upgraded to .NET 4.0. This upgrade was made to advantage of some of the unique platform features. However, this was a drawback for those users who were still bound to .NET 3.5 and would like to use (or continue using) the framework.

So, some weeks ago I did finish creating a compatibility layer providing .NET3.5 support for the newest version of the framework. While some features had to be left aside, most of the framework is currently functional. A few features, however, had to be removed. Those include the Accord.Audio modules,  support for CancellationTokens on processor-intensive learning algorithms, some parallel processing capabilities and some of the Expression Tree manipulation in some numeric optimization algorithms and in Decision Trees.

With time, some of those could be re-implemented to work correctly under 3.5 as needed. You can leave a request, if you would like. Stay tuned!

The Mistery of Failing Tests – Or how to implement IDisposable correctly


Some days ago my Visual Studio tests started failing randomly. Curiously, the same tests which failed in a previous run would succeed if ran again. There seemed to be nothing wrong with the tests, except for the message “The agent process was stopped while the test was running” appearing the Test Results window.

It turns out this message tends to appear whenever an object throws an exception inside its destructor.

Luckly, only about a dozen classes in Accord.NET do need to implement finalizers. After a few searches, it turned down the symptoms were being caused by a NullReferenceException being thrown in one of those classes.

The solution is to always implement the Disposable pattern correctly.

Disposing managed code correctly

To do so, first, implement IDisposable. Consider, for example, the Metronome class, which makes use of timers:

Then, implement the required Dispose method, but do not write the dispose code right away. Instead, write:

Always check if a member is not null before disposing it. Alternatively, after disposing, always set it to null so you can’t accidentaly dispose it twice. The GC.SuppressFinalize instructs the Garbage Collector that the object has already been disposed, so it won’t be required to call its finalizer. The GC will be trusting you have already cleaned the room; so don’t fail on it!

Finally, write the finalizer as:

Disposing unmanaged code

The disposal of unmanaged code follows exactly the same pattern as above, but some care is required when disposing unmanaged resources. Consider the Signal class, which represents an audio signal in PCM format:

Note that this class uses an unmanaged memory block to store the audio signal. To dispose it properly, the Disposable pattern has to be implemented as it follows:

Here we have almost the same pattern. However, note that unmanaged code should always be freed, even if Dispose has not been called explictly. So the code for releasing unmanaged resources is executed unconditionally in the Dispose(bool) method.

Haar Feature Face Detection in C#


A new article has been published in CodeProject! This article details the Viola-Jones face detection algorithm available in the Accord.NET Framework. The article page also provides a standalone package for face detection which can be reused without instantiating the entire framework.


The article is also participating in CodeProject’s monthly competition under the categories Best C# Article and Best Overall Article of the month! Feel free to cast your vote if you liked and otherwise found them useful!  
This month we have other great articles participating in the competition: Marcelo de Oliveira‘s Social News excels in the Web category; Roy‘s Inline MSIL in C# also presents a very interesting reading in the C# category. The later may eventually be extremely useful to leverage performance in managed applications. 
For those who don’t know, CodeProject is a amazingly useful site which publishes user-created articles and news. Every month, the best articles among all submissions are selected to win prizes and gain popularity as well!

Always test your hypothesis…


… but also always make sure to interpret results correctly! This post presents a quick intro on how to perform statistical hypothesis testing and power analysis with the Accord.NET Framework in C#.


  1. Hypothesis testing
    1. Statistical hypothesis testing
    2. My test turned insignificant.
      Should I accept the null hypothesis?
    3. Further criticism
  2. The Accord.NET Framework
    1. Available tests
    2. Example problems
  3. Suggestions
  4. References


Hypothesis testing

What does hypothesis testing means in statistics (and should also mean everywhere, for that matter)? You may recall from Karl Popper’s theory of falsiability that good theories can rarely be accurately proven, but you may gain a considerable confidence on them by constantly challenging and failing to refute them.

This comes from the fact that it is often easier to falsify something than to prove it. Consider for instance the white-swan/black-swan example: Let’s say a theory states that all swans are white. This is a very strong claim; it does not applies to one or a few particular observations of a swan, but all of them. It would be rather difficult to verify if all of the swans in Earth are indeed white. It is thus almost impossible to prove this theory directly.

However, the catch is that it will take only a single contrary example to refute it. If
we find a single swan that is black, the entire theory should be rejected, so alternate theories could be raised. It should be fairly easy to attempt to prove a theory wrong. If a theory continuously survives those attempts to be proven wrong, it becomes stronger. This does not necessarily means it is correct, only that it is very unlikely to be wrong.

This is pretty much how the science method works; it also provides a solution to the demarcation problem originally proposed by Kant: the problem of separating the sciences from the pseudo-sciences (i.e. astronomy from astrology). A “good” theory should be easy to attack so we can try to refute it; and by constantly challenging it and failing to prove it wrong, we gain further confidence that this theory may,
indeed, be right. In short:

Often the most interesting theories can’t be proven right, they can only be proven wrong. By continuously refuting alternatives, a theory becomes stronger (but most likely never reaching the ‘truth’).

Answering the question in the first phrase of this section, hypothesis testing means verifying if a theory holds even when confronted with alternative theories. In statistical hypothesis testing, this often means checking if a hypothesis holds even when confronted with the fact that it may have just happened to be true by pure chance or plain luck.

Statistical hypothesis testing

Fisher (1925) also noted that we can’t always prove a theory but we can attempt to refute it. Therefore, statistical hypothesis testing includes stating a hypothesis, which is the hypothesis we are trying to invalidade; and check if we can confidently reject it by confronting it with data from a test or experiment. This hypothesis to be pickpocketed is often called the null hypothesis (commonly denoted H0). It receives this name as it is usually the hypothesis of no change: there is no difference, nothing changed after the experiment, there is no effect.

The hypotheses verified by statistical hypothesis tests are often theories about whether or not a random sample from a population comes from a given probability distribution. This seems weird, but several problems can be cast in this way. Suppose, for example, we would like to determine if students from a classroom have significantly different grades than students from another room. Any difference could
possibly be attributed to chance, as some students may just perform better on a exam because of luck.

An exam was applied to both classrooms. The exam results (by each student) are written below:

Classroom A Classroom B
8.12, 8.34, 7.54,
8.98, 8.24, 7.15,
6.60, 7.84, 8.68,
9.44, 8.83, 8.21,
8.83, 10.0, 7.94,
9.58, 9.44, 8.36,
8.48, 8.47, 8.02,
8.20, 10.0, 8.66,
8.48, 9.17, 6.54,
7.50, 6.70, 8.55,
7.84, 9.23, 6.10,
8.45, 8.27, 7.01,
7.18, 9.05, 8.18,
7.70, 7.93, 8.20,
8.19, 7.65, 9.25,
8.71, 8.34, 7.47,
7.47, 8.24, 7.10,
7.87, 10.0, 8.26,
6.82, 7.53
Students: 28
Mean: 8.416
Students: 29
Mean: 7.958

We have two hypothesis:

  • Results for classroom A are not significantly different from the results from classroom B. Any difference in means could have been explained due to chance alone.
  • Results are indeed different. The apparent differences are very unlikely to have occurred by chance.

Since we have less than 30 samples, we will be using a Two-Sample T-Test to test the hypothesis that the population’s mean values of the two samples are not equal. Besides, we will not be assuming equal variances. So let’s we create our test object:

And now we can query it:

Which reveals the test is indeed significant. And now we have at least two problems to address…


Problem 1: Statistical significance does not imply practical significance

So the test was significant. But would this mean the difference itself is significant?
Would this mean there any serious problem with the school teaching method?

No – but it doesn’t mean the contrary either. It is impossible to tell just by looking at the p-level.

The test only said there was a difference, but it can not tell the importance of this difference. Besides the two classes having performed so differently they could trigger statistical significance, we don’t know if this difference has any practical significance. A statistical test being significant is not a proof; it is just an evidence to be balanced together with other pieces of information in order to drawn a conclusion.

Perhaps one of best examples illustrating this problem is given by Martha K. Smith:

Suppose a large clinical trial is carried out to compare a new medical treatment with a standard one. The statistical analysis shows a statistically significant difference in lifespan when using the new treatment compared to the old one. But the increase in lifespan is at most three days, with average increase less than 24 hours, and with poor quality of life during the period of extended life. Most people would not consider the improvement practically significant.

In our classroom example, the difference in means is about 0.46 points. If principals believe a difference of less than 0.50 in a scale from 0.00 to 10.00 is not that critical, there may be no need to force students from the room with lower grades to start taking extra lessons after school. In other words, statistical hypothesis testing does not lead to automatic decision making. A statistically significant test is just another evidence which should be balanced with other clues in order to take a decision or
draw a conclusion

Problem 2: Powerless tests can be misleading

The p-level reported by the significance test is the probability of the extreme data we found be occurring given the null hypothesis is correct. Often, this is not the case. We
must also know the probability that the test will reject the null hypothesis when the null hypothesis is false. To do so, we must compute the power of our test, and, better yet, we should have used this information to conclude how many samples we would need to achieve a more informative test before we conducted our experiment. The power is then the probability of detecting a difference if this difference indeed exists (Smith, 2011). So let’s see:

The test we performed had astonishingly small power; so if the null hypothesis is false (and there is actually a difference between the classrooms) we have only about 50% chance of correctly rejecting it. Therefore, this would also lead to a 50% chance of producing a false negative – incorrectly saying there is no difference when actually there is.  The table below exemplifies the different errors we can get by rejecting or not the null hypothesis.

Null hypothesis is true Null hypothesis is false
Fail to reject the
null hypothesis
True negative
Type II error (beta)
False negative
Reject the null
Type I error (alpha)
False positive
Correct outcome
True positive

Tests with little statistical power are  often inconsistent in the literature. Suppose, for example, that the score from the first student from classroom B had earned a  7.52 instead of 7.50. Due to the low power of the test, this little change would already be sufficient to render the test nonsignificant, and we will not be able to reject the null hypothesis that the two population means aren’t different anymore. Due to the low power of the test, we can’t distinguish between a correct true negative and a type II error. This is why powerless tests can be misleading and should never be relied upon for decision making (Smith, 2011b).

The power of a test increases with the sample size. To obtain a power of at least 80%, let’s see how many samples should have been collected:

So, we would actually need 57 students in each classroom to confidently affirm whether there was an difference or not. Pretty disappointing, since in the real world we wouldn’t be able to enroll more students and wait years until we could perform another exam just to adjust the sample size. On those situations, the power for the test can be increased by increasing the significance threshold (Thomas, Juanes, 1996), although clearly sacrificing our ability to detect false positives (Type I errors) in the process.

My test turned insignificant. Should I accept the null hypothesis?

The short answer is ‘only if you have enough power‘. Otherwise, definitively no.

If you have reason to believe the test you performed had enough power to detect a difference within the given Type-II error rate, and it didn’t, then accepting the null would most likely be acceptable. The acceptance should also be accompanied of an analysis of confidence intervals or effect sizes. Consider, for example, that some actual scientific discoveries were  indeed made by accepting the null hypothesis rather than by contradicting it; one notable example being the discovery of the X-Ray (Yu, 2012).

Further criticism

Much of the criticism associated with statistical hypothesis testing is often related not to the use of statistical hypothesis testing per se, but on how the significance outcomes from such tests are interpreted. Often it boils down to incorrectly believing a p-value is the probability of a null hypothesis being true, when, in fact, it is only the probability of obtaining a test statistic as extreme as the one calculated from the data, always within the assumptions of the test under question.

Moreover, another problem may arise when we chose a null hypothesis which is obviously false. There is no point on hypothesis testing when the null hypothesis couldn’t possibly be true. For instance, it is very difficult to believe a parameter of a continuous distribution is exactly equal to an hypothesized value, such as zero. Given enough samples, it will always be possible to find a difference, as small as it gets, and the test will invariably turn significant. Under those specific circumstances, statistical testing can be useless as it has relationship to practical significance. That is why analyzing the effect size is important in order to determine the practical significance of an hypothesis test. Useful hypothesis would also need to be probable, plausible and falsifiable (Beaulieu-Prévost, 2005).

The following links also summarize much of the criticism in statistical hypothesis testing. The last one includes very interesting (if not enlightening) comments (in the comment section) on common criticisms of the hypothesis testing method.

The Accord.NET Framework

Now that we presented the statistical hypothesis testing framework, and now that the reader is aware of its drawbacks, we can start talking about performing those tests with the aid of a computer. The Accord.NET Framework is a framework for scientific computing with wide support for statistical and power analysis tests, without entering the merit if they are valid or no. In short, it provides some scissors;
feel free to run with them

Tests available in the framework (at the time of this writing)

As it may already have been noticed, the sample code included in the previous section was C# code using the framework. In the aforementioned example, we created a T-Test for comparing the population means of two samples drawn from Normal distributions. The framework, nevertheless, includes  many other tests, some with support for power analysis. Those include:

Parametric tests Nonparametric tests

Tests marked with a * are available in versions for one and two samples. Tests on the second row can be used to test hypothesis about contingency tables. Just remembering, the framework is open source and all code is available on GitHub.

A class diagram for the hypothesis testing module is shown in the picture below. Click for a larger version.

Class diagram for the Accord.Statistics.Testing namespace.

Framework usage should be rather simple. In order to illustrate it, the next section brings some example problems and their solution using the hypothesis testing module.

Example problems and solutions

Problem 1. Clairvoyant card game.

This is the second example from Wikipedia’s page on hypothesis testing. In this example, a person is tested for clairvoyance (ability of gaining information about something through extra sensory perception; detecting something without using the known human senses.

Problem 2. Worried insurance company

This is a common example with variations given by many sources. Some of them can be found here and here.

Problem 3. Differences among multiple groups (ANOVA)

This example comes from Wikipedia’s page on the F-test. Suppose we would like to study the effect of three different levels of a factor ona response (such as, for example, three levels of a fertilizer on plant growth. We have made 6 observations for each of the three levels a1, a2 and a3, and have written the results as in the table below.

The last line in the example shows the ANOVA table using the framework’s DataGridBox object. The DataGridBox is a convenience class for displaying DataGridViews just as one would display a message using MessageBox.
The table is shown below:

 Problem 4. Biased bees

This example comes from the stats page of the College of Saint Benedict and Saint John’s University (Kirkman, 1996). It is a very interesting example
as it shows a case in which a t-test fails to see a difference between the samples because of the non-normality of of the sample’s distributions. The Kolmogorov-Smirnov nonparametric test, on the other hand, succeeds.

The example deals with the preference of bees between two nearby blooming trees in an empty field. The experimenter has colelcted data measurinbg how much time does a bee spents near a particular tree. The time starts to be measured when a bee first touches the tree, and is stopped when the bee moves more than 1 meter far from it. The samples below represents the measured time, in seconds, of the observed
bees for each of the trees.

Problem 5. Comparing classifier performances

The last example comes from (E. Ientilucci, 2006) and deals with comparing the performance of two different raters (classifiers) to see if their performance are significantly different.

Suppose an experimenter has two classification systems, both trained to classify observations into one of 4 mutually exclusive categories. In order to measure the performance of each classifier, the experimenter confronted their classification labels with the ground truth for a testing dataset, writing the respective results in the form of contingency tables.

The hypothesis to be tested is that the performance of the two classifiers are the same.

In this case, the test didn’t show enough evidence to confidently reject the null hypothesis. Therefore, one should restrain from affirming anything about differences between the two systems, unless the power for the test is known.

Unfortunately I could not find a clear indication in the literature about the power of a two matrix Kappa test. However, since the test statistic is asymptotically normal, one would try checking the power for this test by analysis the power of the underlying Z-test. If there is enough power, one could possibly accept the null hypothesis that there are no large differences between the two systems.


As always, I expect the above discussion and examples could be useful for interested readers and users. However, if you believe you have found a flaw or would like to discuss any portion of this post, please feel free to do so by posting on the comments section.

PS: The classroom example uses a T-test to test for differences in populations means. The T-Test assumes a normal distribution. The data, however, it not exactly normal, since it is crippled between 0 and 10. Suggestions for a better example would also be appreciated!


R. A. Fisher, 1925. Statistical Methods for Research Workers. Available online from:

M. K. Smith, 2011. Common mistakes in using statistics: Spotting and Avoiding Them – Power of a
Statistical Procedure. Available online from:

M. K. Smith, 2011b. Common mistakes in using statistics: Spotting and Avoiding Them – Detrimental
Effects of Underpowered or Overpowered Studies. Available online from:

L. Thomas, F. Juanes, 1996. The importance of statistical power analysis: an example from animal behaviour,
Animal Behaviour, Volume 52, Issue 4, October., Pages 856-859. Available online from:

C. H. (Alex) Yu, 2012. Don’t believe in the null hypothesis? Available online from:

D. Beaulieu-Prévost, 2005. Statistical decision and falsification in science : going beyond the null
hypothesis. Séminaires CIC. Université de Montréal.

T.W. Kirkman, 1996. Statistics to Use. Acessed July 2012. Avalilable online from

E. Ientilucci, 2006. “On Using and Computing the Kappa Statistic”. Available online from

Accord.NET has a new home!


As some might have noticed, Origo is going away until the end of the month. Origo maintainers were kind enough to support previous project owners with a tarball containing the entire contents from the hosted projects upon request; this means all Accord.NET content previously hosted on Origo can be gradually restored at Google Code one the next months.

The Accord.NET Framework is now at

With the hosting change there also comes a few updates. Version 2.7 is about to be released soon; this version now includes the highly requested Bag of Visual Words (BoW) based on the SURF feature point detector. This feature can be used to extract fixed-length feature vectors from images and train standard methods such as Support Vector Machines to perform image classification. Initial experiments with the χ² (chi-square) kernel have shown promising results.

The Accord.NET Framework project page, now hosted at Google Code.

Other hopefully useful features include new learning algorithms for hidden Markov models, such as the Viterbi training and the Maximum Likelihood Estimation training; and the introduction of the ISampleable interface for specifying distributions which can be sampled efficiently (i.e. now it is possible to drawn random samples from some distributions – including hidden Markov models). Other nice addition is the inclusion of the ScatterplotBox class for displaying data points in a scatterplot in a similar way to MessageBox.Show. It comes at hand to quickly inspect some spacial data in the middle of other calculations or even in live demonstrations.

As others might also have noticed, this site’s design also has changed, hopefully for the better.

As usual, I hope someone will find the new release useful 🙂

Quadratic Programming in C#

I have manually translated and adapted the QuadProg solver for quadratic programming problems made by Berwin A. Turlach. His code was originally published under the GNU Library License, which has now been superseded by the GNU Lesser License. This adapted version honors the original work and is thus distributed under the same license.

The code in this article is part of the Accord.NET Framework. Download the full framework through NuGet and use this code in your own applications.


Despite the name, the terms linear or quadratic programming have little resemblance to the set of activities most people now know as programming. Those terms usually usually refers to a specific set of function optimization methods, i.e. methods which can be used to determine the maximum or minimum points of special kinds of functions under a given number of solution constraints. For example, suppose we would like to determine the minimum value of the function:

f(x, y) = 2x + y + 4

Under the constraints that x and y must be non-negative (i.e. either positive or zero). This may seem fairly simple and trivial, but remember that practical linear programming problems may have hundreds or even thousands of variables and possibly million constraints.

When the problem to be solved involves a quadratic function instead of a linear function, but still presents linear constraints, this problem can be cast as a quadratic programming problem. Quadratic functions are polynomial functions in each each term may have at most a total degree of 2. For example, consider the function

f(x, y, z) = 2x² + 5xy + y² – z² + x – 5.

Now let’s check the sum of the degrees for each variable on the polynomial terms. We start by writing the missing terms of the polynomial

f(x, y, z) = 2 x2y0z0 + 5 x1y1z0 + 2 x0y2z0x0y0z2 + x1y0z0 – 5 x0y0z0

and then proceed to check the sum of the degrees at each term. In the first term, 2+0+0 = 2. For the second, 1+1+0 = 2, and so on. Those functions have a nice property that they can be expressed in a matrix form

f(x) = 1/2 xT Q x + cTx.

Here, x and c are vectors. The matrix Q is a symmetric matrix specifying how the variables combine in the quadratic terms of the function. If this matrix is positive definite, then the function is convex, and the optimization has a single, unique optimum (we say it has a global optimum). The coefficients c specify the linear terms of our function.

Source code

The available source code is based on a translation of the Fortran code written by Berwin A. Turlach. However, some modifications have been made. Fortran uses column-major ordering for matrices, meaning that matrices are stored in memory in sequential order of column elements. Almost all other languages use row-major ordering, including C, C++, Java and C#. In order to improve data locality, I have modified the code to use the transpose of the original matrices D and A. I have also modified the QP formulation adopted in the Goldfarb and Idnani paper to reflect the form presented in the introduction.

In this article, we will be using the GoldfarbIdnani class from the Accord.NET Framework, as well as the QuadraticObjectiveFunction. Their source code are available at GoldfarbIdnani.cs and QuadraticObjectiveFunction.cs, respectively. Please note that the sample application version available in this blog post will most likely not be the most recently, updated, fixed and enhanced version of the code. For the latest version, be sure to download the latest version of the framework on the project site or through a NuGet package.

Using the code

The first step in solving a quadratic programming problem is, well, specifying the problem. To specify a quadratic programming problem, one would need two components: a matrix D describing the relationship between the quadratic terms, and a vector d describing the linear terms. Perhaps this would work better with an example.

Suppose we are trying to solve a minimization problem. Given a function, the goal in such problems is to find the correct set of function arguments which would result in the minimum possible value for the function. An example of a quadratic minimization problem is given below:

min f(x, y) = 2x² – xy + 4y² – 5x – 6y

subject to the constraints:

x – y = 5
x = 10

(generated with Wolfram Alpha)

However, note that this problem involves a set of constraints. The required solution for this minimization problem is required to lie in the interval specified by the constraints. More specifically, any x and y pair candidate for being a minimal of the function must respect the relations x – y = 5 and x >= 10. Thus, instead of lying in the unconstrained minimum of the function surface shown above, the solution lies slightly off the center of the surface. This is an obvious easy problem to solve manually, but it will fit for this demonstration.


As it can be seen (and also live demonstrated by asking Wolfram Alpha) the solution lies on the point (10,5), and the constrained minimum of the function is given by 170. So, now that we know what a quadratic programming problem looks like, how can we actually solve it?

Specifying the objective function

The first step in solving a quadratic programming problem is to specify the objective function. Using this code, there are three ways to specify it. Each of them has their own advantages and disadvantages.

1. Manually specifying the QP matrix.

This is the most common approach for numerical software, and probably the most cumbersome for the user. The problem matrix has to be specified manually. This matrix is sometimes denoted Q, D or H as it actually denotes the Hessian matrix for the problem.

The matrix Q is used to describe the quadratic terms of our problem. It is a n x n matrix, in which n corresponds to the number of variables in our problem, covering all possible combinations of variables. Recall our example given on the start of this section. We have 2 variables, x and y. Thus, our matrix Q is 2 x 2. The possible combinations for x and y are expressed in the table below.

x y
x x*x x*y
y y*x y*y

To form our matrix Q, we can take all coefficients associated with each pair mentioned on the table above. The diagonal elements should also be multiplied by two (this is actually because the matrix is the Hessian matrix of the problem: it is the matrix of all second-order derivatives for the function. Since we have only at most quadratic terms, the elementary power rule of derivation “drops” the ² from the x² and y² terms – I think a mathematician would hit me with a stick for explaining it like this, but it serves well for a quick, non-technical explanation).

Remember our quadratic terms were 2x²  – 1xy + 4y². Writing the terms on their proper position and differentiating, we have:

As it can be seen, the matrix is also symmetric (and often, but not always, positive definite). The next step, more trivial, is to write a vector d containing the linear terms. The linear terms are –5x –6y, and thus our vector d can be given by:

Therefore our C# code can be created like this:

2. Using lambda expressions

This approach is a bit more intuitive and less error prone. However, it involves lambdas functions and some people find it hard to follow them. Another disadvantage is that we will lose the edit & continue debugging ability of visual studio. The advantage is that the compiler may catch some obvious syntax errors automatically. Below we are using the QuadraticObjectiveFunction from Accord.NET to specify our target objective function.

Note that the x and y variables could have been initialized to any value. They are only used as symbols, and not used in any computations.

3. Using text strings

This approach is more intuitive but a bit more error prone. The function can be specified using strings, as in a standard mathematical formula. However, since all we have are strings, there is no way to enforce static, compile time checking.

Couldn’t be easier.

Specifying the constraints

The next step in specifying a quadratic programming problem is to specify the constraints. The constraints can be specified in almost the same way as the objective function.

1. Manually specifying the constraints matrix

The first option is to manually specify the constraints matrix A and vector b. The constraint matrix expresses the way the variables should be combined when compared to corresponding value on vector b. It is possible to specify either equality constraints or inequality constraints. The formulation used in this code is slightly different from the one used in Turlach’s original implementation. The constraints are supposed to be in the form:

A1 x = b1
A2 x = b2

This means that each line of matrix A expresses a possible combination of variables x which should be compared to the corresponding line of b. An integer variable m can be specified to denote how many of the first rows of matrix A should be treated as equalities rather than inequalities. Recall that in our example the constraints are given by 1x –1y = 5 and 1x = 10. Lets write this down in a tabular form:

# x y ? b
q1 1 -1 = 5
q2 1 0 = 10

Thus our matrix A and vector b can be specified as:

And not forgetting that m = 1, because the first constraint is actually an equality.

2. Using classes and objects

A more natural way to specify constraints is using the classes and objects of the Accord.NET Framework. The LinearConstraint class allows one to specify a single constraint using an object-oriented approach. It doesn’t have the most intuitive usage on earth, but has much more expressiveness. It can also be read aloud, it that adds anything! 🙂

The specification is centered around the notion that variables are numbered and have an associated index. For example, x is the zero-th variable of the problem. Thus x has an index of 0 and y has an index of 1. So for example, reading aloud the last constraint, it is possible to express how the variables at indices 0 and 1, when combined as 1x and –1y, should be equal to value 5.

2. Using lambda expressions

A more intuitive way to express constraints is again using lambda expressions. And again the problems are the same: some people find it hard to follow and we lose edit & continue.

3. Using text strings

Same as above, but with strings.

Finally, creating and solving the problem

Once we have specified what do we want, we can now ask the code for a solution. In case we have opted for manually specifying the objective function Q and d, and the constraint matrix A, vector b and integer m, we can use:

In case we have opted for creating a QuadraticObjectiveFunction object and a list of constraints instead, we can use:

After the solver object has been created, we can call Minimize() to solve the problem. In case we have opted for manually specifying Q and d, we can use:

The solution will be available in the Solution property of the solver object, and will be given by:

Sample application

The Accord.NET Framework now includes a sample application demonstrating the use of the Goldfarb-Idnani Quadratic Programming Solver. It can be downloaded at the Accord.NET Framework site, and also comes together with recent versions of the framework (> 2.6).


Solver sample application included in the Accord.NET Framework.


Because the code has been translated by hand (in contrast of using automatic translators such as f2c) there could be potential bugs in the code. I have tested the code behavior against R’s quadprog package and still didn’t find errors. But this does not mean the code is bug-free. As always, as is the case of everything else in this blog, this code is published in good faith, but I can not guarantee the correctness of everything. Please read the disclaimer for more information.


Limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) for Unconstrained Optimization in C#


The Limited-memory Broyden-Fletcher-Goldfarb-Shanno method is an optimization method belonging to the family of quasi-Newton methods for unconstrained non-linear optimization. In short terms, it is an off-the-shelf optimizer for seeking either minimum or maximum points of a any differentiable and possibly non-linear function, requiring only an expression of the function and its gradient. The goal of this article is to provide and demonstrate a C# port of the L-BFGS method originally written by Jorge Nocedal in Fortran.


Function optimization is a common problem found in many numerical applications. Suppose we have a differentiable function f : Rn → R and we would like to obtain its minimal or maximum value while traversing its space of possible input values. Those kind of problems arise often in parameter optimization of machine learning models and other statistic related applications (and in a myriad of other applications too – however, we will pay special attention to cases related to machine learning).

The problem of maximizing or minimizing a function can be simply stated as max f(x) or min f(x). When there aren’t any constraints limiting possible values for x, it is widely known from calculus that the maximum or a minimum of such a function would occur when the first derivatives f’(x) of the function f(x) in respect to x are zero. Newton’s method is a common method for finding the roots of a differentiable function and is indeed a suitable method for finding those values such that first derivatives of a function are zero.


Surface of the example function. Contour lines of the example function.

Example of a convex, but non-linear function f(x,y) = exp{-(x-1)²} + exp{-(y-2)²/2}. Images have been created using Wolfram Alpha.


However, while Newton’s method may work very well with functions of a single variable, the generalization to higher dimensions will require the replacement of the first derivative f’(x) with the function’s gradient vector g of partial derivatives, and the second derivative f’’(x) with the inverse Hessian matrix H-1. The problem is that the Hessian is a square matrix with as many rows and columns as parameters of our function f, and, besides being very costly and challenging to compute, even its size may be infeasible to accommodate in main memory when dealing with very high dimensional problems.

To overcome those limitations, several methods attempt to replace the Hessian matrix with an approximation extracted from the first partial derivatives alone. Those methods are called quasi-Newton methods, as they replace H in Newton’s method by an approximation instead of using the full Hessian. In case the function being optimized has any particular form or special characteristic which could be exploited, it is also possible to use more specialized methods such as the Gauss-Newton, Levenberg-Marquardt or even lower-bound approximation methods to avoid computing the full Hessian.

The quasi-Newton BFGS method for non-linear optimization builds an approximation for the Hessian matrix based on estimates extracted solely from first order information. However, even being cheaper to compute, the memory requirements are still high as the method still needs to accommodate a dense N x N matrix for the inverse Hessian approximation (where N is the number of variables for the function). To overcome this problem, the L-BFGS method, or the limited-memory version of the BFGS method, never forms or stores the approximation matrix, but only a few vectors which can be used to represent it implicitly.

Source code

The code presented here is based upon a direct translation of the original code by Jorge Nocedal. His original code was released under the permissive BSD license, and honoring the original license and the author’s considerations, this code is released under the BSD as well.

L-BFGS Class diagrams

The L-BFGS method is implemented within the BroydenFletcherGoldfarbShanno class. Upon object creation, it is possible to specify a function and its gradient through either the use of Lambda expressions or by specifying handles for named functions. After the object has been initialized, a simple call to Minimize runs the standard L-BFGS method until convergence. The details for the Minimize method is given below.

The version of the code detailed here only supports Minimization problems. However, it is possible to note that any minimization problem can be converted into a maximization problem and vice-versa by taking the opposite of the function and its gradient.

Using the code

Code usage is rather simple. Suppose we would like to maximize the function g(x,y) = exp{-(x-1)²} + exp{-(y-2)²/2}:

Since we would like to perform a maximization, we first have to convert it to a minimization problem. The minimization version of the function is simply given by taking f(x,y) = –g(x,y):

Which is a convex function, as can be seen by plotting its surface. As it can be seen, the minimum of the function lies on the point (x,y) = (1,2). As expected, this point coincides with the roots of the partial derivative functions, as shown in the line plots below:


Example optimization cast as an minimization problem.

The minimization problem min f(x,y) = -(exp(-(x-1)²) + exp(-(y-2)²/2)), computed by Wolfram Alpha.

Partial derivative for x. Partial derivative for y.

The roots of the partial derivatives in respective to x (left) and y (right). It is possible to see that the roots occur at 1 and 2, respectively. Images computed using Wolfram Alpha. 


The first step in solving this optimization problem automatically is first to specify the target function. The target function f, in this case, can be specified through a lambda function in the form:

The next step is to specify the gradient g for the function f. In this particular case, we can manually compute the partial derivatives to be df/dx = -2 e^(-(x-1)^2) (x-1) and df/dy = -e^(-1/2 (y-2)^2) (y-2), in respect to x and y, respectively. Writing a lambda function to compute the gradient vector g, we have:

We can also note that this example was rather simple, so the gradient vector was easy to calculate. However, the gradient could also have been computed automatically using Accord.NET‘s FiniteDifferences class. In either case, all we have to do now is to create our L-BFGS solver and call its Minimize() method to begin optimization:

After the optimization is complete, the solution parameter will be available in the Solution property of the lbfgs object, and will be equal to { 1.0, 2.0 }. And also in case one is interested in progress reports (such as in the case of optimizing very large functions), it is also possible to register a listener to the Progress event handler. The complete version of the sample application accompanying the source code is given below:


In this post, we showed how to use a reduced set of the Accord.NET Framework‘s to perform non-linear optimization. This routine is also used by the Conditional Random Fields and Hidden Conditional Random Fields trainers to optimize parameters for such models. The solver routines have been adapted from the original Fortran’s source code from Nocedal, which, tracing back to a 2001 message from the author, also have been reported to be available under the public domain. Nevertheless, the reduced set of the framework available for download within this post is available under a BSD license, as an alternative to the version available within the Accord.NET Framework which is available only under a LGPL license.

As always, I hope someone will find it useful 🙂


Decision Trees in C#


Decision trees are simple predictive models which map input attributes to a target value using simple conditional rules. Trees are commonly used in problems whose solutions must be readily understandable or explainable by humans, such as in computer-aided diagnostics and credit analysis.


Decision Trees give a direct and intuitive way for obtaining the classification of a new instance from a set of simple rules. Because of this characteristic, decision trees find wide use in situations in which the interpretation of the obtained results and of the reasoning process is crucial, such as in computer-aided diagnostics (CAD) and in financial credit analysis. Consumer credit analysis is an interesting example because, in many countries, one can not simply reject credit without giving a justification, justification of which is trivial to extract from a decision tree.

The tree on the right has been generated using the Context Free software based on the grammar shared by davdrn.

Learning decision trees

Decision trees can be simply drawn by hand based on any prior knowledge the author may have. However, their real power becomes apparent when trees are learned automatically, through some learning algorithm.

The most notable and classics examples to decision tree learning are the algorithms ID3 (Quinlan, 1986) and the C4.5 (Quinlan, 1993). Both are examples of greedy algorithms, performing local optimum decisions in the hope of producing a most general tree. Such algorithms are based on the principle of the Occam’s razor, favoring simpler or smaller trees in the hope that such smaller trees could retain more generalization power. This preference is formalized through the specification of an hypothesis ordering criteria such as the information gain. The information gain measures the, as the name implies, gain of information in using each of the attributes as a next factor to consider during decision. The information gain can be defined as:



However, the information gain has a bias towards attributes with a high number of possible values (Mitchell, 1997). A way to overcome this bias is to select new selection attributes based on alternative criteria, such as the gain ratio (Quinlan, 1986), defined as:



In the GainRatio, the SplitInformation term attenuates the importance given to attributes with many, uniformly distributed, possible values.

Iterative Dichotomizer 3 (ID3)

The algorithm presented below is a slightly different version of the original ID3 algorithm as presented by Quinlan. The modifications are to support multiple output labels. In each recursion of the algorithm, the attribute which bests classifiers the set of instances (or examples, or input-output pairs, or data) is selected according to some selection criteria, such as the InfoGain or the GainRatio.

  • ID3(instances, target_attribute, attributes)
    • Create a new root node to the tree.
    • If all instances have the target_attribute belonging to the same class c,
      • Return the tree with single root node with label c.
    • If attributes is empty, then
      • Return the tree with single root node with the most common label of the target_attribute in instances.
    • Else
      • A ← the attribute in attributes which best classifies instances
      • root decision attribute ← A
      • Foreach possible value vi of A,
        • Add a new ramification below root, corresponding to the test A = vi
        • Let instancesvi be the subset of instances with the value vi for A
        • If instancesvi is empty then
          • Below this ramification, add a new leaf node with the most common value of target_attribute in instances.
        • Else below this ramification, add the subtree given by the recursion:
          ID3(instancesvi, target_attribute, attributes – { A })
    • End
  • Return root

Difficulties and disadvantages of decision tree learning

Despite relying on the Occam’s Razor to guide the learning, neither ID3 or C4.5 algorithms are not guaranteed to produce the smaller or more general tree possible. This happens because their learning is based solely on heuristics and not in truly optimality criteria. The following example, from (Esmeir & Markovitch, 2007) illustrates the learning of the concept xor (exclusive or) by the ID3 algorithm. In this example, A3 and A4 are irrelevant attributes, having absolutely no influence on the target answer. However, yet being irrelevant, ID3 will select one of them to belong to the tree. In fact, ID3 will select the irrelevant attribute A4 to be the root of the learned tree.

A1 A2 A3 A4 label
1 0 0 1 +
0 1 0 0 +
0 0 0 0
1 1 0 0
0 1 1 1 +
0 0 1 1
1 0 1 1 +

One complication of decision tree learning is that the problem to find the smaller consistent tree (in the sense of classifying correctly all training samples) is known to be NP-complete (Hyafil & Rivest, 1976). Moreover, the separating decision surface generated by such trees are always formed by parallel cuts alongside the attribute space axis, which could be a severely suboptimal solution (Bishop, 2007, p. 666). The example given by Bishop illustrates well this problem: for example, to separate classes whose optimum decision margin are on 45 degrees from one of the axis, it will be needed a high number of parallel cuts in comparison with a single cut on the diagonal such as could be given by any linear decision machine. Another disadvantage of traditional decision tree learning algorithms is that most methods require only a constant learning time, and, as such, do not allow for trading extra training time for a better solutions. The work of (Esmeir & Markovitch, 2007) is dedicated to overcome such problem.

The following picture shows an example on how learning by decision trees is limited to cuts parallel to the axis, as described by Bishop. The Ying-Yang classification problem is a classical example of a non-linearly separable decision problem. Decision trees, albeit not being linear classifiers, have difficulty classifying this set with simple thresholds.


Top-Left: Ying-Yang non-linear classification problem. Top-Right: Decision surface extracted by a decision tree. Bottom: Decision threshold rules extracted from the tree.

However, despite all of those shortcomings, decision trees plays major roles as base of many successful algorithms. One interesting application and of notorious success is given in the field of object detection in images. The first real-time face detecting algorithm (Viola & Jones, 2001) is based on a degenerated decision tree (also known as a cascade). The body recognition and segmentation algorithm used by the Xbox 360 gaming platform used to process depth images generated by its companion sensor Kinect is equally based on the use of decision trees (Shotton, et al., 2011). As well is the case of the FAST algorithm for interest point detection in images (Rosten & Drummond, 2006).

I should also make the note that both the Viola-Jones and the FAST algorithms are available in the Accord.NET Framework for immediate use (the Viola-Jones (HaarObjectDetector) has also been recently updated to support multiple threads, so feel free to take a look and experiment with it!).

In sum, its possible to say that great part of the applicability of decision trees came from the simple fact that they are extremely fast to evaluate. One of the reasons for this feature is being easily translated to sets of conditional instructions in a conventional computer program. The decision trees now available in the Accord.NET Framework make full use of this fact and enables the user to actually compile the decision trees to native code on-the-fly, augmenting even more its performance during classification.

Source code

The code presented in this section is actually part of the Accord.NET Framework. The Accord.NET Framework is a framework extension to the already very popular AForge.NET Framework, adding and providing new algorithms and techniques for machine learning, computer vision and even more.

The Accord.MachineLearning.DecisionTree namespace is comprised of the following classes:

  • DecisionTree, the main class representing a decision tree, with methods such as Compute to compute the tree classification given an input vector;
  • DecisionNode, the node class for the decision tree, which may or may not have children nodes contained under a collection of children represented by a DecisionBranchNodeCollection;
  • DecisionVariable, a class to specify the nature of each variable processable by the tree, such as if the variable is continuous, discrete, which are their expected or valid ranges;
  • DecisionBranchNodeCollection, a class to contain children nodes together with information about which attribute of the data should be compared with the child nodes during reasoning.


Whose relationships can be seen in the following class diagram:

class diagram

The learning algorithms available are either the ID3 algorithm discussed above, or its improved version C4.5 (which can handle continuous variables, but at this time does not yet support pruning), both proposed and published by Ross Quinlan.


Despite the bit complicated look, usage is rather simple as it will be shown in the next section.

Using the code

Consider, for example, the famous Play Tennis example by Tom Mitchell (1998):

In the aforementioned example, one would like to infer if a person would play tennis or not based solely on four input variables. Those variables are all categorical, meaning that there is no order between the possible values for the variable (i.e. there is no order relationship between Sunny and Rain, one is not bigger nor smaller than the other, but are just distinct). Moreover, the rows, or instances presented above represent days on which the behavior of the person has been registered and annotated, pretty much building our set of observation instances for learning.

In order to try to learn a decision tree, we will first convert this problem to a more simpler representation. Since all variables are categories, it does not matter if they are represented as strings, or numbers, since both are just symbols for the event they represent. Since numbers are more easily representable than text string, we will convert the problem to use a discrete alphabet through the use of a codebook.

A codebook effectively transforms any distinct possible value for a variable into an integer symbol. For example, “Sunny” could as well be represented by the integer label 0, “Overcast” by “1”, Rain by “2”, and the same goes by for the other variables. So:

Now we should specify our decision tree. We will be trying to build a tree to predict the last column, entitled “PlayTennis”. For this, we will be using the “Outlook”, “Temperature”, “Humidity” and “Wind” as predictors (variables which will we will use for our decision). Since those are categorical, we must specify, at the moment of creation of our tree, the characteristics of each of those variables. So:

Let’s now proceed and create our DecisionTree:

Now we have created our decision tree. Unfortunately, it is not really very useful, since we haven’t taught it the problem we are trying to predict. So now we must instantiate a learning algorithm to make it useful. For this task, in which we have only categorical variables, the simplest choice is to use the ID3 algorithm by Quinlan. Let’s do it:

At this point, the tree has been created. In order to ask the tree to classify new samples, we can use:

Please note that, in case any of the steps in this post doesn’t work out, it might be because the most up-to-date version in the Accord.NET Framework may have undergone some changes. In this case, please refer to the official documentation at the Accord.NET website.

Going further, a suitable representation of the learned tree could be given by the following diagram:


However, more than just a diagram, we can also go and generate a .NET Expression Tree describing our decision tree. Expression trees represent code in the form of a tree of expressions, which can then be read, modified or compiled. By compiling the DecisionTree‘s expression, we are able to generate code on-the-fly and let the JIT compile it down to native code at runtime, greatly improving the performance of our decision tree:

And here is the resulting C# code obtained by compiling the expression into a lambda function, dumping the function into an dynamic assembly, opening and inspecting this assembly using ILSpy:


Decision trees are useful tools when the problem to be solved needs to be quickly interpreted and understood by humans. Another suitable use is when the decision needs to be fast. However, decision trees, at least those trained by simple training algorithms such as ID3 and C4.5 can perform quite poorly depending on the problem. As it happens with all machine learning techniques, it is futile to believe there is a one true classifier which would act perfectly on all possible imaginable scenarios. As always, it is important to know our tools and know in which situation each technique would work better.


  • Bishop, C. M., 2007. Pattern Recognition and Machine Learning (Information Science and Statistics). 1st ed. 2006. Corr. 2nd printing ed. s.l.:Springer
  • Fayyad, U. M. & Irani, K. B., 1992. On the Handling of Continuous-Valued Attributes in Decision Tree Generation. Machine Learning, January, 8(1), pp. 87-102.
  • Quinlan, J. R., 1986. Induction of decision trees. Machine Learning, 1(1), pp. 81-106.
  • Quinlan, J. R., 1993. C4.5: Programs for Machine Learning (Morgan Kaufmann Series in Machine Learning). 1 ed. s.l.:Morgan Kaufmann.
  • Shotton, J. et al., 2011. Real-Time Human Pose Recognition in Parts from Single Depth Images. s.l., s.n.
  • Viola, P. & Jones, M., 2001. Robust Real-time Object Detection. s.l., s.n.
  • Mitchell, T. M., 1997. Decision Tree Learning. In:: Machine Learning (McGraw-Hill Series in Computer Science). s.l.:McGraw Hill.
  • Mitchell, T. M., 1997. Machine Learning (McGraw-Hill Series in Computer Science). Boston(MA): WCB/McGraw-Hill.
  • Esmeir, S. & Markovitch, S., 2007. Anytime Learning of Decision Trees. J. Mach. Learn. Res., May, Volume 8, pp. 891-933.
  • Hyafil, L. & Rivest, R. L., 1976. Constructing Optimal Binary Decision Trees is NP-complete. Information Processing Letters, 5(1), pp. 15-17.

Accord.NET Framework: Gesture Controller Components


A new version (2.2.0) of the Accord.NET Framework has just been released. This new version introduces many new features, fixes and improvements. The most interesting additions are certainly the HeadController and FaceController .NET components.


Accord.NET Framework sample application for Gesture Controller Components

The Accord.NET Controller components can be used to generate events based on webcam motion. By using a combination of HaarCascadeClassifiers, Camshift and Template-based Tracking, those components are able to detect when a face enters scene, leaves the scene, and moves across a scene.

The video above shows only the sample application which comes together with the framework. However, the interesting part is that this is just a sample of what can be accomplished using the real controller components. The controller components are .NET components, similar to Button, Label or Timer, and can be dragged and dropped from Visual Studio’s ToolBox directly into any application.

Accord.NET Controller Components in Visual Studio

Once inside an application, it will be possible to set event actions just as in any other .NET component:

Setting events for head movements using Accord.NET Controller Components

The controls have built-in support for calibration. All values except tilting angle are passed to the hosting application in the [-1;+1] range, in which -1 indicates either a total left/down/backwards position and +1 indicates a total right/up/forward position. The tilting angle is given in radians. Please note that the face controller is still a bit experimental and still requires some tuning.


This new version also introduces HSL Color Range object trackers, more default Haar Cascades, an experimental version of linear-chain Conditional Random Fields, and the ability to generate hardcoded C# definitions of any Haar cascade available in the OpenCV XML format. There is also initial support for finger detection using new implementations for Border-Following contour extraction, K-Curvatures and Convex Hull Defects extraction. On the statistics side, there has been the inclusion of the Von-Mises distribution, Moving and Running Normal distributions and improvements in the Multivariate Gaussian implementation. The full release notes are available in the release’s download page.

Also, a special thanks to Professor Dr. Modesto Castrillón Santana to let me embed some of his Haar definitions into the framework under the LGPL license. Please be sure to include a reference to his work if you plan to use this in an academic publication.


As always, I hope those additions and improvements will be useful to everyone 🙂