Tag Archives: gradient descent

A simple Python program for an ANN to cover the MNIST dataset – VII – EBP related topics and obstacles

Posted on by

I continue with my series about a Python program to build simple MLPs:

A simple Python program for an ANN to cover the MNIST dataset – VI – the math behind the „error back-propagation“
A simple program for an ANN to cover the Mnist dataset – V – coding the loss function
A simple program for an ANN to cover the Mnist dataset – IV – the concept of a cost or loss function
A simple program for an ANN to cover the Mnist dataset – III – forward propagation
A simple program for an ANN to cover the Mnist dataset – II – initial random weight values
A simple program for an ANN to cover the Mnist dataset – I – a starting point

On our tour we have already learned a lot about multiple aspects of MLP usage. I name forward propagation, matrix operations, loss or cost functions. In the last article of this series
A simple program for an ANN to cover the Mnist dataset – VI – the math behind the „error back-propagation“
I tried to explain some of the math which governs “Error Back Propagation” [EBP]. See the PDF attached to the last article.

EBP is an algorithm which applies the “Gradient Descent” method for the optimization of the weights of a Multilayer Perceptron [MLP]. “Gradient Descent” itself is a method where we step-wise follow short tracks perpendicular to contour lines of a hyperplane in a multidimensional parameter space to hopefully approach a global minimum. A step means a change of of parameter values – in our context of weights. In our case the hyperplane is the surface formed by the cost function over the weights. If we have m weights we get a hyperplane in an (m+1) dimensional space.

To apply gradient descent we have to calculate partial derivatives of the cost function with respect to the weights. We have discussed this in detail in the last article. If you read the PDF you certainly have noted: Most of the time we shall execute matrix operations to provide the components of the weight gradient. Of course, we must guarantee that the matrices’ dimensions fit each other such that the required operations – as an element-wise multiplication and the numpy.dot(X,Y)-operation – become executable.

Unfortunately, there are some challenges regarding this point which we have not covered, yet. One objective of this article is to get prepared for these potential problems before we start coding EBP.

Another point worth discussing is: Is there really just one cost function when we use mini-batches in combination with gradient descent? Regarding the descriptions and the formulas in the PDF of the last article this was and is not fully clear. We only built sums there over cost contributions of all the records in a mini-batch. We did NOT use a loss function which assigned to costs to deviations of the predicted result (after forward propagation) from known values for all training data records.

This triggers the question what our
code in the end really does if and when it works with mini-batches during weight optimization … We start with this point.

In the following I try to keep the writing close to the quantity notations in the PDF. Sorry for a bad display of the δs in HTML.

Gradient descent and mini-batches – one or multiple cost functions?

Regarding the formulas given so far, we obviously handle costs and gradient descent batch-wise. I.e. each mini-batch has its own cost function – with fewer contributions than a cost function for all records would have. Each cost function has (hopefully) a defined position of a global minimum in the weights’ parameter space. Taking this into consideration the whole mini-batch approach is obviously based on some conceptually important assumptions:

  • The basic idea is that the positions of the global minima of all the cost-functions for the different batches do not deviate too much from each other in the basic parameter space.
  • If we additionally defined a cost function for all test data records (over all batches) then this cost function should display a global minimum positioned in between the ones of the batches’ cost functions.
  • This also means that there should be enough records in each batch with a really statistical distribution and no specialties associated with them.
  • Contour lines and gradients on the hyperplanes defined by the loss functions will differ from each other. On average over all mini-batches this should not hinder convergence into a common optimum.

To understand the last point let us assume that we have a batch for MNIST dataset where all records of handwritten digits show a tendency to be shifted to the left border of the basic 28×28 pixel frames. Then this batch would probably give us other weights than other batches.

To get a deeper understanding, let us take only two batches. By chance their cost functions may deviate a bit. In the plots below I have just simulated this by two assumed “cost” functions – each forming a hyperplane in 3 dimensions over only two parameter (=weight) dimensions x and y. You see that the “global” minima of the blue and the red curve deviate a bit in their position.

The next graph shows the sum, i.e. the full “cost function”, in green in comparison to the (vertically shifted and scaled) original functions.

Also here you clearly see the differences in the minimas’ positions. What does this mean for gradient descent?

Firstly, the contour lines on the total cost function would deviate from the ones on the cost function hyperplanes of our 2 batches. So would the directions of the different gradients at the point presently reached in the parameter space during optimization! Working with batches therefore means jumping around on the surface of the total cost function a bit erratically and not precisely along the direction of steepest descent there. By the way: This behavior can be quite helpful to overcome local minima.

Secondly, in our simplified example we would in the end not converge completely, but jump or circle around the minimum of the total cost function. Reason: Each batch forces the weight corrections for x,y into different directions namely those of
its own minimum. So, a weight correction induced by one bath would be countered by corrections imposed by the optimization for the other batch. (Regarding MNIST it would e.g. be interesting to run a batch with handwritten digits of Europeans against a batch with digits written by Americans and see how the weights differ after gradient descent has converged for each batch.)

This makes us understand multiple things:

  • Mini-batches should be built with a statistical distribution of records and their composition should be changed statistically from epoch to epoch.
  • We need a criterion to stop iterating over too many epochs senselessly.
  • We should investigate whether the number and thus the size of mini-batches influences the results of EBP.
  • At the end of an optimization run we could invest in some more iterations not for the batches, but for the full cost function of all training records and see if we can get a little deeper into the minimum of this total cost function.
  • We should analyze our batches – if we keep them up and do not create them statistically anew at the beginning of each epoch – for special data records whose properties are off of the normal – and maybe eliminate those data records.

Repetition: Why Back-propagation of 2 dimensional matrices and not vectors?

The step wise matrix operations of EBP are to be performed according to a scheme with the following structure:

  • On a given layer N apply a layer specific matrix “NW.T” (depending on the weights there) by some operational rule on some matrix “(N+1)δS“, which contains some data already calculated for layer (N+1).
  • Take the results and modify it properly by multiplying it element-wise with some other matrix ND (containing derivative expressions for the activation function) until you get a new NδS.
  • Get partial derivatives of the cost function with respect to the weights on layer (N-1) by a further matrix operation of NδS on a matrix with output values (N-1)A.TS on layer (N-1).
  • Proceed to the next layer in backward direction.

The input into this process is a matrix of error-dependent quantities, which are defined at the output layer. These values are then back-propagated in parallel to the inner layers of our MLP.

Now, why do we propagate data matrices and not just data vectors? Why are we allowed to combine so many different multiplications and summations described in the last article when we deal with partial derivatives with respect to variables deep inside the network?

The answer to the first question is numerical efficiency. We operate on all data records of a mini-batch in parallel; see the PDF. The answer to the second question is 2-fold:

  • We are allowed to perform so many independent operations because of the linear structure of our cost-functions with respect to contributions coming from the records of a mini-batch and the fact that we just apply linear operations between layers during forward propagation. All contributions – however non-linear each may be in itself – are just summed up. And propagation itself between layers is defined to be linear.
  • The only non-linearity occurring – namely in the form of non-linear activation functions – is to be applied just on layers. And there it works only node-wise! We do not
    couple values for nodes on one and the same layer.

In this sense MLPs are very simple by definition – although they may look complex! (By the way and if you wonder why MLPs are nevertheless so powerful: One reason has to do with the “Universal Approximation Theorem”; see the literature hint at the end.)

Consequence of the simplicity: We can deal with δ-values (see the PDF) for both all nodes of a layer and all records of a mini-batch in parallel.

Results derived in the last article would change dramatically if we had rules that coupled the Z- or A-values of different nodes! E.g. if the squared value at node 7 in layer X must always be the sum of squared values at nodes 5 an 6. Believe me: There are real networks in this world where such a type of node coupling occurs – not only in physics.

Note: As we have explained in the PDF, the nodes of a layer define one dimension of the NδS“-matrices,
the number of mini-batch records the other. The latter remains constant. So, during the process the δ-matrices change only one of their 2 dimensions.

Some possible pitfalls to tackle before EBP-coding

Now, my friends, we can happily start coding … Nope, there are actually some minor pitfalls, which we have to explain first.

Special cost-, activation- and output-functions

I refer to the PDF mentioned above and its formulas. The example explained there referred to the “Log Loss” function, which we took as an example cost function. In this case the outδS and the 3δS-terms at the nodes of the outermost layer turned out to be quite simple. See formula (21), (22), (26) and (27) in the PDF.

However, there may be other cost functions for which the derivative with respect to the output vector “a” at the outermost nodes is more complicated.

In addition we may have other output or activation functions than the sigmoid function discussed in the PDF’s example. Further, the output function may differ from the activation function at inner layers. Thus, we find that the partial derivatives of these functions with respect to their variables “z” must be calculated explicitly and as needed for each layer during back propagation; i.e., we have to provide separate and specific functions for the provision of the required derivatives.

At the outermost layer we apply the general formulas (84) to (88) with matrix ED containing derivatives of the output-function Eφ(z) with respect to the input z to find EδS with E marking the outermost layer. Afterwards, however, we apply formula (92) – but this time with D-elements referring to derivatives of the standard activation-function φ used at nodes of inner layers.

The special case of the Log Loss function and other loss functions with critical denominators in their derivative

Formula (21) shows something interesting for the quantity outδS, which is a starting point for backward propagation: a denominator depending on critical factors, which directly involve output “a” at the outer nodes or “a” in a difference term. But in our one-hot-approach “a” may become zero or come close to it – during training by accident or by convergence! This is a dangerous thing; numerically we absolutely want to avoid any division by zero or by small numbers close to the numerical accuracy of a programming language.

What mathematically saves us in the special case of Log Loss are formulas (26) and (27), where due to some “magic” the dangerous denominator is cancelled by a corresponding factor in the numerator when we evaluate EδS.

In the general case, however,
we must investigate what numerical dangers the functional form of the derivative of the loss function may bring with it. In the end there are two things we should do:

  • Build a function to directly calculate EδS and put as much mathematical knowledge about the involved functions and operations into it as possible, before employing an explicit calculation of values of the cost function’s derivative.
  • Check the involved matrices, whose elements may appear in denominators, for elements which are either zero or close to it in the sense of the achievable accuracy.

For our program this means: Whether we calculate the derivative of a cost function to get values for “outδS” will depend on the mathematical nature of the cost function. In case of Log Loss we shall avoid it. In case of MSE we shall perform the numerical operation.

Handling of bias nodes

A further complication of our aspired coding has its origin in the existence of bias nodes on every inner layer of the MLP. A bias node of a layer adds an additional degree of freedom whilst adjusting the layer’s weights; a bias node has no input, it produces only a constant output – but is connected with weights to all normal nodes of the next layer.

Some readers who are not so familiar with “artificial neural networks” may ask: Why do we need bias nodes at all?

Well, think about a simple matrix operation on a 2 dim-vector; it changes its direction and length. But if we want to approximate a function for regression or a separation hyperplanes for classification by a linear operation then we need another element which corresponds to a constant translation part in a linear transformation: z = w1*x1 + w2*x2 + const.. Take a simple function y=w*x + c. The “c” controls where the line crosses the y axis. We need such a parameter if our line should separate clusters of points separably distributed somewhere in the (x,y)-plane; the w is not sufficient to orientate and position the hyperplane in the (x,y)-plane.

This very basically what bias neurons are good for regarding the basically linear operation between two MLP-layers. They add a constant to an otherwise linear transformation.

Do we need a bias node on all layers? Definitely on the input layer. However, on the hidden layers a trained network could by learning evolve weights in such a way that a bias neuron comes about – with almost zero weights on one side. At least in principle; however, we make it easier for the MLP to converge by providing explicit “bias” neurons.

What did we do to account for bias nodes in our Python code so far? We extended the matrices describing the output arrays ay_A_out of the activation function (for input ay_Z_in) on the input and all hidden layers by elements of an additional row. This was done by the method “add_bias_neuron_to_layer()” – see the codes given in article III.

The important point is that our weight matrices already got a corresponding dimension when we built them; i.e. we defined weights for the bias nodes, too. Of course, during optimization we must calculate partial derivatives of the cost function with respect to these weights.

The problem is:

We need to back propagate a delta-matrix Nδ for layer N via
( (NW.T).dot(Nδ) ). But then we can not apply a simple element-wise matrix multiplication with the (N-1)D(z)-matrix at layer N-1. Reason: The dimensions do not fit, if we calculate the elements of D only for the existing Z-Values at layer N-1.

There are two solutions for coding:

  • We can add a row artificially and intermediately to the Z-matrix to calculate the D-matrix, then calculate NδS as
    ( (NW.T).dot(Nδ) ) * (N_1)D
    and eliminate the first artificial row appearing in NδS afterwards.
  • The other option is to reduce the weight-matrix (NW) by a row intermediately and restore it again afterwards.

What we do is a matter of efficiency; in our coding we shall follow the first way and test the difference to the second way afterwards.

Check the matrix dimensions

As all steps to back-propagate and to circumvent the pitfalls require a bit of matrix wizardry we should at least check at every step during EBP backward-propagation that the dimensions of the involved matrices fit each other.

Outlook

Guys, after having explained some of the matrix math in the previous article of this series and the problems we have to tackle whilst programming the EBP-algorithm we are eventually well prepared to add EBP-methods to our Python class for MLP simulation. We are going to to this in the next article:
A simple program for an ANN to cover the Mnist dataset – VIII – coding Error Backward Propagation

Literature

“Machine Learning – An Applied Mathematics Introduction”, Paul Wilmott, 2019, Panda Ohana Publishing

A simple Python program for an ANN to cover the MNIST dataset – IV – the concept of a cost or loss function

Posted on by

We continue with our effort to write a Python class for a Multilayer Perceptron [MLP] – a simple form of an artificial neural network [ANN]. In the previous articles of this series

A simple program for an ANN to cover the Mnist dataset – III – forward propagation
A simple program for an ANN to cover the Mnist dataset – II – initial random weight values
A simple program for an ANN to cover the Mnist dataset – I – a starting point

I have already explained

  • what parts of an MLP setup we need to parameterize; e.g. the number of layers, the number of nodes per layer, the activation and output functions;
  • how we create node layers and the corresponding weight arrays,
  • how (and also a bit of why) we work with “mini-batches” of test data during training,
  • how we can realize a “vectorized” form of the required “Feed Forward Propagation” algorithm [FFP]. A vectorized form enables us to process all training data records of a mini-batch in parallel. We used Linear Algebra functions provided by Numpy for this purpose; these functions are supported by the the OpenBlas library on a Linux system.

We also set up a basic loop over a number of epochs during training. (Remember: An epoch corresponds to a training step over all training data records). The number of epochs is handled as a parameter to the class’s interface. By artificially repeating the FFP algorithm up to a thousand times, we already got an impression of the code’s performance and its dependence on the number of CPU cores and the size of a mini-batch.

A special method of our class MyANN controls the handling of a mini-batch of multiple input data records via two major steps so far:

  • Step 1: Extract the data records for the mini-batch from the input data.
  • Step 2: Apply FW-propagation to all data records of the mini-batch.

The next natural step would be to encode a training algorithm which optimizes the weight parameters of our MLP. However, in this article we shall not code anything. Instead, I shall discuss some aspects of the so called “cost function” of a MLP. I think this to be useful to get a basic understanding of what training of an ANN actually means and what the differences are in comparison to other ML-algorithms as e.g. the SVM approach. Understanding the cost function’s role for the training of a MLP will also help to better understand the origin and the mathematical form of the back-propagation-algorithm used for training and discussed in a later article.

I simplify a lot below; more details can be found in the literature on machine Learning [ML]; see the section “Links” for some references. Note that if you know all about the theoretical concepts behind ANN training you will not learn anything new here. This is for beginners (and for later reference in this article series).

The concept of a cost function: Turning a classification problem into an optimization problem

What do we mean by training an ANN? Training means to optimize the weights of the ANN such that the “Feed Forward Propagation” in the end delivers correct predictions for new datasets. A cost function is a central concept of the so called “gradient descent method” used for this optimization. By the way: A synonym for cost function is “loss function”. We use both terms
alike below.

The relation between ANN-training based on a loss function and the classification task, which we want to solve with an ANN, is a subtle one. Let us first discuss what we understand by “classification”:

Classification means to separate the input data into categories; i.e.: finding categorical separation surfaces in the multidimensional vector space of input data. In case of the MNIST dataset such separation interfaces should discriminate between 10 different clusters of data points.

I have discussed the problem of finding a separation surface for the case of the moons dataset example in previous articles in this blog. We then used SVM-algorithms to solve this particular problem. Actually, we determined parameters of (non-linear) polynomials to define a separation surface with a (soft) maximum distance from category related clusters of data points in an extended feature space (=input vector space). The extended feature space covered not only basic features of the input data but also powers of it.

All in all we worked directly in an multidimensional extension of the input vector space and optimized parameters describing linear separation interfaces there. If we had several categories instead of 2 we could use a so called “one versus all”-strategy to calculate 10 linear separation interfaces and determine the distance of any new data point towards the separation surfaces as a confidence measure (score) for a prediction. The separation with the highest score would be used to discriminate between the 10 possible solutions and choose the optimal one. Yes, working in an extended input vector space and with parameters of multiple linear separation surfaces was a bit difficult.

Actually, working with ANNs and cost functions corresponds to a more elegant way of optimizing; it starts with measuring distances in the output vector space of the ANN/MLP:

In the context of classification tasks (with known results for training data) a loss function provides a fictitious cost value which weighs the deviations (or distances) of calculated result values (of the ML-algorithm under training) from the already known correct result for training records. I.e. it measures the errors for the training data records in the output space. The optimization task then means to minimize the cost function and thereby minimize a kind of mean error for all input data records.
The hope is that the collection of resulting weight values allows for predictions of other unknown input data, too.

The result of an ANN/MLP for a training data record is the outcome of a complex transformation performed by the ANN. In case of an MLP the transformation of input into output data is done by the “Feed Forward Propagation” algorithm [FFP]. Thus a reasonably designed cost function becomes dependent on the parameters of the FFP-algorithm – predominantly on the weights given at the nodes of the MLP’s layers. We concentrate on this type of parameter below; but note that in special ANN cases there may be additional other parameters to be varied for training and ANN optimization.

The MLP’s weights can in principle be varied continuously during training. The parameter (vector) space thus can be described by multiple real value axes – one for each of the weights. The parameter space of a MLP is a multi-dimensional one with a dimension equal to or bigger than the space of input data – and of course also the result space. (That the dimension is bigger follows from the required node number in the input layer.)

With the help of a suitable cost function we can pose a mathematical optimization problem for the weight parameters:

Find a point in the weight vector space for which the cost
function gives us a minimum, which in turn corresponds to an overall minimum of the deviation distances.

A simple example for a cost function would be a sum of square values for the length of the difference vectors in the output space for all training data.

There are several things to mention:

  1. The result space is a multidimensional vector space (in case of MNIST a 10 dimensional one); so the distance between points there has to be defined via a mathematical norm over components.
  2. The result space in classification problems typically has a much smaller dimension “m” than the dimension “n” of the space of the input data (m < n).
  3. It makes almost no sense to display the cost function over the multidimensional space of input data – as a working ML-algorithm should deliver small cost values for all input data. However, it makes a lot of sense to display the costs over the multi-dimensional vector space of continuous weight values.
  4. We deal with batches of many training data records; it follows that a reasonable cost function in this case must combine deviations of individual records from optimal values. This is very often done via some kind of sum over individual cost contributions from each training record.

A continuous differentiable cost function defines a hyperplane for gradient-descent

In many MLP cases the cost function will be a function of the weight parameters only; this requires a reasonable node independent form of the activation functions. A loss function with a continuous dependency on all ANN parameters (as the weights) provides a multidimensional hyperplane in an (n+1)-dimensional space – with “n” being the number of FFP variables. The (n+1)-th dimension is for the cost values. As the the FFP-algorithm depends on a multitude of linear and non-linear operations we expect that the hyperplane-surface will have a rather complex form – with maxima and minima as well as so called saddle points.

However, if we construct the cost function cleverly the optimum values for the ANN’s weights will lead to a global minimum of this hyperplane – which then in turn corresponds to a minimum of distances between the propagation results and the known values for the training data:

The task to find categorical separation surfaces in the vector space of input data is reformulated as an optimization task in the cost-weight vector space: There it means finding a (global) minimum of the cost hyperplane.

Let us assume we sit at some point on a yet unexplored hyperplane. A quite general way to find the (global) minimum of this hyperplane is to follow a path indicated by the (tangential) gradient vector at the local point: The gradient is vertically oriented with respect to contour lines of constant cost values on the hyperplane. It thus gives us the direction along which a maximum cost change occurs per unit change of some weights. Calculating corrections of the weights translates into following the gradient with small steps. Geometrically speaking:

We follow the direction the overall gradient points to – and translate the movement into to small components along each weight axis – which gives us the individual weight corrections. Our hope is that the overall gradient points into the direction of the global minimum. (In case of local minima or large planes of the hyperplane we would have to adopt the step size somehow.)

This is called the “gradient descent method“. In one of the next contributions to this article series we shall see how this in turn efficiently translates into the backward propagation of errors through the network via matrix operations. Our optimization task is thus reduced to a
systematic variation of the weights during gradient descent with a series of mathematical operations determining gradient components and resulting weight corrections.

Smooth or stochastic gradient descent?

The cost function absorbs complexity stemming from the large amount of all training data rather smoothly by summing up the individual contributions of training data records. Let us look a bit at the gradient: Normally we would have to calculate partial derivatives of all cost contributions off all data sets with respect to all individual weights. For big training data sets this corresponds to a lot of mathematical operations – both matrix operations (linear algebra) and value calculations of nonlinear (activation and output) functions.

What happens if we took not all data records but concentrated on the contributions of selected input data, only? And corrected afterwards again for another disjunctive set of selected data points? I.e. what if we calculated the full required correction only piece-wise for different collections (mini-batches) of input data records?

Then the reduced gradient components would guide us into a direction on the hyperplane which deviates from the overall gradients direction. Taking the next data record would correct this movement a bit into another direction again. If we perform gradient correction for batches of different data records or in the extreme case for individual records we would move somewhat erratically around the overall gradient’s direction; we speak of a “stochastic gradient descent” [SGD].

The erratic movement of SGD helps to overcome local overall minima. But all in all it may take more steps to come to a global minimum or at least close to it – as the a stochastic movement may never converge into the overall minimum’s point in the weight space – but hop instead around it.

The question of how many input data we include in the cost function determining one single weight correction step during an epoch leads to the choice between the following cases:

  • stochastic gradient descent (sequence of weight corrections during an epoch – each based on just one training data record at a time and for all weights),
  • full batch gradient descent (one weight correction per epoch – based on all training data records and for all weights),
  • mini-batch gradient descent (sequence of weight corrections during an epoch – each based on a batch of multiple training data records and for all weights).

A stochastic or mini-bath based gradient descent may mean much faster corrections in terms of a reduced number of (vectorized) mathematical operations and CPU consumption – at least at the beginning of the descent. The CPU time of the training process for large amounts of input data may actually be reduced by factors!

In the case of mini-batches we can, therefore, optimize the performance by varying the mini-batch size. The required matrix operations can be performed vectorized over all data records of the batch; i.e. the operations can be performed “in parallel”. Fortunately, we do not need to care about the necessary CPU register handling whilst coding – optimized libraries will take care of this. As we have seen already in this blog, also threading for a reasonable amount of CPU cores may influence the performance on a specific system a lot.

For our Python class we will therefore provide parameters for the size of a mini-batch – and adapt both the calculation of cost-contributions and respective weight corrections accordingly.

Note that we do not only hope for that the weights determined by gradient descent provide reasonable result values for the training data but also for any other data later on provided to the ANN/MLP. Solving the optimization problem in the end must provide reliable and complex separation surfaces in
the multidimensional input vector space (for MNIST with a dimension of n=784). The mathematical equivalence of the problem of finding separation surfaces in the input vector space to the optimization problem in the result space can be proven for regression problems. (Actually, I do not know whether a mathematical equivalence has been proven for general problems. So, for some ML classification tasks gradient descent may not work sufficiently well.)

Choosing a cost function

Cost functions should be designed carefully. A “cost function” must have certain properties for the so called “gradient descent method” to work successfully:

  • For convenience the global extremum should be a minimum.
  • The cost function must be continuous and differentiable with respect to the ANN’s weights.
  • The requirement of differentiability translates back to the requirement of differentiable activation and output functions – as we shall see in detail in a later article.
  • It should expose a basic convex form in the surroundings of the global minimum (second partial derivatives > 0).
  • The “cost function” must have certain properties for making use of an efficient way to calculate gradients, i.e. partial derivatives. We shall see that some reasonable cost functions turn this task into a back propagation of errors. The efficiency comes via similar matrix operations as those used in the forward propagation algorithm.

Besides choosing a cost function carefully also the choice of the activation function is important for the success of gradient descent. The path to global minimum on a hyperplane may also depend on the starting point (defined by the statistically chosen initial weight values) as well as on an adaptive step size (called learning rate).

Most Machine Learning algorithms can incorporate a variety of reasonable “cost functions. For classification tasks often the following cost functions are used:

  • Categorial Cross-Entropy
  • Log Loss ( = Logistic Regression Loss )
  • Relative Entropy,
  • Exponential Loss
  • MSE (Mean Square Error)

Each of these functions is more or less appropriate for a specific type of classification problem. See the literature for more information on each of these cost functions.

In our code for MNIST-problem we will only include two of these functions as a starting point – Log Loss and MSE. MSE is e.g. used by T. Rashid in his book (see section Links) on building an MLP with Python for the MNIST case. Information on the Log Loss function are provided by the book of Rashka and the book of Geron; see the references in the section “Links” below.

Do we need cost function values at all?

The training of an ANN – i.e. the optimization of weights – does not require the explicit calculation of cost values. The reason for this is of course that gradient descent first of all works with partial derivatives with respect to weights. To calculate them we must use the chain rule with respect to the activation function, the output of lower layers and so on. But the cost values themselves are nowhere required. As a consequence in all of the book of T. Rashid on “Make your Own Neural network” the calculation of costs is never encoded.

Nevertheless, in the next article of this series we shall discuss the code for cost calculations of mini-batches. The reason for this is that we can use the cost values to study the progress of training and the convergence into a minimum: The change of total “costs” provides a way to control and watch the
success of training through its epochs.

Summary and conclusion

The concept of a cost function is central to MLPs and classification tasks: Classification means to separate the input data into categories. The task to find categorical separation surfaces in the vector space of input data is reformulated as an optimization task. This in turn requires us to find a minimum of the cost/loss hyperplane over the multidimensional space of potential weight-parameters. Calculating corrections of the weights during following a gradient guided path to a minimum in turn efficiently translates into the backward propagation of errors through the network via matrix operations.

Links and Literature

https://www.python-course.eu/matrix_arithmetic.php

Gradient descent and cost functions
towardsdatascience.com understanding-the-mathematics-behind-gradient-descent-dde5dc9be06e
ml-cheatsheet readthedocs – gradient-descent.html
page.mi.fu-berlin.de
neural chapter K7.pdf

Regularization
chunml.github.io tutorial on Regularization/

Books
“Neuronale Netze selbst programmieren”, Tariq Rashid, 2017, O’Reilly Media Inc. + dpunkt.verlag GmbH
“Machine Learning mit SciKit-Learn & TensorFlow, Aurelien Geron, 2018, O’Reilly Media Inc. + dpunkt.verlag GmbH
“Python machine Learning”, Seb. Raschka, 2016, Packt Publishing, Birmingham, UK
“Machine Learning mit Sckit-Learn & TensorFlow”, A. Geron, 2018, O’REILLY, dpunkt.verlag GmbH, Heidelberg, Deutschland