Tag Archives: CNN

Autoencoders and latent space fragmentation – VIII – approximation of the latent vector distribution by a multivariate normal distribution and ellipses

Posted on by

This post series is about creative abilities of convolutional Autoencoders [AE] which have been trained on a set of human face images. The objectives of this series and its numerical experiments are:

  • We want to create images with human faces from statistical z-vectors and related z-points in the AE’s latent space [z-space or LS]. Image creation will be done with the help of the AE’s Decoder after a training on the CelebA dataset.
  • We work with a standard Autoencoder, only. I.e., we do NOT add any artificial layers and cost terms to the Autoencoder’s layer structure (as it is done e.g. in Variational Autoencoders).
  • We analyze the position, shape and internal structure of the multidimensional z-vector distribution created by the AE’s Encoder after training. We assume that generated statistical z-vectors must point to respective regions of the latent space to guarantee images with reasonable content.
  • We raise the question whether simple statistical generator algorithms are sufficient to cover these regions with statistical z-vectors.

Our numerical experiments gave us some indications that such an endeavor is indeed feasible. In addition the third objective may give us some insight into the rules a trained AE follows when it encodes information about human faces into vectors of its latent space.

We have already studied the “natural” z-vector distribution created by a convolutional Autoencoder for CelebA images after a thorough training. The related z-point distribution fortunately filled just one confined and coherent off-center region of the AE’s latent space. Our experiments have furthermore shown that we must indeed restrict the statistical z-vector creation such that the vectors point to this particular region. Otherwise we will not get reasonable images. For details see the previous posts.

Autoencoders and latent space fragmentation – VII – face images from statistical z-points within the latent space region of CelebA
Autoencoders and latent space fragmentation – VI – image creation from z-points along paths in selected coordinate planes of the latent space
Autoencoders and latent space fragmentation – V – reconstruction of human face images from simple statistical z-point-distributions?
Autoencoders and latent space fragmentation – IV – CelebA and statistical vector distributions in the surroundings of the latent space origin
Autoencoders and latent space fragmentation – III – correlations of latent vector components
Autoencoders and latent space fragmentation – II – number distributions of latent vector components
Autoencoders and latent space fragmentation – I – Encoder, Decoder, latent space

The frustrating point so far was that simple methods for creating statistical vectors fail to put the end-points of the z-vectors into the relevant latent space region. In particular methods based on constant probability distributions within a common value interval for all z-vector components are doomed to miss the interesting region due to intricate mathematical reasons.

Afterward we tried to restrict the component values of test vectors to intervals defined by the shape of the number distribution for the values of each component of the CelebA related z-vectors. Such a distribution is nothing else than a one-dimensional probability density function for our special set of encoded CelebA samples: The function describes the probability that a component of a z-vector for human face images gets a value within a certain small value range. The probability distributions for all z-vector components were bell shaped and showed clear transitions to flat wings with very low values. See the plots below. This allowed us to define a value range

d_j_l   <   x_j   <   d_j_h

for each vector component x_j.

But keeping statistical values per component within the identified respective interval was not a sufficient restriction. We saw this clearly in the last post from significant irregular fluctuations in the reconstructed images. Obviously the components of statistically generated z-vectors must in addition fulfill correlation conditions.

The questions which I want to answer in this post are:

  • Can we approximate the 1-dimensional probability density functions for the z-vector components by some simple and common mathematical function?
  • What kind of correlations do we find between the components of the z-vectors encoding the information of human face images?
  • Can we derive some mathematical description of the multivariate z-vector distribution created by convolutional AEs for human face images in the AE’s multidimensional latent space?

Correlations are to be expected …

Please note: We deal with a multidimensional problem. A single latent vector encodes information about a human face image via all of its component values and by relations between these values. Regarding the purpose and the task an AE has to fulfill, it would be naive to assume that the components of our multi-dimensional z-vectors were independently organized. A z-vector encodes information for a convolutional Decoder to combine patterns detected by the Encoder and represented in neural (feature) maps of the networks to create an image. This is a subtle business. Just think about what you do when you draw a sketch of a human face. There are a lot of rules you follow.

When you think about the properties of basic feature patterns in a human face you would certainly assume that the pixel data of a corresponding image show strong correlations. This is among other things due to obvious symmetries – not excluding fluctuations of basic parameters describing human face features. But a nose tends to be at a position below the eyes and at a mid-distance of the eyes. In additions fluctuations of face features would on average respect certain limits given by natural proportions of a face. It would therefore be unreasonable to assume that the input for a Decoder to create a superposition of elementary patterns consists of un-correlated data. Instead the patterns in the original data should not only lead to well adjusted weights in the convolutional networks’ feature maps, but also to well regulated structural elements in the data distribution in the target space of the information encoding, namely in the latent space.

If the relations of the vector components were of a complex, highly non-linear kind and involved many dimensions at the same time we might be lost. But the results we have gained so far indicate a proper common structure of at least the density function for the individual components. This gives us some hope that the multidimensional problem somehow involves well defined 1-dimensional constituents. Whether this a sign that the multidimensional structure of the z-vector distribution can be decomposed into low-dimensional relations remains to be seen.

Observations regarding the z-vector distribution created by convolutional Autoencoders for human face images

Coordinate values of the z-points are identical to z-vector component values when we fix the end of each vector to the origin of the latent space coordinate system. The z-vector distribution thus directly corresponds to a z-point density distribution in the orthogonal coordinate system of the AE’s multi-dimensional LS. We have already made three interesting observations regarding these distributions:

  • The individual probability density function for a selected component of the latent vectors has a bell-shaped form. One , therefore, is tempted to think of a Gaussian function. This would indicate a possible normal distribution for the coordinate values of the z-points along each of the selected coordinate axes.
    Note: This does not exclude that the probability distributions for the components are correlated in some complex way.
  • When we plotted the projection of the z-point distribution onto 2-dimensional coordinate planes (for selected pairs of coordinate axes) then almost all of the resulting 2-dimensional density distributions seemed to have a defined core with an ellipsoidal form of its boundary.
  • For certain component- or axis-pairs the main axes of the apparent ellipses for pair-wise density function appeared rotated against the coordinate axes. The elongated regular and more or less symmetric forms showed a diagonal orientation (with different angles). This alone signals a strong correlation between related two vector components. Indeed we found high values for certain elements of the matrix of normalized Pearson correlation coefficients for the multi-dimensional distribution of z-vector component values.

These observations are not unrelated; they indicate a clear pattern of dependencies and correlations of the distributions for the variables in place. Regarding the data basis we have to keep five things in mind:

  • We treat the z-point distribution for CelebA images as a multi-dimensional probability density distribution. During the analysis we look in particular at 2-dimensional projections of this distribution onto planes spanned by a selected pair of orthogonal axes of the LS coordinate system. We also consider the one-dimensional value distributions for z-vector components. In this sense we regard the z-vector components as logically separate variables.
  • The data used are numbers of z-points counted in finite 1d-intervals, 2d-rectangles or multidimensional cuboids. We fit idealized functions to the respective discrete bar plots. Even if there is a good 1d-fit fluctuations may especially get visible in multidimensional plots for correlated data. A related probability density requires a normalization. We drop the resulting constant factors in the qualitative discussions below.
  • Statistical (un-)correlation of statistical variable distributions must NOT to be confused with underlying variable (in-) dependency. Linear correlations can be reduced to zero by coordinate transformations without eliminating the original variable dependencies.
  • Pearson correlation coefficients are sensitive to linear elements in the relations of logically separate variable distributions. They can not fully cover non-linear distribution relations or covered variable dependencies.
  • A transformation to a local coordinate system whose axes are aligned to the so called main axes of the multidimensional distributions does not remove the original data relations – but there may exist a coordinate system in which the distribution data can be described in a simple, factorized form corresponding to a composition of seemingly un-correlated data distributions.

Anyway – by discussing density distributions we work on overall and large scale average relations between statistical value distributions for our variables, namely the z-vector components. We do not cover local micro-relations that may be in place in addition.

The relation of ellipses with Gaussian probability densities

Probability density functions for two logically separate, but maybe not un-correlated variables have to be multiplied. In our case this reflects the following point: First we determine the probability that the value of component x_i lies in a certain (infinitesimal) interval and then we determine the probability that (for the given value of x_i) the component x_j falls into another value range. The distributions for a specific variable can include variable relations and thus the probability density g(x_j) can include a dependency g(x_j(x_i)).

In the case of uncorrelated normal distributions per coordinate we can just multiply the individual Gaussians g(x_i) * g(x_j). Due to the quadratic terms in the exponent of the Gaussians we then get a sum of quadratic expressions in the common exponent, having the form fac1 * (x_i-mu_i)**2 + fac2 * (x_j-mu_j)**2.

By setting this expression to a constant value we get contour lines of the probability density distribution for the (x_i, x_j)-distribution. Quadratic sums correspond to the definition of an ellipse having main axes which are aligned with the x_i- and x_j-axes of the coordinate system. Thus the contour lines of a 2-dimensional distribution composed of un-correlated Gaussians are ellipses having an orientation aligned with the coordinate axes.

This was for un-correlated density-distributions of two vector components. Mathematically a linear correlation between a pair of Gaussians-distributions corresponds to an affine transformation of the contour-ellipses. The transformation can be expressed by a defined sequence of matrix operations describing a translation, rotations (in a defined order) and a dilation.

This means: The contour lines for a 2-dimensional probability density composed of linearly correlated Gaussians are still ellipses. But these ellipses will appear to be shifted, rotated and stretched along the main axes in comparison with their originally un-correlated Gaussian counterparts. The angle of rotation depends on details of the correlation function and the original standard deviations. The Pearson correlation matrix for linearly correlated distributions is a positive-definite one and, of course, shows off-diagonal elements different from zero. This result can be extended to multivariate normal distributions in spaces with many dimensions and related affine transformations of the coordinate system.

A multivariate normal distribution with linear correlations between the Gaussians results in elliptic contour lines for pair-wise density distributions in the respective 2D-coordinate planes of an orthogonal coordinate system. When we define the contours via multiples of the standard deviations of the underlying Gaussian functions we arrive at so called confidence ellipses.

A really nice mathematical aspect is that the basic parameters of the confidence ellipses can be derived from the normalized correlation coefficients of the Pearson matrix of the multivariate probability distribution. I will come back to this point in forthcoming posts in more detail. For now we just need to know that a multidimensional probability density comes along with confidence ellipses which can be calculated with the help of Pearson correlation coefficients.

Before we go on a word of caution: For a general multi-variate distribution it is not at all clear that it should decompose into a factorized form. However, for a multivariate normal distribution with un-correlated or only linearly correlated components this is by definition different. In this case a transformation to a coordinate system can be found which leads to a complete decomposition into a product of (seemingly) un-correlated Gaussians per component. The latter point lies at the center of PCA and SVD algorithms, which diagonalize the Pearson correlation matrix.

Do we really have Gaussian probability distributions for the individual z-vector-components?

After this short tour into the world of (multi-variate) normal distributions, Gaussian functions and related ellipses we are a bit better equipped to understand the density distributions in the latent space of our Autoencoders for human face images.

Let me remind you about the shapes of the number distribution for our concrete z-vector components resulting for for CelebA face images. The first plot shows the number densities on sampling intervals of width 0.25 for selected vector components resulting for case I of our experiments. The second plot shows the number densities for the values of selected components of case II.

Ok, these curves do resemble Gaussians and some fluctuations are normal. But can we prove the Gaussian properties of the curves a bit better?

Well, for case II I have drawn the best fits by Gaussian functions with the help of SciPy’s optimize.curve_fit() for 3 and yet another 4 selected components of the latent vectors and the respective number distribution curves. The dashed lines show the approximations by Gaussian functions:

The selected components are part of the list of around 20 dominant component distributions – due to their relatively large standard deviations. But the Gaussian form is consistently found for all components (with some small deviations regarding the symmetry of the curves).

So all in all it looks like as if our convolutional AE has indeed created a multivariate normal z-point distribution in the latent space. As said: This does not exclude correlations …

Pairwise linear correlations of the (normal) probability distributions for the latent vector components?

Now we are a bit bold – and assume the best case for us: Could the approxiate Gaussians distributions for the component values be pair-wise and linearly correlated? What would be a clear indication of a pair-wise linear correlation of our component distributions?

Well, we should find an elliptic form of contour lines in the 2-dimensional distribution for the component pair in the respective coordinate plane of the basic orthogonal LS coordinate system. This imposes quite strong symmetry conditions on the contour lines. The ellipses can be shifted and rotated – but they should remain being ellipses. If non-linear contributions to the correlation had a significant impact this would not be the case.

Practically it is not trivial to prove that we have approximately rotated ellipses in 2 dimensions. Satter plos alone do not help: Ellipses fit a lot of plotted distributions of discrete data points quite well. We really need to count number densities to get reliable contour lines. The following plots show such contour lines based on number sampling in rectangles and local smoothing operations with the help of scipy.stats.gaussian_kde().

The fat red and dark orange lines show corresponding confidence ellipses derived from the original CelebA distribution. See below for some remarks on confidence ellipses.

The contours are basically of elliptic shape although they do not show the complete symmetry expected for pure and linearly dependent Gaussian distributions. But overall the confidence ellipses fit quite well into the general form and orientation of the distributions. We also see that for higher σ-levels the coincidence with nearby contours is quite good. The wiggles in the contour change with the z-vector selection a bit.

We conclude that our basic impression regarding an elliptic shape of the z-point distributions is basically consistent with only linearly correlated Gaussian probability density distributions for the component values of the latent vectors.

Approximation of the core of the multivariate z-point distribution by confidence ellipses for component pairs

Above I referred to the boundary of a core of the probability density for two selected vector components. But how would we define the “boundary” of a continuous distribution in the coordinate planes? Answer: As we like – but based on the decline of the approximate Gaussian curves.

We can e.g. pick two times the half-width in each direction or we can use contours defined by confidence levels.
For 2 ≤ fact * σ ≤ 3 we saw already that the contour lines could well be fitted by confidence ellipses. A 3-sigma level covers around 97% of all data points or more. A 2 sigma-level ellipse encircles between 70% and 90% of all data points, depending on the eccentricity of the ellipse. Note that the numbers are smaller for ellipses than for rectangles. I.e. the standard 68-95-99.7 rule does not apply.

The plots below give you an impression of how well ellipses for a 2σ-confidence level approximate the core of the CelebA distribution in selected 2D coordinate planes of the latent space:

Each of the sub-plots was based on 10,000 statistically selected vectors of the 170,000 available in my test runs. This is a relative low number. Therefore, for a certain diameter of the points in the scatter plot only the inner core appears to be densely populated. The next plots shows the results for a 3 σ-level of the ellipse – but this time for 50,000 vectors. With more vectors we could visually fill the outer regions of the core.

The orange points mark the center of the multidimensional distributions derived from the one-dimensional distribution curves for the components. We see that it does not always appear to be optimally centered. There are multiple reasons: Our functions are not fully symmetric as ideal Gaussians. And equally important: The accuracy of the position depends on the sampling resolution which was coarse. Outliers of the distribution do have an impact.

And how would we explain the appearance of Gaussians and ellipses?

This all looks quite good, despite some notable deviations regarding symmetry and maxima. Gaussians fit at least most of the important probability density curves very well, though not by a 100%. The appearance of an elliptic shape of the inner core of the distribution and the appearance of overall elliptic contour curves can be explained by linear correlations of the Gaussian distributions for the components.

The appearance of normal distributions per component and basically linear correlations is something that really should be explained. I mean, dwell a bit on what we have found:

A convolutional Autoencoder network with more than 10 million adjustable parameters encoded information about human face images in the form of a roughly multivariate normal distribution of z-points in its latent space – with basically linear correlations between the Gaussian curves describing the probability densities functions for the component values of the z-vectors.

I find this astonishing and not at all self-evident. It is one of the most simple solutions for a multidimensional situation one can imagine. The following questions automatically came to my mind:

Does such a result only appear for training images of defined objects with some Gaussian variation in their features? Are the normal distributions a reflection of variations of relevant features in the original data?
Is this a typical result for (convolutional) AEs? How does it depend on the dimensionality of the latent space? Does it automatically come with a large number of z-space dimensions? Is it an efficient way to encode feature differences in the latent space, which (convolutional) AEs in general tend to use due to their structure?

Do I personally have a convincing explanation? No. Especially not, as the data shown above stem from convolutional neural networks [CNNs] without any batch-normalization layers.

A first idea would be that the dominant features of a human face themselves show variations described by Gaussian normal distributions already in the original data and that convolutional filtering does not destroy such distributions during optimization. A problem of this idea lies in the (non-) linear activation functions used at the nodes of the neural maps. Though ReLU, Leaky ReLU and SeLU contain linear parts.

The other problem is the linear form of the correlations. This is a rather simple kind of correlations. But why should an AE choose this simple form into its mapping of image information to latent space vectors after training?

How to generate statistical vectors for the creation of human face images?

The positive message which comes with the above results is that our problem of how to create proper statistical z-vectors decomposes into a sequence of two-dimensional problems. We can use the data of the ellipses appearing in the density-distributions for pairs of vector components to confine the components of statistically generated z-vectors to the relevant region in the latent space. All ellipses together restrict the component values in a well defined form. In the next post I will shortly outline some methods of how we can use the information contained in the ellipses with available algorithms.

Conclusion

In this post we have seen that for the case of a convolutional Autoencoder trained on CelebA human face images the latent vector distributions showed some remarkable properties:

The probability density functions for all component values can roughly be approximated by Gaussian functions. The components appear to be pairwise linearly correlated – at least to first order analysis. This automatically implies elliptic contour curves for the pairwise number density functions of coordinate values. Such contour curves were indeed found with first order accuracy. The core of the probability density for the z-points in the latent space could therefore be approximated by confidence ellipses for a σ-level above σ = 2.5.
The elliptic conditions correspond to a multivariate normal distribution with linear correlations of the variables.

Before we get to enthusiastic about these findings we should be careful and await a further test. All statements refer to a first order approximations. A real multivariate normal distribution would decompose into un-correlated Gaussians and 2D-ellipses of probability densities of component pairs after a PCA transformation.

In the next post

Autoencoders and latent space fragmentation – IX – PCA transformation of the z-point distribution for CelebA

I shall present the results of a PCA analysis. In later posts I will introduce a related method to restrict the components of statistical vectors to the relevant region in the latent space of our Autoencoder.

Links and literature

On first sight my short description of the relation between multivariate Gaussian normal distributions and ellipses as the contour lines for the projected density distributions on coordinate planes may appear plausible. But in the general multi-dimensional case the question of linear correlations requires some more math than indicated. For details I just refer to some articles on the Internet – but any good book on multivariate analysis will give you the relevant information
https://de.wikipedia.org/ wiki/ Multivariate_ Normalverteilung
https://de.wikipedia.org/ wiki/ Mehrdimensionale_ Normalverteilung
http:// www.mi.uni-koeln.de/ ~jeisenbe/ Vortrag2.pdf
https://methodenlehre.uni-mainz.de/ files/ 2019/06/ Multivariate-Distanz-Normalverteilung-MDC-Bayes.pdf
https://en.wikipedia.org/ wiki/ Multivariate_ normal_ distribution
https://en.wikipedia.org/ wiki/ Confidence_region
https://users.cs.utah.edu/ ~tch/ CS6640F2020/ resources/ How to draw a covariance error ellipse.pdf
https://biotoolbox.binghamton.edu/ Multivariate Methods/ Multivariate Tools and Background/ pdf files/ MTB%20070.pdf

Regarding the intimate relation between the ellipses’ main axes to normalized Pearson correlation coefficients I also refer to
https://carstenschelp.github.io/ 2018/09/14/ Plot_ Confidence_ Ellipse_ 001.html
I am very grateful that the author Carsten Schelp saved me a lot of time when trying to find a way to program a solution for confidence ellipses. Thank you, Mr. Schelp for the great work.

 

Autoencoders, latent space and the curse of high dimensionality – I

Posted on by

Recently, I had to give a presentation about standard Autoencoders (AEs) and related use cases. Whilst preparing examples I stumbled across a well-known problem: The AE solved tasks as to reconstruct faces hidden in extreme noisy or leaky input images perfectly. But the reconstruction of human faces from arbitrarily chosen points in the so called “latent space” of a standard Autoencoder did not work well.

In this series of posts I want to discuss this problem a bit as it illustrates why we need Variational Autoencoders for a systematic creation of faces with varying features from points and clusters in the latent space. But the problem also raises some fundamental and interesting questions

  • about a certain “blindness” of neural networks during training in general, and
  • about the way we save or conserve the knowledge which a neural network has gained about patterns in input data during training.

This post requires experience with the architecture and principles of Autoencoders.

Note, 02/14/2023: I have revised and edited this post to get consistent with new insights from extended experiments with AEs and VAEs.

Standard tasks for conventional Autoencoders

For preparing my talk I worked with relatively simple Autoencoders. I used Convolutional Neural Networks [CNNs] with just 4 convolutional layers to create the Encoder and Decoder parts of the Autoencoder. As typical applications I chose the following:

  • Effective image compression and reconstruction by using a latent space of relatively low dimensionality. The trained AEs were able to compress input images into latent vectors with only few components and reconstruct the original image from the compressed format.
  • Denoising of images where the original data were obscured by the superposition of statistical noise and/or statistically dropped pixels. (This is my favorite task for AEs which they solve astonishingly well.)
  • Recolorization of images: The trained AE in this case transforms images with only gray pixels into colorful images.

Such challenges for AEs are discussed in standard ML literature. In a first approach I applied my Autoencoders to the usual MNIST and Fashion MNIST datasets. For the task of recolorization I used the Cifar 10 dataset. But a bit later I turned to the Celeb A dataset with images of celebrity faces. Just to make all of the tasks a bit more challenging.

Standard Autoencoders and low dimensions of the latent space for (Fashion) MNIST and Cifar10 data

My Autoencoders excelled in all the tasks named above – both for MNIST, CELEB A and, regarding recolorization, CIFAR 10.

Regarding MNIST and MNIST/Fashion 4-layer CNNs for the Encoder and Decoder are almost an overkill. For MNIST the dimension z_dim of the latent space can be chosen to be pretty small:

z_dim = 12 gives a really good reconstruction quality of (test) images compressed to minimum information in the latent space. z_dim=4 still gave an acceptable quality and even with z_dim = 2 most of test images were reconstructed well enough. The same was true for the reconstruction of images superimposed with heavy statistical noise – such that the human eye could no longer guess the original information. For Fashion MNIST a dimension number 20 < z_dim < 40 gave good results. Also for recolorization the results were very plausible. I shall present the results in other blog posts in the future.

Face reconstructions of (noisy) Celeb A images require a relative high dimension of the latent space

Then I turned to the Celeb A dataset. By the way: I got interested in Celeb A when reading the books of David Foster on “Generative Deep Learning” and of Tariq Rashi “Make Your First GANs with PyTorch” (see the complete references in the last section of this post).

The Celeb A data set contains images of around 200,000 faces with varying contours, hairdos and very different, in-homogeneous backgrounds. And the faces are displayed from very different viewing angles.

For a good performance of image reconstruction in all of the named use cases one needs to raise the number of dimensions of the latent space significantly. Instead of 12 dimensions of the latent space as for MNIST we now talk about 200 up to 1200 dimensions for CELEB A – depending on the task the AE gets trained for and, of course, on the quality expectations. For reconstruction of normal images and for the reconstruction of clear images from noisy input images higher numbers of dimensions z_dim ≥ 512 gave visibly better results.

Actually, the impressive quality for the reconstruction of test images of faces, which were almost totally obscured by the superimposition of statistical noise or the statistical removal of pixels after a self-supervised training on around 100,000 images surprised me. (Totalitarian states and security agencies certainly are happy about the superb face reconstruction capabilities of even simple AEs.) Part of the explanation, of course, is that 20% un-obscured or un-blurred pixels out of 30,000 pixels still means 6,000 clear pixels. Obviously enough for the AE to choose the right pattern superposition to compose a plausible clear image.

Note that we are not talking about overfitting here – the Autoencoder handled test images, i.e. images which it had never seen before, very well. AEs based on CNNs just seem to extract and use patterns characteristic for faces extremely effectively.

But how is the target space of the Encoder, i.e. the latent space, filled for Celeb A data? Do all points in the latent space give us images with well recognizable faces in the end?

Face reconstruction after a training based on Celeb A images

To answer the last question I trained an AE with 100,000 images of Celeb A for the reconstruction task named above. The dimension of the latent space was chosen to be z_dim = 200 for the results presented below. (Actually, I used a VAE with a tiny amount of KL loss by a factor of 1.e-6 smaller than the standard Binary Cross-Entropy loss for reconstruction – to get at least a minimum confinement of the z-points in the latent space. But the results are basically similar to those of a pure AE.)

My somewhat reworked and centered Celeb A images had a dimension of 96×96 pixels. So the original feature space had a number of dimensions of 27,648 (almost 30000). The challenge was to reproduce the original images from latent data points created of test images presented to the Encoder. To be more precise:

After a certain number of training epochs we feed the Encoder (with fixed weights) with test images the AE has never seen before. Then we get the components of the vectors from the origin to the resulting points in the latent space (z-points). After feeding these data into the Decoder we expect the reproduction of images close to the test input images.

With a balanced training controlled by an Adam optimizer I already got a good resemblance after 10 epochs. The reproduction got better and very acceptable also with respect to tiny details after 25 epochs for my AE. Due to possible copyright and personal rights violations I do not dare to present the results for general Celeb A images in a public blog. But you can write me a mail if you are interested.

Most of the data points in the latent space were created in a region of 0 < |x_i| < 20 with x_i meaning one of the vector components of a z-point in the latent space. I will provide more data on the z-point distribution produced by the Encoder in later posts of this mini-series.

Face reconstruction from randomly chosen points in the latent space

Then I selected arbitrary data points in the latent space with randomly chosen and uniformly distributed components 0 < |x_i| < boundary. The values for boundary were systematically enlarged.

Note that most of the resulting points will have a tendency to be located in outer regions of the multidimensional cube with an extension in each direction given by boundary. This is due to the big chance that one of the components will get a relatively high value.

Then I fed these arbitrary z-points into the Decoder. Below you see the results after 10 training epochs of the AE; I selected only 10 of 100 data points created for each value of boundary (the images all look more or less the same regarding the absence or blurring of clear face contours):

boundary = 0.5

boundary = 2.5

boundary = 5.0

boundary = 8.0

boundary = 10.0

boundary = 15.0

boundary = 20.0

boundary = 30.0

boundary = 50

This is more a collection of face hallucinations than of usable face images. (Interesting for artists, maybe? Seriously meant …).

So, most of the points in the latent space of an Autoencoder do NOT represent reasonable faces. Sometimes our random selection came close to a region in latent space where the result do resemble a face. See e.g. the central image for boundary=10.

From the images above it becomes clear that some arbitrary path inside the latent space will contain more points which do NOT give you a reasonable face reproduction than points that result in plausible face images – despite a successful training of the Autoencoder.

This result supports the impression that the latent space of well trained Autoencoders is almost unusable for creative purposes. It also raises the interesting question of what the distribution of “meaningful points” in the latent space really looks like. I do not know whether this has been investigated in depth at all. Some links to publications which prove a certain scientific interest in this question are given in the last section of this posts.

I also want to comment on an article published in the Quanta Magazine lately. See “Self-Taught AI Shows Similarities to How the Brain Works”. This article refers to “masked” Autoencoders and self-supervised learning. Reconstructing masked images, i.e. images with a superposition of a mask hiding/blurring pixels with a reasonably equipped Autoencoder indeed works very well. Regarding this point I totally agree. Also with the term “self-supervised learning”.

But to suggest that an Autoencoder with this (rather basic) capability reflects methods of the human brain is in my opinion a massive exaggeration. On the contrary, in my opinion an AE reflects a dumbness regarding the storage and usage of otherwise well extracted feature patterns. This is due to its construction and the nature of its mapping of image contents to the latent space. A child can, after some teaching, draw characteristic features of human faces – out of nothing on a plain white piece of paper. The Decoder part of a standard Autoencoder (in some contrast to a GAN) can not – at least not without help to pick a meaningful point in latent space. And this difference is a major one, in my opinion.

A first interpretation – the curse of many dimensions of the latent space

I think the reason why arbitrary points in the multi-dimensional latent space cannot be mapped to images with recognizable faces is yet another effect of the so called “curse of high dimensionality”. But this time also related to the latent space.

A normal Autoencoder (i.e. one without the Kullback-Leibler loss) uses the latent space in its vast extension to produce points where typical properties (features) of faces and background are encoded in a most unique way for each of the input pictures. But the distinct volume filled by such points is a pretty small one – compared to the extensions of the high dimensional latent space. The volume of data points resulting from a mapping-transformation of arbitrary points in the original feature space to points of the latent space is of course much bigger than the volume of points which correspond to images showing typical human faces.

This is due to the fact that there are many more images with arbitrary pixel values already in the original feature space of the input images (with lets say 30000 dimensions for 100×100 color pixels) than images with reasonable values for faces in front of some background. The points in the feature space which correspond to reasonable images of faces (right colors and dominant pixel values for face features), is certainly small compared to the extension of the original feature space. Therefore: If you pick a random point in latent space – even within a confined (but multidimensional) volume around the origin – the chance that this point lies outside the particular volume of points which make sense regarding face reproduction is big. I guess that for z_dim > 200 the probability is pretty close to 1.

In addition: As the mapping algorithm of a neural Encoder network as e.g. CNNs is highly non-linear it is difficult to say how the boundary hyperplanes of mapping areas for faces look like. Complicated – but due to the enormous number of original images with arbitrary pixel values – we can safely guess that they enclose a rather small volume.

The manifold of data points in the z-space giving us recognizable faces in front of a reasonably separated background may follow a curved and wiggly “path” through the latent space. In principal there could even be isolated unconnected regions separated by areas of “chaotic reconstructions”.

I think this kind of argumentation line holds for standard Autoencoders and variational Autoencoders with a very small KL loss in comparison to the reconstruction loss (BCE (binary cross-entropy) or MSE).

Why do Variational Autoencoders [VAEs] help?

The fist point is: VAEs reduce the total occupied volume of the latent space. Due to mu-related term in the Kullback-Leibler loss the whole distribution of z-points gets condensed into a limited volume around the origin of the latent space.

The second reason is that the distribution of meaningful points are smeared out by the logvar-term of the Kullback-Leibler loss.

Both effects enforce overlapping regions of meaningful standard Gaussian-like z-point distributions in the latent space. So VAEs significantly increase the probability to hit a meaningful z-point in latent space – if you chose points around the origin within a distance of “1” per coordinate (or vector component).

The total distance of a point and its vector in z-space has to be measured with some norm, e.g. the Euclidian one. Actually we should get meaningful reconstructions around a multidimensional sphere of radius “16”. Why this is reasonable will be discussed in forthcoming posts.

Please, also look at the series on the technical realization of VAEs in this blog. The last posts there prove the effects of the KL-loss experimentally for Celeb A data. Below you find a selection of images created from randomly chosen points in the latent space of a Variational Autoencoder with z_dim=200 after 10 epochs.

Conclusion

Enough for today. Whilst standard Autoencoders solve certain tasks very well, they seem to produce very specific data distributions in the latent space for CelebA images: Only certain regions seem to be suitable for the reconstruction of “meaningful” images with human faces.

This problem may have its origin already in the feature space of the original images. Also there only a small minority of points represents humanly interpretable face images. This becomes obvious when you look at the vast amount of possible pixel values in a feature space of lets say 96x96x3 = 27,648. Each of these dimension can get a value between 0 and 255. This gives us more than 7 million combinations. Only a tiny fraction of these possible images will show reasonable faces in the center with a reasonably structured background around.

From a first experiment the chance of hitting a data point in latent space which gives you a meaningful image seems to be small. This result appears to be a variant of the curse of high dimensionality – this time including the latent space.

In a forthcoming post
Autoencoders, latent space and the curse of high dimensionality – II – a view on fragments and filaments of the latent space for CelebA images
we will investigate the z-point distribution in latent space with a variety of tools. And find that this distribution is fragmented and that the z-points for CelebA images are arranged in certain regions of the latent space. In addition we will get indications that the distribution contains filament-like structures.

Links

https://towardsdatascience.com/ exploring-the-latent-space-of-your-convnet-classifier-b6eb862e9e55

Felix Leeb, Stefan Bauer, Michel Besserve,Bernhard Schölkopf, “Exploring the Latent Space of Autoencoders with
Interventional Assays”, 2022,
https://arxiv.org/abs/2106.16091v2 // https://arxiv.org/pdf/2106.16091.pdf
https://wiredspace.wits.ac.za/ handle/10539/33094?show=full
https://www.elucidate.ai/post/ exploring-deep-latent-spaces

Books:
T. Rashid, “GANs mit PyTorch selbst programmieren”, 2020, O’Reilly, dpunkt.verlag, Heidelberg, ISBN 978-3-96009-147-9
D. Foster, “Generatives Deep Learning”, 2019, O’Reilly, dpunkt.verlag, Heidelberg, ISBN 978-3-96009-128-8