Variational Autoencoder with Tensorflow 2.8 – XIII – Does a VAE with tiny KL-loss behave like an AE? And if so, why?

This post continues my series on Variational Autoencoders [VAE] with some considerations regarding a VAE whose settings allow only for a tiny amount of the so called Kullback-Leibler [KL] loss.

Variational Autoencoder with Tensorflow 2.8 – I – some basics
Variational Autoencoder with Tensorflow 2.8 – II – an Autoencoder with binary-crossentropy loss
Variational Autoencoder with Tensorflow 2.8 – III – problems with the KL loss and eager execution
Variational Autoencoder with Tensorflow 2.8 – IV – simple rules to avoid problems with eager execution
Variational Autoencoder with Tensorflow 2.8 – V – a customized Encoder layer for the KL loss
Variational Autoencoder with Tensorflow 2.8 – VI – KL loss via tensor transfer and multiple output
Variational Autoencoder with Tensorflow 2.8 – VII – KL loss via model.add_loss()
Variational Autoencoder with Tensorflow 2.8 – VIII – TF 2 GradientTape(), KL loss and metrics
Variational Autoencoder with Tensorflow 2.8 – IX – taming Celeb A by resizing the images and using a generator
Variational Autoencoder with Tensorflow 2.8 – X – VAE application to CelebA images
Variational Autoencoder with Tensorflow 2.8 – XI – image creation by a VAE trained on CelebA
Variational Autoencoder with Tensorflow 2.8 – XII – save some VRAM by an extra Dense layer in the Encoder

So far, most of the posts in this series have covered a variety of methods (provided by Tensorflow and Keras) to control the KL loss. One of the previous posts (XI) provided (indirect) evidence that also GradientTape()-based methods for KL-loss calculation work as expected. In stark contrast to a standard Autoencoder [AE] our VAE trained on CelebA data proved its ability to reconstruct humanly interpretable images from random z-points (or z-vectors) in the latent space. Provided that the z-points lie within a reasonable distance to the origin.

We could leave it at that. One of the basic motivations to work with VAEs is to use the latent space “creatively”. This requires that the data points coming from similar training images should fill the latent space densely and without gaps between clusters or filaments. We have obviously achieved this objective. Now we could start to do funny things like to combine reconstruction with vector arithmetic in the latent space.

But to look a bit deeper into the latent space may give us some new insights. The central point of the KL-loss is that it induces a statistical element into the training of AEs. As a consequence a VAE fills the so called “latent space” in a different way than a simple AE. The z-point distribution gets confined and areas around z-points for meaningful training images are forced to get broader and overlap. So two questions want an answer:

  • Can we get more direct evidence of what the KL-loss does to the data distribution in latent space?
  • Can we get some direct evidence supporting the assumption that most of the latent space of an AE is empty or only sparsely populated? in contrast to a VAE’s latent space?

Therefore, I thought it would be funny to compare the data organization in latent space caused by an AE with that of a VAE. But to get there we need some solid starting point. If you consider a bit where you yourself would start with an AE vs. VAE comparison you will probably come across the following additional and also interesting questions:

  • Can one safely assume that a VAE with only a very tiny amount of KL-loss reproduces the same z-point distribution vs. radius which an AE would give us?
  • In general: Can we really expect a VAE with a very tiny Kullback-Leibler loss to behave as a corresponding AE with the same structure of convolutional layers?

The answers to all these questions are the topics of this post and a forthcoming one. To get some answers I will compare a VAE with a very small KL-loss contribution with a similar AE. Both network types will consist of equivalent convolutional layers and will be trained on the CelebA dataset. We shall look at the resulting data point density distribution vs. radius, clustering properties and the ability to create images from statistical z-points.

This will give us a solid base to proceed to larger and more natural values of the KL-loss in further posts. I got some new insights along this path and hope the presented data will be interesting for the reader, too.

Below and in following posts I will sometimes call the target space of the Encoder also the “z-space“.

CelebA data to fill the latent vector-space

The training of an AE or a VAE occurs in a self-supervised manner. A VAe or an AE learns to create a point, a z-point, in the latent space for each of the training objects (e.g. CelebA images). In such a way that the Decoder can reconstruct an object (image) very close to the original from the z-point’s coordinate data. We will use the “CelebA” dataset to study the KL-impact on the z-point distribution.CelebA is more challenging for a VAE than MNIST. And the latent space requires a substantially higher number of dimensions than in the MNIST case for reasonable reconstructions. This makes things even more interesting.

The latent z-space filled by a trained AE or VAE is a multi-dimensional vector space. Meaning: Each z-point can be described by a vector defining a position in z-space. A vector in turn is defined by concrete values for as many vector components as the z-space has dimensions.

Of course, we would like to see some direct data visualizing the impact of the KL-loss on the z-point distribution which the Encoder creates for our training data. As we deal with a multidimensional vector space we cannot plot the data distribution. We have to simplify and somehow get rid of the many dimensions. A simple solution is to look at the data point distribution in latent space with respect to the distance of these points from the origin. Thereby we transform the problem into a one-dimensional one.

More precisely: I want to analyze the change in numbers of z-points within “radius“-intervals. Of course, a “radius” has to be defined in a multidimensional vector space as the z-space. But this can easily be achieved via an Euclidean L2-norm. As we expect the KL loss to have a confining effect on the z-point distribution it should reduce the average radius of the z-points. We shal later see that this is indeed the case.

Another simple method to reduce dimensions is to look at just one coordinate axis and the data distribution for the calculated values in this direction of the vector space. Therefore, I will also check the variation in the number of data points along each coordinate axis in one of the next posts.

A look at clustering via projections to a plane may be helpful, too.

The expected similarity of a VAE with tiny KL-loss to an AE is not really obvious

Regarding the answers to the 3rd and 4th questions posed above your intuition tells you: Yes, you probably can bet on a similarity between a VAE with tiny KL-loss and an AE.

But when you look closer at the network architectures you may get a bit nervous. Why should a VAE network that has many more degrees of freedom than an AE not use both of its layers for “mu” and “logvar” to find a different distribution solution? A solution related to another minimum of the loss hyperplane in the weight configuration space? Especially as this weight-related space is significantly bigger than that of a corresponding AE with the same convolutional layers?

The whole point has to do with the following facts: In an AE’s Encoder the last flattening layer after the Conv2D-layers is connected to just one output layer. In a VAE, instead, the flattening layer feeds data into two consecutive layers (for mu and logvar) across twice as many connections (with twice as many weight parameters to optimize).

In the last post of this series we dealt with this point from the perspective of VRAM consumption. Now, its the question in how far a VAE will be similar to an AE for a tiny KL-loss.

Why should the z-points found be determined only by mu-values and not also by logvar-values? And why should the mu values reproduce the same distribution as an AE? At least the architecture does not guarantee this by any obvious means …

Well, let us look at some data.

Structure of an AE for CelebA and its total loss after some epochs

Our test AE contains the same simple sequence of four Conv2D layers per Encoder and four 4 Conv2DTranspose layers as our VAE. See the AE’s Encoder layer structure below.

A difference, however, will be that I will not use any BatchNormalizer layers in the AE. But as a correctly implemented BatchNormalization should not affect the representational powers of a VAE network for very principle reasons this should not influence the comparison of the final z-point distribution in a significant way.

I performed an AE training run for 170,000 CelebA training images over 24 epochs. The latent space has a dimension if z_dim=256. (This relatively low number of dimensions will make it easier for a VAE to confine z_points around the origin; see the discussion in previous posts).

The resulting total loss of our AE became ca. 0.49 per pixel. This translates into a total value of

AE total loss on Celeb A after 24 epochs (for a step size of 0.0005): 4515

This value results from a summation over all geometric pixels of our CelebA images which were downsized to 96×96 px (see post IX). The given value can be compared to results measured by our GradientTape()-based VAE-model which delivers integrated values and not averages per pixel.

This value is significantly smaller than values we would get for the total loss of a VAE with a reasonably big KL-loss of contribution in the order of some percent of the reconstruction loss. A VAE produces values around 4800 up to 5000. Apparently, an AE’s Decoder reconstructs originals much better than a VAE with a significant KL-loss contribution to the total loss.

But what about a VAE with a very small KL-loss? You will get the answer in a minute.

Where does a standard Autoencoder [AE] place the z-points for CelebA data?

We can not directly plot a data point distribution in a 256-dimensional vector-space. But we can look at the data point density variation with a calculated distance from the origin of the latent space.

The distance R from the origin to the z-point for each image can be measured in terms of a L2 (= Euclidean) norm of the latent vector space. Afterward it is easy to determine the number of images within all radius intervals with e.g. a length of 0.5 e.g. between radii R

0  <  R  <  35 .

We perform the following steps to get respective numbers. We let the Encoder of our trained AE predict the z-points of all 170,000 training data

z_points  = AE.encoder.predict(data_flow) 

data_flow was created by a Keras DataImageGenerator to send batches of training data to the GPU (see the previous posts).

Radius values are then calculated as

print("NUM_Images_Train = ", NUM_IMAGES_TRAIN)
ay_rad_z = np.zeros((NUM_IMAGES_TRAIN,),  dtype='float32')
for i in range(0, NUM_IMAGES_TRAIN):
    sq = np.square(z_points[i]) 
    sqrt_sum_sq = math.sqrt(sq.sum())
    ay_rad_z[i] = sqrt_sum_sq

The numbers vs. radius relation then results from:

li_rad      = []
li_num_rad  = []
int_width = 0.5
for i in range(0,70):
    low   = int_width * i
    high  = int_width * (i+1) 
    num   = np.count_nonzero( (ay_rad_z >= low) & (ay_rad_z < high ) )
    li_rad.append(0.5 * (low + high))
    li_num_rad.append(num)

The resulting curve is shown below:

There seems to be a peak around R = 16.75. So, yet another question arises:

>What is so special about the radius values of 16 or 17 ?

We shall return to this point in the next post. For now we take this result as god-given.

Clustering of CelebA z-point data in the AE’s latent space?

Another interesting question is: Do we get some clustering in the latent space? Will there be a difference between an AE and a VAE?

A standard method to find an indication of clustering is to look for an elbow in the so called “inertia” curve for different assumed numbers of clusters. Below you find an inertia plot retrieved from the z-point data with the help of MiniBatchKMeans.

This result was achieved for data taken at every second value of the number of clusters “num_clus” between 2 ≤ num_clus ≤ 80. Unfortunately, the result does not show a pronounced elbow. Instead the variation at some special cluster numbers is relatively high. But, if we absolutely wanted to define a value then something between 38 and 42 appears to be reasonable. Up to that point the decline in inertia is relatively smooth. But do not let you get misguided – the data depend on statistics and initial cluster values. When you change to a different calculation you may get something like the following plot with more pronounced spikes:

This is always as sign that the clustering is not very clear – and that the clusters do not have a significant distance, at least not in all coordinate directions. Filamental structures will not be honored well by KMeans.

Nevertheless: A value of 40 is reasonable as we have 40 labels coming with the CelebA data. I.e. 40 basic features in the face images are considered to be significant and were labeled by the creators of the dataset.

t-SNE projections

We can also have a look at a 2-dimensional t-SNE-projection of the z-point distribution. The plots below have been produced with different settings for early exaggeration and perplexity parameters. The first plot resulted from standard parameter values for sklearn’s t-SNE variant.

tsne = TSNE(n_components=2, early_exaggeration=12, perplexity=30, n_iter=1000)

Other plots were produced by the following setting:

tsne = TSNE(n_components=2, early_exaggeration=16, perplexity=10, n_iter=1000)

Below you find some plots of a t-SNE-analysis for different numbers and different adjusted parameters for the resulting scatter plot. The number of statistically chosen z-point varies between 20,000 and 140,000.

Number of statistical z-points: 20,000 (non-standard t-SNE-parameters)

Actually we see some indication of clustering, though it is not very pronounced. The clusters in the projection are not separated by clear and broad gaps. Of course a 2-dimensional projection can not completely visualize the original separations in a 256-dim space. However, we get the impression that clusters are located rather close to each other. Remember: We already know that almost all points are locates in a multidimensional sphere shell between 12 < R < 24. And more than 50% between 14 ≤ R ≤ 19.

However, how the actual distribution of meaningful z-points (in the sense of a recognizable face reconstruction) really looks like cannot be deduced from the above t-SNE analysis. The concentration of the z-points may still be one which follows thin and maybe curved filaments in some directions of the multidimensional latent space on relatively small or various scales. We shall get a much clearer picture of the fragmentation of the z-point distribution in an AE’s latent space in the next post of this series.

Number of statistical z-points: 80,000

For the higher number of selected z-points the room between some concentration centers appears to be filled in the projection. But remember: This may only be due to projection effects in the presently chosen coordinate system. Another calculation with the above non-standard data for perplexity and early_exaggeration gives us:

Number of statistical z-points: 140,000

Note that some islands appear. Obviously, there is at least some clustering going on. However, due to projection effects we cannot deduce much for the real structure of the point distribution between possible clusters. Even the clustering itself could appear due to overlapping two or more broader filaments along a projection line.

Whether correlations would get more pronounced and therefore could also be better handled by t-SNE in a rotated coordinate system based on a PCA-analysis remains to be seen. The next post will give an answer.

At least we have got a clear impression about the radial distribution of the z-points. And thereby gathered some data which we can compare to corresponding results of a VAE.

Total loss of a VAE with a tiny KL-loss for CelebA data

Our test VAE is parameterized to create only a very small KL-loss contribution to the total loss. With the Python classes we have developed in the course of this post series we can control the ratio between the KL-loss and a standard reconstruction loss as e.g. BCE (binary-crossentropy) by a parameter “fact“.

For BCE

fact = 1.0e-5

is a very small value. For a working VAE we would normally choose something like fact=5 (see post XI).

A value like 1.0e-5 ensures a KL loss around 0.0178 compared to a reconstruction loss of 4550, which gives us a ratio below 4.e-6. Now, what is a VAE going to do, when the KL-loss is so small?

For the total loss the last epochs produced the following values:

AE total loss on Celeb A after 24 epochs for a step size of 0.0005: 4,553

Output of the last 6 of 24 epochs.

Epoch 1/6
1329/1329 [==============================] - 120s 90ms/step - total_loss: 4557.1694 - reco_loss: 4557.1523 - kl_loss: 0.0179
Epoch 2/6
1329/1329 [==============================] - 120s 90ms/step - total_loss: 4556.9111 - reco_loss: 4556.8940 - kl_loss: 0.0179
Epoch 3/6
1329/1329 [==============================] - 120s 90ms/step - total_loss: 4556.6626 - reco_loss: 4556.6450 - kl_loss: 0.0179
Epoch 4/6
1329/1329 [==============================] - 120s 90ms/step - total_loss: 4556.3862 - reco_loss: 4556.3682 - kl_loss: 0.0179
Epoch 5/6
1329/1329 [==============================] - 120s 90ms/step - total_loss: 4555.9595 - reco_loss: 4555.9395 - kl_loss: 0.0179
Epoch 6/6
1329/1329 [==============================] - 118s 89ms/step - total_loss: 4555.6641 - reco_loss: 4555.6426 - kl_loss: 0.0178

This is not too far away from the value of our AE. Other training runs confirmed this result. On four different runs the total loss value came to lie between

VAE total loss on Celeb A after 24 epochs: 4553 ≤ loss ≤ 4555 .

VAE with tiny KL-loss – z-point density distribution vs. radius

Below you find the plot for the variation of the number density of z-points vs. radius for our VAE:

Again, we get a maximum close to R = rad = 16. The maximum value lies a bit below the one found for a KL-loss-free AE. But all in all the form and width of the distribution of the VAE are very comparable to that of our test AE.

Can this result be reproduced?
Unfortunately not at a 100% of test runs performed. There are two main reasons:

  1. Firstly, we can not be sure that a second minimum does not exist for a distribution of points at bigger radii. This may be the case both for the AE and the VAE!
  2. Secondly, we have a major factor of statistical fluctuation in our game:
    The epsilon value which scales the logvar-contribution to the loss in the sampling layer of the Encoder may in very seldom cases abruptly jump to an unreasonable high value. A Gaussian covers extreme values, although the chances to produce such a value are pretty small. and a Gaussian is invilved in the calculation of z-points by our VAE.

Remember that the z-point coordinates are calculated via the the mu and logvar tensors according to

 
z = mu + B.exp(log_var / 2.) * epsilon

See Variational Autoencoder with Tensorflow 2.8 – VIII – TF 2 GradientTape(), KL loss and metrics for respective code elements of the Encoder.

So, a lot depends on epsilon which is calculated as a statistically fluctuating quantity, namely as

epsilon = B.random_normal(shape=B.shape(mu), mean=0., stddev=1.)

Is there a chance that the training process may sometimes drive the system to another corner of the weight-loss configuration space due to abrupt fluctuations? With the result for the z-point distribution vs. radius that it may significantly deviate from a distribution around R = 16? I think: Yes, this is possible!

From some other training runs I actually have an indication that there is a second minimum of the cost hyperplane with similar properties for higher average radius-values, namely for a distribution with an average radius at R ≈ 19.75. I got there after changing the initialization of the weights a bit.

Another indication that the cost surface has a relative rough structure and that extreme fluctuations of epsilon and a resulting gradient-fluctuation can drive the position of the network in the weight configuration space to some strange corners. The weight values there can result in different z-point distributions at higher average radii. This actually happened during yet another training run: At epoch 22 the Adam optimizer suddenly directed the whole system to weight values resulting in a maximum of the density distribution at R = 66 ! This appeared as totally crazy. At the same time the KL-loss also jumped to a much higher value.

When I afterward repeated the run from epoch 18 this did not happen again. Therefore, a statistical fluctuation must have been the reason for the described event. Such an erratic behavior can only be explained by sudden and extreme changes of z-point data enforcing a substantial change in size and direction of the loss gradient. And epsilon is a plausible candidate for this!

So far I had nothing in our Python classes which would limit the statistical variation of epsilon. The effects seen spoke for a code change such that we do not allow for extreme epsilon-values. I set limits in the respective part of the code for the sampling layer and its lambda function

        # The following function will be used by an eventual Lambda layer of the Encoder 
        def z_point_sampling(args):
            '''
            A point in the latent space is calculated statistically 
            around an optimized mu for each sample 
            '''
            mu, log_var = args # Note: These are 1D tensors !
            epsilon = B.random_normal(shape=B.shape(mu), mean=0., stddev=1.) 
            if abs(epsilon) >= 5: 
                epsilon *= 5. / abs(epsilon)       
            return mu + B.exp(log_var / 2.) * epsilon * self.enc_v_eps_factor

This stabilized everything. But even without these limitations on average three out of 4 runs which I performed for the VAE ran into a cost minimum which was associated with a pronounced maximum of the z-point-distribution around R ≈ 16. Below you see the plot for the fourth run:

So, there is some chance that the degrees of freedom associated with the logvar-layer and the statistical variation for epsilon may drive a VAE into other local minima or weight parameter ranges which do not lead to a z-point distribution around R = 16. But after the limitation of epsilon fluctuations all training runs found a loss minimum similar to the one of our simple AE – in the sense that it creates a z-point density distribution around R ≈ 16.

VAE with tiny KL-loss: Inertia and clustering of the CelebA data?

Our VAE gives the following variation of the inertia vs. the number of assumed clusters:

This also looks pretty similar to one of the plots shown for our AE above.

t-SNE for our VAE with a tiny KL loss

Below you find t-SNE plots for 20,000, 80,000 and 140,000 images:

Number of statistical z-points: 20,000 (non-standard t-SNE-parameters)

This is quite similar to the related image for the AE. You just have to rotate it.

Number of statistical z-points: 80,000

Number of statistical z-points: 140,000

All in all we get very similar indications as from our AE that some clustering is going on.

VAE with tiny KL-loss: Should its logvar values become tiny, too?

Besides reproducing a similar z-point distribution with respect to radius values, is there another indication that a VAE behaves similar to an AE? What would be a clear sign that the similarity really exists on a deeper level of the layers and their weights?

The z-vector is calculated from the mu and logvar-vectors by:

z = mu + exp(logvar/2)*epsilon

with epsilon coming from a normal distribution. Please note that we are talking about vectors of size z_dim=256 per image.

If a VAE with a tiny KL-loss really becomes similar to an AE it should define and set its z-points basically by using mu-values, only, and not by logvar-values. I.e. the VAE should become intelligent enough to ignore the degrees of freedom associated with the logvar-layer. Meaning that the z-point coordinates of a VAE with a very small Kl-loss should in the end be almost identical to the mu-component-values.

Ok, but to me it was not self-evident that a VAE during its training would learn

  • to produce significant mu-related weight-values, only,
  • and to keep the weight values for the connections to the logvar-layer so small that the logvar-impact on the z-space position gets negligible.

Before we speculate about reasons: Is there any evidence for a negligible logvar-contribution to the z-point coordinates or, equivalently, to the respective vector components?

A VAE with tiny KL-loss produces tiny logvar values …

To get some quantitative data on the logvar impact the following steps are appropriate:

  1. Get the size and algebraic sign of the logvar-values. Negative values logvar < -3 would be optimal.
  2. Measure the deviation between the mu- and z_points vector components. There should only be a few components which show significant values &br; abs(mu – z) > 0.05
  3. Compare the the radius-value determined by z-components vs. the radius values derived from mu-components, only, and measure the absolute and relative deviations. The relative deviation should be very small on average.

Some values of logvar, (z – mu), z-radii and z-radius-deviations for a VAE with small KL-loss

Regarding the maximum value of the logvar’s vector-components I found

3.4 ≥ max(logvar) ≥ -3.2. # for 1 up to 3 components out of a total 45.52 million components

The first value may appear to be big for a component. But typically there are only 2 (!) out of 170,000 x 256 = 43.52 million vector components in an interval of [-3, 5]. On the component level I found the following minimum, maximum and average-values

Maximum value for logvar:  -2.0205276
Minimum value for logvar:  -24.660698
Average value for logvar:  -13.682616

The average value of logvar is pretty pleasing: Such big negative values indeed render the logvar-impact on the position of our z-points negligible. So we should only find very small deviations of the mu-components from the z-point components. And, actually, the maximum of the deviation between a z_point component and a mu component was delta_mu_z = 0.26:

Maximum (z_points - mu) = delta_mu_z = 0.26  # on the component level 

There were only 5 out of the 45.52 million components which showed an absolute deviation in the interval

0.05 < abs(delta_mu_z) < 2. 

The rest was much, much smaller!

What about radius values? Here the situation looks, of course, even better:

max radius defined by z  :  33.10274
min radius defined by z  :  6.4961233
max radius defined by mu :  33.0972
min radius defined by mu :  6.494558

avg_z:      16.283989  
avg_mu:     16.283972

max absolute difference :  0.018045425 
avg absolute difference :  0.00094899256

max relative difference  :   0.00072147313
avg relative difference  :   6.1240215e-05

As expected, the relative deviations between z- and mu-based radius values became very small.

In another run (the one corresponding to the second density distribution curve above) I got the following values:

Maximum value for logvar:  3.3761873
Minimum value for logvar:  -22.777826
Average value for logvar:  -13.4265175

max radius z :  35.51387
min radius z :  7.209168
max radius mu :  35.515926
min radius mu :  7.2086616

avg_z:  17.37412
avg_mu: 17.374104

max delta rad relative :   0.012512478
avg delta rad relative :   6.5660715e-05

This tells us that the z-point distributions may vary a bit in their width, their precise center and average values. But, overall they appear to be similar. Especially with respect to a relative negligible contribution of logvar-terms to the z-point position. The relative impact of logvar on the radius value of a z-point is of the order 6.e-5, only.

All the above data confirm that a trained VAE with a very small KL-loss primarily uses mu-values to set the position of its z-points. During training the VAE moves along a path to an overall minimum on the loss hyperplane which leads to an area with weights that produce negligible logvar values.

Explanation of the overall similarity of a VAE with tiny KL-loss to an AE

o far we can summarize: Under normal conditions the VAE’s behavior is pretty close to that of a similar AE. The VAE produces only small logvar values. z-point coordinates are extremely close to just the mu-coordinates.

Can we find a plausible reason for this result? Looking at the cost-hyperplane with respect to the Encoder weights helps:

The cost surface of a VAE spans across a space of many more weight parameters than a corresponding AE. The reason is that we have weights for the connection to the logvar-layer in addition to the weights for the mu-layer (or a single output layer as in a corresponding AE). But if we look at the corner of the weight-vector-space where the logvar-related values are pretty small, then we would at least find a local (if not global) loss minimum there for the same values of the mu-related weight parameters as in the corresponding AE (with mu replacing the z-output).

So our question reduces to the closely related question whether the old minimum of an AE remains at least a local one when we shift to a VAE – and this is indeed the case for the basic reason that the KL-contributions to the height of the cost-hyperplane are negligibly small everywhere (!) – even for higher logvar-related values.

This tells us that a gradient descent algorithm should indeed be able to find a cost minimum for very small values of logvar-related weights and for weight-values related to the mu-layer very close to the AE’s weight-values for direct connections to its output layer. And, of course, with all other weight parameter of the VAE-Decoder being close to the values of the weights of a corresponding AE. At least under the condition that all variable quantities really change smoothly during training.

Does a VAE with small KL-loss produce reasonable face images?

A last test to confirm that a VAE with a very small KL-loss operates as an comparable AE is a trial to create images with recognizable human faces from randomly chosen points in z-space. Such a trial should fail! I just show you three results – one for a normal distribution of the z-point components. And two for equidistant distribution of component values up to 3, 8 and 16:

z-point coordinates from normal distribution

z-point coordinates from equidistant distribution in [-2,2]

z-point coordinates from equidistant distribution in [-10,10]

This reminds us very much about the behavior of an AE. See: Autoencoders, latent space and the curse of high dimensionality – I.

The z-point distribution in latent space of a VAE with a very small KL-loss obviously is as complicated as that of an AE. Neighboring points of a z-point which leads to a good image produce chaotic images. The transition path from good z-points to other meaningful z-points is confined to a very small filament-like volume.

Conclusion

A trained VAE with only a tiny KL-loss contribution will under normal circumstances behave similar to an AE with a the same hidden (convolutional) layers. It may, however, be necessary to limit the statistical variation of the epsilon factor in the z-point calculation based on mu– and logvar-values.

The similarity is based on very small logvar-values after training. The VAE creates a z-point distribution which shows the same dependency on the radius as an AE. We see similar indications and patterns of clustering. And the VAE fails to produce human faces from random z-points in the latent space – as a comparable AE.

We have found a plausible reason for this similarity by comparing the minimum of the loss hyperplane in the weight-loss parameter space with a corresponding minimum in the weight-loss space of the VAE – at a position with small weights for the connection to the logvar layers.

The z-point density distribution shows a maximum at a radius between 16 and 17. The z-point distribution basically has a Gaussian form. In the next post we shall look a bit closer at these findings – and their origin in Gaussian distributions along the coordinate axes of the latent space. After an application of a PCA analysis we shall furthermore see that the z-point distribution in an AE’s latent vector space is indeed fragmented and shows filaments on certain length scales. A VAE with a tiny KL-loss will show the same fragmentation.

In further forthcoming post we shall afterward investigate the confining and at the same time blurring impact of the KL-loss on the latent space. Which will make it usable for creative purposes.

And let us all who praise freedom not forget:
The worst fascist, war criminal and killer living today is the Putler. He must be isolated at all levels, be denazified and sooner than later be imprisoned. Long live a free and democratic Ukraine!

 

KMeans as a classifier for the WIFI and MNIST datasets – V – cluster based classification of the MNIST dataset

In this series about KMeans

KMeans as a classifier for the WIFI and MNIST datasets – I – Cluster analysis of the WIFI example
KMeans as a classifier for the WIFI and MNIST datasets – II – PCA in combination with KMeans for the WIFI-example
KMeans as a classifier for the WIFI and MNIST datasets – III – KMeans as a classifier for the WIFI-example
KMeans as a classifier for the WIFI and MNIST datasets – IV – KMeans on PCA transformed data

we have so far studied the application of KMeans to the WIFI dataset of the UCI Irvine. We now apply the Kmeans clustering algorithm to the MNIST dataset – also in an extended form, namely as a classifier. The MNIST dataset – a collection of 28x28px images of handwritten numbers – has already been discussed in other sections of this blog and is well documented on the Internet. I, therefore, do not describe its basic properties in this post. A typical image of the collection is

Load MNIST – dimensionality of the feature space and scaling of the data

Due to the ease of use, I loaded the MNIST data samples via TF2 and the included Keras interface. Otherwise TF2 was not used for the following experiments. Instead the clustering algorithms were taken from “sklearn”.

Each MNIST image can be transformed into a one-dimensional array with dimension 784 (= 28 * 28). This means the MNIST feature space has a dimension of 784 – which is much more than the seven dimensions we dealt with when analyzing the WIFI data in the last post. All MNIST samples were shuffled for individual runs.

Scaling of MNIST data for clustering?

A good question is whether we should scale or normalize the sample data for clustering – and if so by what formula. I could not answer this question directly; instead I tested multiple methods. Previous experience with PCA and MNIST indicated that Sklearn’s “Normalizer” would be helpful, but I did not take this as granted.

A simple scaling method is to just divide the pixel values by 255. This brings all 784 data array elements of each image into the value range [0,1]. Note that this scaling does not change relative length differences of the sample vectors in the feature space. Neither does it shift or change the width of the data distribution around its mean value. Other methods would be to standardize the data or to normalize them, e.g. by using respective algorithms from Scikit-Learn. Using either method in combination with a cluster analysis corresponds to a theory about the cluster distribution in the feature space. Normalization would mean that we assume that the clusters do not so much depend on the vector length in the feature space but mainly on the angle of the sample vectors. We shall later see what kind of scaling helps when we classify the MNIST data based on clusters.

In a first approach we leave the data as they are, i.e. unscaled.

Parameters for clustering

All the following cluster calculations were done on 3 out of 8 available (hyperthreaded) CPU cores. For Kmeans and MiniBatchKMeans we used

n_init       = 100       # number of initial cluster configurations to test 
max_iter     = 100       # maximum number of iterations  
tol          = 1.e-4     # final deviation of subsequent results (= stop condition)  
random_state = 2         # a random state nmber for repeatable runs
mb_size      = 200       # size of minibatches (for MiniBatchKMeans) 

The number of clusters “num_clus” was defined individually for each run.

Analysis by KMeans? Too expensive …

A naive approach to perform an elbow analysis, as we did for the WIFI-data, would be to apply KMeans of Sklearn directly to the MNIST data. But a test run on the CPU shows that such an endeavor would cost too much time. With 3 CPU cores and only a very limited number of clusters and iterations

n_init   = 10      # only a few initial configurations
max_iter = 50 
tol      = 1.e-3  
num_clus = 25      # only a few clusters

a KMeans fit() run applied to 60,000 training samples [len(X_train) => 60,000]

kmeans.fit(X_train)

requires around 42 secs. For 200 clusters the cluster analysis requires around 214 secs. Doing an elbow analysis would therefore require many hours of computational time.
To overcome this problem I had to use MiniBatchKMeans. It is by factors > 80 faster.

Elbow analysis with the help of MiniBatchKmeans

When we use the following setting for MiniBatchKMeans

n_init = 50 # only a few initial configurations
max_iter = 100
tol = 1.e-4
mb_size = 200 

I could perform an elbow analysis for all cluster-numbers 1 < k <= 250 in less than 20 minutes. The following graphics shows the resulting intertia curve vs. cluster number:

The “elbow” is not very pronounced. But I would say that by using a cluster number around 200 we are on the safe side. By the way: The shape of the curve does not change very much when we apply Sklearn’s Normalizer to the MNIST data ahead of the cluster analysis.

Classifying unscaled data with the help of clusters

We now perform a prediction of our adapted cluster algorithm regarding the cluster membership for the training data and for k=225 clusters:

n_clu    = 225
mb_size  = 200
max_iter = 120
n_init   = 100
tol      = 1.e-4

Based on the resulting data we afterward apply the same type of algorithm which we used for the WIFI data to construct a “classifier” based on clusters and a respective predictor function (see the last post of this series).

The data distribution for the 10 different digits of the training set was:

class 0 :  5905
class 1 :  6721
class 2 :  6031
class 3 :  6082
class 4 :  5845
class 5 :  5412
class 6 :  5917
class 7 :  6266
class 8 :  5860
class 9 :  5961

How good is the cluster membership of a sample for a digit class defined?
Well, out of 225 clusters there were only around 15 for which I got an “error” above 40%, i.e. for which the relative fraction of data samples deviating from the dominant class of the cluster was above 40%. For the vast majority of clusters, however, samples of one specific digit class dominated the clusters members by more than 90%.

The resulting confusion matrix of our new “cluster classifier” for the (unscaled) MNIST data looks like

[[5695    4   37   21    7   57   51   15   15    3]
 [   0 6609   33   21   11    2   15   17    2   11]
 [  62   45 5523  120   14   10   27  107  116    7]
 [  11   43  114 5362   15  153    8   60  267   49]
 [   5   60   62    2 4752    3   59   63    5  834]
 [  54   18  103  777   25 4158  126    9  110   32]
 [  49   20   56    4    6   38 5736    0    8    0]
 [   5   57   96    2   86    1    0 5774    7  238]
 [  30   76  109  416   51  152   39   35 4864   88]
 [  25   20   37   84  706   14    6  381   46 4642]]

This confusion matrix comes at no surprise: The digits “4”, “5”, “8”, “9” are somewhat error prone. Actually, everybody familiar with MNIST images knows that sometimes “4”s and “9”s can be mixed up even by the human eye. The same is true for handwritten “5”s, “8”s and “3”s.

Another representation of the confusion matrix is:

The calculation for the matrix elements was done in a standard way – the sum over percentages in a row gives 100% (the slight deviation in the matrix is due to rounding). I.e. we look at erors of the type TN (True Negatives).

The confusion matrix for the remaining 10,000 test data samples is:

The relative errors we get for our classifier when applied to the train and test data is

rel_err_train = 0.115 ,
rel_err_test = 0.112

All for unscaled MNIST data. Taking into account the crudeness of the whole approach this is a rather convincing result. It proves that it is worth the effort to perform a cluster analysis on high dimensional data:

  • It provides a first impression whether the data are structured in the feature space such that we can find relatively good separable clusters with dominant members belonging to just one class.
  • It also shows that a cluster based classification for many datasets cannot reach accuracy levels of CNNs, but that it may still deliver good results. Without any supervised training …

The second point also proves that the distance of the data points to the various cluster centers contains valuable information about the class membership. So, a MLP or CNN based classification could be performed on transformed MNIST data, namely distance vectors of sample datapoints to the different cluster centers. This corresponds to a dimension reduction of the classification problem. Actually, in a different part of this blog, I have already shown that such an approach delivers accuracy values beyond 98%.

For MNIST we can say that the samples define a relatively well separable cluster structure in the feature space. The granularity required to resolve classes sufficiently well means a clsuter number of around 200 < k < 250. Then we get an accuracy close to 90% for cluster based classification.

t-SNE representation of the MNIST data

Can we somehow confirm this finding about a good cluster-class-relation independently? Well, in a limited way. The t-SNE algorithm, which can be used to “project” multidimensional data onto a 2-dimensional plane, respects the vicinity of vectors in the original feature space whilst deriving a 2-dim representation. So, a rather well structured t-SNE diagram is an indication of clustering in the feature space. And indeed for 10,000 randomly selected samples of the (shufffled) training data we get:

The colorization was done by classes, i.e. digits. We see a relatively good separation of major “clusters” with data points belonging to a specific class. But we also can identify multiple problem zones, where data points belonging to different classes are intermixed. This explains the confusion matrix. It also explains why we need so many fine-grained clusters to get a reasonable resolution regarding a reliable class-cluster-relation.

Classifying scaled and normalized MNIST data with the help of clusters

Can we improve the accuracy of our cluster based classification a bit? This would, e.g., require some transformation which leads to a better cluster separation. To see the effect of two different scalers I tried the “Normalizer” and then also the “StandardScaler” of Sklearn. Actually, they work in opposite direction regarding accuracy:

The “Normalizer” improves accuracy by more than 1.5%, while the “Standardizer” reduces it by almost the same amount.

I only discuss results for “Normalization” below. The confusion matrix for the training data becomes:

and for the test data:

The relative error for the test data is

Error for trainings data:
avg_err_train = 0.085 :: num_err_train = 5113
 
Error for test data:
avg_err_test = 0.083 :: num_err_test = 832

So, the relative accuracy is now around 91.5%.
The result depends a bit on the composition of the training and the test dataset after an initial shuffling. But the value remains consistently above 90%.

Data compression by Autoencoders and clustering

Just for interest I also had a look at a very different approach to invoke clustering:

I first applied a simple CNN-based AutoEncoder [AE] to compress the MNIST data into a 25-dimensional space and applied our clustering methods afterwards.

I shall not discuss the technology of autoenconders in this post. The only relevant point in our context is that an autoencoder provides an efficient non-linear way of data compression and dimensionality reduction. Among many other useful properties and abilities … . Note: I did not use a “Variational Autoencoder” which would have allowed for even better results. The loss function for the AE was a simple quadratic loss. The autoencoder was trained on 50,000 training samples and for 40 epochs.

A t-SNE based plot of the “clusters” for test data in the 25-dimensional space looks like:

We see that the separation of the data belonging to different classes is somewhat better than before. Therefore, we expect a slightly better classification based on clusters, too. Without any scaling we get the following confusion data:

[[5817    7   10    3    1   14   15    2   27    1]
 [   3 6726   29    2    0    1   10    5   12   10]
 [  49   35 5704   35   14    4   10   61   87    7]
 [   8   78   48 5580   22  148    2   40  111   29]
 [  47   27   18    0 4967    0   44   38    3  673]
 [  32   20   10  150    8 5039   73    4   43   28]
 [  31   11   23    2    2   47 5746    0   15    1]
 [   6   35   35    6   32    0    1 5977    7  163]
 [  17   67   22   86   16  217   24   22 5365   52]
 [  35   32   11   92  184   15    1  172   33 5406]]

Error averaged over (all) clusters :  6.74

The resulting relative error for the test data was:

avg_err_test = 0.0574 :: num_err_test = 574

With Normalization:

Error for test data:
avg_err_test = 0.054 :: num_err_test = 832

So, after performing the autoencoder training on normalized data we consistently get

an accuracy of around 94%.

This is not too much of a gain. But remember:
We performed a cluster analysis on a feature space with only 25 dimensions – which of course is much cheaper. However, we paid a prize, namely the Autoencoder training which lasted about 150 secs on my old Nvidia 960 GTX.

And note: Even with only 100 clusters we get above 92% on the AE-compressed data.

Conclusion

We have shown that using a non-supervised cluster analysis of the MNIST data with around 225 clusters allows for classifying images with an accuracy around 90.5%. In combination with an Autoencoder compression we even reaches values around 94%. This is comparable with other non-optimized standard algorithms aside of neural networks.

This means that the MNIST data samples are organized in a well separable cluster structure in their feature space. A test run with normalized data showed that the clusters (and their centers) differ mostly by their direction relative to the origin of the feature space and not so much by their distance from the origin. With a relatively fine grained resolution we could establish a simple cluster-class-relation which allowed for cluster based classification.

The accuracy is, of course, below the values reachable with optimized MLPS (98%) and CNNs (above 99%). But, clustering is a fast, reliable and non-supervised method. In addition in combination with t-SNE we can create plots which can easily be understood by the customers. So, even for more complex data I would always recommend to try a cluster based classification approach if you need to provide plots and quick results. Sometimes the accuracy may even be sufficient for your customer’s purposes.

KMeans as a classifier for the WIFI and MNIST datasets – IV – KMeans on PCA transformed data

In the last posts of this series

KMeans as a classifier for the WIFI and MNIST datasets – I – Cluster analysis of the WIFI example
KMeans as a classifier for the WIFI and MNIST datasets – II – PCA in combination with KMeans for the WIFI-example
KMeans as a classifier for the WIFI and MNIST datasets – III – KMeans as a classifier for the WIFI-example

we applied the KMeans algorithm to perform a cluster analysis of the WIFI dataset of the UCI Irvine. The results gave us insights into the spatial grouping and the separability of the data samples in their 7-dimensional feature space. An additional PCA analysis helped to understand why projections of the data into some selected 2-dimensional sub-spaces of the feature space revealed the four or five dominant clusters very well. In the third post I discussed a simple method to transform KMeans into a classifier. In the WIFI case a set of 9 to 11 clusters provided a good resolution of the data distribution and we reached a convincing classifier accuracy.

What we have not done, yet, is to transform and project the WIFI data into the coordinate system of the most important main components and afterward apply clustering by the help of KMeans. We know already that three primary components fit the data very well and give us around 90% of the “explained variance“. See the second post for these basic PCA results. We, therefore, expect comparably accurate prediction results of a cluster classifier for the PCA transformed data as the accuracy values given in the last post. For 500 test samples after a KMeans fit of 1500 training samples in the original feature space we found a prediction accuracy of around 98%.

In this post we first perform a PCA analysis for three primary components of the WIFI data distribution and then transform the vectors of 1500 randomly selected training samples to the 3-dimensional main component space. Then we apply KMeans onto the data in the reduced vector space and establish a classifier predictor based on the methods described in the last article. Eventually, we check the accuracy and display the resulting confusion matrix for the 500 test samples.

As a side-step for readers who look for real world use cases regarding signals I want to mention an article in “Nature”, which I found today via a newspaper podcast. There neural firing rates of a brain region, i.e. some very special signals, were used to enable an ALS patient to select letters from presented sequences and form statements – by his “thoughts”. This looks like an environment where Machine Learning really could contribute more in the future.

KMeans as a classifier on the PCA transformed WIFI data

Below I give you the results for the WIFI data transformed and projected to the most important three primary components and 11 clusters:

Results for 1500 training samples

  
Confusion matrix for training data - 11 clusters, 3 PCA components 
A confusion matrix for the classes according to the clustering
[[374   0   1   0]
 [  0 362  13   0]
 [  4   7 359   5]
 [  1   0   0 374]]

Number of wrongly predicted train samples:  31  :: avg_err =  0.020

So, just from counting wrongly classified examples the average error is measured to be around 2% and the relative accuracy is something like 98%.

And for the test data I got:

Number of wrongly predicted test samples:  6  :: avg_err =  0.012

This gives us the following confusion matrix:

This actually proves that our assumption about combining a PCA transformation with a KMeans classifier was correct. The reduction of the dimensionality of the problem did not affect the prediction accuracy very much.

Just for completeness the data for only 2 primary components:

The accuracy of around 97% is still convincing. The reason is that the two most important primary components already deliver around 85% of the “explained variance”.

Why is the Wifi-example not so boring as one may think?

A reader wrote me that he finds the WIFI example too simple and boring. OK, but … The principles and methods remain the same when more complex data are analyzed for clusters. Especially in the case of binary classification. But are there interesting real world use cases for other types of signals? Oh, yes. I just want to refer to an interesting example which I read about this morning.

The WIFI example works with samples which describe 7 signals. Now, imagine that such signals come from a sensor implant measuring electric potentials of a human brain and that we do not analyze for the location of rooms but for the selection of letters by “Yes/No” decision-“imaginations” – made by a human who was trained via frequency based audio-feedback for the brain regions covered by the implants. Science fiction? No, reality. And of huge help for ALS patients. See

Chaudhary, U., Vlachos, I., Zimmermann, J.B. et al. Spelling interface using intracortical signals in a completely locked-in patient enabled via auditory neurofeedback training. Nat Commun 13, 1236 (2022). https://doi.org/10.1038/s41467-022-28859-8

and
https://www.nature.com/articles/s41467-022-28859-8

There, signals were measured from two implant arrays with 64 electrodes. OK, these are somewhat more signals than just 7. But if I understood the text correctly not all channels were used or useful. Just a few. Reminds us of PCA? In addition the time structure of the signal (firing rates) are important – but these are just different signal characteristics. And we have different labels. But, at least in principle, we speak of nothing else than pattern detection based on signal values.

I only had a brief look into the supplementary data of the experiment (an Excel file) and I am not at all familiar with the the experimental setup – but from reading my impression was that just threshold values for the firing rate of some channels were used to distinguish “Yes” from “No”. Maybe we could do a bit better with AI (PCA and classifying according to multidimensional pattern analysis)? Does this look like an interesting use case?

Conclusion

In the case of the WIFI example KMeans can be used as an efficient classifier for samples in a feature space which describes characteristics of multiple signal sources. We have seen that the basic concept also works when we apply KMeans after a PCA based transformation to the most important primary components.

The question is: Does this work equally well for other data sets? The answer depends upon the accuracy by which clusters reside completely within regions of the feature space filled by samples of a specific label.
A data set whose samples show grouping in a multidimensional feature space and appear relatively well separable by their labels is the MNIST data set. In the next post of this series we shall therefore try and apply a clustering algorithm to the MNIST data ensemble.

Stay tuned …

Ceterum censeo: The worst fascist today who must be isolated and denazified is the Putler.