How to Access the Latent Dimension of an Autoencoder

I'm quite new to machine learning. I've only taken one ML class, but I'm hoping to expand on what we did there, which was mainly different ways to cluster data. I've mainly tried the built-in clustering methods so far.

I came across this article: https://www3.ntu.edu.sg/home/EXDJiang/spl20.pdf

And I wanted to try clustering with a VAE. From what I understand from the paper, a clustering method is going to be applied to the latent dimensions? I've looked at the VAE examples in MATLAB, but they all deal with image data and recreating images. Plus, the latent dimensions weren't explicitly accessed there. How can you access the latent dimensions?

Or am I going in the wrong direction with this idea? Are there more suited neural architectures for clustering?

Thank you in advance, and sorry if this is a vague and terrible question.

[Umar]

Hi Eunice,

It took me a while to research answers to your questions because Mathworks is a highly respectable platform. Many OPs who have posted their questions in the past had strong background in electrical, computer, biomedical, data science , RF etc and some of them asked intuitive and intellectual questions regarding their problems. My goal has always been staying humble and embrace humility when answering questions.

Based on your background in machine learning and interest in exploring clustering methods beyond the basics, it's great that you're looking to delve into more advanced techniques like clustering with a Variational Autoencoder (VAE). The paper you referenced indeed discusses applying a clustering method to the latent dimensions generated by a VAE, which can be a powerful approach for unsupervised learning tasks. Now, let me provide solutions to your questions.

Question#1: How can you access the latent dimensions?

To access latent dimensions in a VAE model, you typically need to modify the architecture or code to extract and analyze these representations directly. This may involve accessing the encoder part of the VAE model to obtain the mean and variance vectors that parameterize the latent space. I will try to demonstrate it with a simple example because the goal is to understand how to access the latent dimensions in a VAE and apply a clustering method to these dimensions. I will create generic data for visualization, define variables, and plot the data without using functions. So, in the example code snippet, I started by creating random data points with 2 dimensions using the randn function to simulate generic data for visualization. Then, set the number of clusters (num_clusters) to 3 and the maximum number of iterations for clustering (max_iterations) to 100, afterwards apply the K-Means clustering algorithm to the generated data (data) with the specified number of clusters and maximum iterations which returns the cluster indices (idx) and the centroid locations (C). Finally, plot the clustered data using the gscatter function to visualize the clusters based on the latent dimensions. The plot includes labels for the dimensions, clusters, and a legend for cluster identification.

%Snippet Code Example

% Generate generic data for visualization

data = randn(100, 2); % Generating 100 data points with 2 dimensions

% Define variables for clustering

num_clusters = 3; % Number of clusters

max_iterations = 100; % Maximum number of iterations for clustering

% Perform clustering on the latent dimensions

[idx, C] = kmeans(data, num_clusters, 'MaxIter', max_iterations);

% Plot the clustered data

figure;

gscatter(data(:,1), data(:,2), idx);

title('Clustering with VAE Latent Dimensions');

xlabel('Dimension 1');

ylabel('Dimension 2');

legend('Cluster 1', 'Cluster 2', 'Cluster 3');

Please see attached plot along with snippet code.

By following these steps, you can access the latent dimensions in a VAE, apply clustering techniques, and visualize the clustered data in MATLAB without using additional functions.

Question#2: Am I going in the wrong direction with this idea? Are there more suited neural architectures for clustering?

Well, it ultimately depends on your specific data and objectives. In my opinion, VAEs are well-suited for tasks involving generative modeling and dimensionality reduction. However, other neural network architectures like deep autoencoders, self-organizing maps, or deep belief networks may also be effective for clustering tasks.

By experimenting with different methods and staying curious about new developments in the field, you can continue to expand your knowledge and skills in machine learning. Please let me know if you have any further questions.

Good luck with your exploration!

[Eunice Chieng]

Thank you for your response! I'm aware of how to cluster datapoints; my main concern is actually on how to access the data that I am going to cluster from an autoencoder since the sample projects that involve autoencoders don't really seem to show this. That's the part I'm asking about.

[Umar]

Hi Eunice,

You have to use the encoder portion of the trained model to obtain latent representations, which you can utilized for clustering. Let me illustrate it with example.

% Load and preprocess your data

data = load('your_dataset.mat');

% Define and train your autoencoder

autoenc = trainAutoencoder(data, hiddenSize);

% Use the encoder to extract features

encodedFeatures = encode(autoenc, data);

% Perform clustering on the encoded features

[idx, C] = kmeans(encodedFeatures, k);

If you have any further questions or need additional assistance, feel free to ask!

Answer 1

[Kaustab Pal]

Hi @Eunice Chieng,

To retrieve the latent vector from the autoencoder, you can use the encoder network with your input data in the "predict" function. Below is a sample code to illustrate this process:

numLatentChannels = 32;
imageSize = [28 28 1];

% ENCODER PART %
layersE = [
    imageInputLayer(imageSize,Normalization="none")
    convolution2dLayer(3,32,Padding="same",Stride=2)
    reluLayer
    convolution2dLayer(3,64,Padding="same",Stride=2)
    reluLayer
    fullyConnectedLayer(numLatentChannels)];


netE = dlnetwork(layersE);

IMG = randn(28,28,1,1); % S, S, C, B
IMG = dlarray(IMG,"SSCB");
latent_vec = predict(netE, IMG) % IMG needs to be a dlarray
% latent_vec is a 32 x 1 dimensional vector

I hope you find this helpful!