update

Author: mhjensen

doc/pub/week13/html/week13-bs.html

@@ -124,6 +124,8 @@
                None,
                'code-examples-using-keras'),
               ('Code in PyTorch for VAEs', 2, None, 'code-in-pytorch-for-vaes'),
+              ('Pros and cons about VAEs', 2, None, 'pros-and-cons-about-vaes'),
+              ('Pros and cons about VAEs', 2, None, 'pros-and-cons-about-vaes'),
               ('Diffusion models, basics', 2, None, 'diffusion-models-basics'),
               ('Problems with probabilistic models',
                2,
@@ -252,6 +254,8 @@
      <!-- navigation toc: --> <li><a href="#computing-the-gradients" style="font-size: 80%;">Computing the gradients</a></li>
      <!-- navigation toc: --> <li><a href="#code-examples-using-keras" style="font-size: 80%;">Code examples using Keras</a></li>
      <!-- navigation toc: --> <li><a href="#code-in-pytorch-for-vaes" style="font-size: 80%;">Code in PyTorch for VAEs</a></li>
+     <!-- navigation toc: --> <li><a href="#pros-and-cons-about-vaes" style="font-size: 80%;">Pros and cons about VAEs</a></li>
+     <!-- navigation toc: --> <li><a href="#pros-and-cons-about-vaes" style="font-size: 80%;">Pros and cons about VAEs</a></li>
      <!-- navigation toc: --> <li><a href="#diffusion-models-basics" style="font-size: 80%;">Diffusion models, basics</a></li>
      <!-- navigation toc: --> <li><a href="#problems-with-probabilistic-models" style="font-size: 80%;">Problems with probabilistic models</a></li>
      <!-- navigation toc: --> <li><a href="#diffusion-models" style="font-size: 80%;">Diffusion models</a></li>
@@ -1293,6 +1297,26 @@ <h2 id="code-in-pytorch-for-vaes" class="anchor">Code in PyTorch for VAEs </h2>
 </div>
 
 
+<!-- !split -->
+<h2 id="pros-and-cons-about-vaes" class="anchor">Pros and cons about VAEs </h2>
+
+<ol>
+<li> Generative Capability: VAEs are powerful generative models that can produce new data similar to the training data.</li>
+<li> Latent Space Structure: The latent space learned by VAEs is continuous, structured, and often interpretable, which makes it useful for interpolation, clustering, and downstream tasks.</li>
+<li> Theoretical Foundation: VAEs are grounded in probability theory and Bayesian inference, offering a principled approach to modeling uncertainty.</li>
+<li> Efficient Training: Unlike GANs (discussed later), VAEs typically train more stably and don&#8217;t suffer from mode collapse.</li>
+<li> Regularization: The KL divergence term in the loss encourages the latent space to follow a standard normal distribution, aiding generalization.
+<!-- o Semi-supervised Learning: VAEs can be adapted for semi-supervised tasks due to their probabilistic framework. --></li>
+</ol>
+<!-- !split -->
+<h2 id="pros-and-cons-about-vaes" class="anchor">Pros and cons about VAEs </h2>
+<ol>
+<li> Blurry Reconstructions: VAEs often produce blurrier outputs compared to GANs due to the use of a probabilistic decoder and the Gaussian likelihood assumption.</li>
+<li> Posterior Collapse: The model might ignore the latent variables entirely (especially with powerful decoders), leading to <b>posterior collapse</b>.</li>
+<li> Trade-off in Loss Function: Balancing the reconstruction term and the KL divergence can be tricky, and tuning this balance is crucial.</li>
+<li> Assumed Distribution Limitations: The prior (usually Gaussian) and the approximated posterior may not capture the true data distribution well.</li>
+<li> Less Sharp Compared to GANs: In image generation tasks, the samples are usually less sharp and detailed than those generated by GANs.</li>
+</ol>
 <!-- !split -->
 <h2 id="diffusion-models-basics" class="anchor">Diffusion models, basics </h2>
 

doc/pub/week13/html/week13-reveal.html

@@ -1276,6 +1276,30 @@ <h2 id="code-in-pytorch-for-vaes">Code in PyTorch for VAEs </h2>
 </div>
 </section>
 
+<section>
+<h2 id="pros-and-cons-about-vaes">Pros and cons about VAEs </h2>
+
+<ol>
+<p><li> Generative Capability: VAEs are powerful generative models that can produce new data similar to the training data.</li>
+<p><li> Latent Space Structure: The latent space learned by VAEs is continuous, structured, and often interpretable, which makes it useful for interpolation, clustering, and downstream tasks.</li>
+<p><li> Theoretical Foundation: VAEs are grounded in probability theory and Bayesian inference, offering a principled approach to modeling uncertainty.</li>
+<p><li> Efficient Training: Unlike GANs (discussed later), VAEs typically train more stably and don&#8217;t suffer from mode collapse.</li>
+<p><li> Regularization: The KL divergence term in the loss encourages the latent space to follow a standard normal distribution, aiding generalization.
+<!-- o Semi-supervised Learning: VAEs can be adapted for semi-supervised tasks due to their probabilistic framework. --></li>
+</ol>
+</section>
+
+<section>
+<h2 id="pros-and-cons-about-vaes">Pros and cons about VAEs </h2>
+<ol>
+<p><li> Blurry Reconstructions: VAEs often produce blurrier outputs compared to GANs due to the use of a probabilistic decoder and the Gaussian likelihood assumption.</li>
+<p><li> Posterior Collapse: The model might ignore the latent variables entirely (especially with powerful decoders), leading to <b>posterior collapse</b>.</li>
+<p><li> Trade-off in Loss Function: Balancing the reconstruction term and the KL divergence can be tricky, and tuning this balance is crucial.</li>
+<p><li> Assumed Distribution Limitations: The prior (usually Gaussian) and the approximated posterior may not capture the true data distribution well.</li>
+<p><li> Less Sharp Compared to GANs: In image generation tasks, the samples are usually less sharp and detailed than those generated by GANs.</li>
+</ol>
+</section>
+
 <section>
 <h2 id="diffusion-models-basics">Diffusion models, basics </h2>
 

doc/pub/week13/html/week13-solarized.html

@@ -151,6 +151,8 @@
                None,
                'code-examples-using-keras'),
               ('Code in PyTorch for VAEs', 2, None, 'code-in-pytorch-for-vaes'),
+              ('Pros and cons about VAEs', 2, None, 'pros-and-cons-about-vaes'),
+              ('Pros and cons about VAEs', 2, None, 'pros-and-cons-about-vaes'),
               ('Diffusion models, basics', 2, None, 'diffusion-models-basics'),
               ('Problems with probabilistic models',
                2,
@@ -1226,6 +1228,26 @@ <h2 id="code-in-pytorch-for-vaes">Code in PyTorch for VAEs </h2>
 </div>
 
 
+<!-- !split --><br><br><br><br><br><br><br><br><br><br>
+<h2 id="pros-and-cons-about-vaes">Pros and cons about VAEs </h2>
+
+<ol>
+<li> Generative Capability: VAEs are powerful generative models that can produce new data similar to the training data.</li>
+<li> Latent Space Structure: The latent space learned by VAEs is continuous, structured, and often interpretable, which makes it useful for interpolation, clustering, and downstream tasks.</li>
+<li> Theoretical Foundation: VAEs are grounded in probability theory and Bayesian inference, offering a principled approach to modeling uncertainty.</li>
+<li> Efficient Training: Unlike GANs (discussed later), VAEs typically train more stably and don&#8217;t suffer from mode collapse.</li>
+<li> Regularization: The KL divergence term in the loss encourages the latent space to follow a standard normal distribution, aiding generalization.
+<!-- o Semi-supervised Learning: VAEs can be adapted for semi-supervised tasks due to their probabilistic framework. --></li>
+</ol>
+<!-- !split --><br><br><br><br><br><br><br><br><br><br>
+<h2 id="pros-and-cons-about-vaes">Pros and cons about VAEs </h2>
+<ol>
+<li> Blurry Reconstructions: VAEs often produce blurrier outputs compared to GANs due to the use of a probabilistic decoder and the Gaussian likelihood assumption.</li>
+<li> Posterior Collapse: The model might ignore the latent variables entirely (especially with powerful decoders), leading to <b>posterior collapse</b>.</li>
+<li> Trade-off in Loss Function: Balancing the reconstruction term and the KL divergence can be tricky, and tuning this balance is crucial.</li>
+<li> Assumed Distribution Limitations: The prior (usually Gaussian) and the approximated posterior may not capture the true data distribution well.</li>
+<li> Less Sharp Compared to GANs: In image generation tasks, the samples are usually less sharp and detailed than those generated by GANs.</li>
+</ol>
 <!-- !split --><br><br><br><br><br><br><br><br><br><br>
 <h2 id="diffusion-models-basics">Diffusion models, basics </h2>
 

doc/pub/week13/html/week13.html

@@ -228,6 +228,8 @@
                None,
                'code-examples-using-keras'),
               ('Code in PyTorch for VAEs', 2, None, 'code-in-pytorch-for-vaes'),
+              ('Pros and cons about VAEs', 2, None, 'pros-and-cons-about-vaes'),
+              ('Pros and cons about VAEs', 2, None, 'pros-and-cons-about-vaes'),
               ('Diffusion models, basics', 2, None, 'diffusion-models-basics'),
               ('Problems with probabilistic models',
                2,
@@ -1303,6 +1305,26 @@ <h2 id="code-in-pytorch-for-vaes">Code in PyTorch for VAEs </h2>
 </div>
 
 
+<!-- !split --><br><br><br><br><br><br><br><br><br><br>
+<h2 id="pros-and-cons-about-vaes">Pros and cons about VAEs </h2>
+
+<ol>
+<li> Generative Capability: VAEs are powerful generative models that can produce new data similar to the training data.</li>
+<li> Latent Space Structure: The latent space learned by VAEs is continuous, structured, and often interpretable, which makes it useful for interpolation, clustering, and downstream tasks.</li>
+<li> Theoretical Foundation: VAEs are grounded in probability theory and Bayesian inference, offering a principled approach to modeling uncertainty.</li>
+<li> Efficient Training: Unlike GANs (discussed later), VAEs typically train more stably and don&#8217;t suffer from mode collapse.</li>
+<li> Regularization: The KL divergence term in the loss encourages the latent space to follow a standard normal distribution, aiding generalization.
+<!-- o Semi-supervised Learning: VAEs can be adapted for semi-supervised tasks due to their probabilistic framework. --></li>
+</ol>
+<!-- !split --><br><br><br><br><br><br><br><br><br><br>
+<h2 id="pros-and-cons-about-vaes">Pros and cons about VAEs </h2>
+<ol>
+<li> Blurry Reconstructions: VAEs often produce blurrier outputs compared to GANs due to the use of a probabilistic decoder and the Gaussian likelihood assumption.</li>
+<li> Posterior Collapse: The model might ignore the latent variables entirely (especially with powerful decoders), leading to <b>posterior collapse</b>.</li>
+<li> Trade-off in Loss Function: Balancing the reconstruction term and the KL divergence can be tricky, and tuning this balance is crucial.</li>
+<li> Assumed Distribution Limitations: The prior (usually Gaussian) and the approximated posterior may not capture the true data distribution well.</li>
+<li> Less Sharp Compared to GANs: In image generation tasks, the samples are usually less sharp and detailed than those generated by GANs.</li>
+</ol>
 <!-- !split --><br><br><br><br><br><br><br><br><br><br>
 <h2 id="diffusion-models-basics">Diffusion models, basics </h2>