- Generative models are a branch of unsupervised learning because the model aims to capture and model the data without using explicit target labels.
- Generative models that use Supervised learning like cGAN : “Conditional generative adversarial networks” learn and generate labeled data and operate based on specific classes or attributes. Unsupervised generative models learn the data without class labels and generate samples without any conditional information. Examples : GAN and VAE :Variational autoencoders
<aside>
💡 “What I cannot create, I do no understand” Richard Feynman
</aside>
- Machine Learning generative models are based on statistical principles and probability distributions to model the data and generate new samples.
- Deep learning- based generative models has recently become the dominant approach for generative modeling for it’s ability to learn complex mappings and to generate high-quality samples results.
- The objective is to learn patterns, structures, and representations from unlabeled data.
- Neural networks will be used as as generative models
- Neutral networks have a number of parameters significantly smaller than the amount of data we train them on
- “Models have to internalize the essence of data in order to generate it”
Generating Images
ImageNet dataset can be used for non-commercial use
A generative model in this case could be one large neural network that outputs images and we refer to these as “samples from the model”.
DCGAN
Deep Convolutional Generative Adversarial Networks
Input : 100 random drawn numbers from a uniform distribution
Output : Image, generated
Changing code shows the model has learned features to describe the world rather than just memorizing
The network has millions of parameters that we can tweak. The goal is to find a setting of theses parameters that makes samples generated from random code look like the training data
“Learning reusable feature representations from large unlabeled datasets”
