ML2 - M4 Modules

is a type of machine learning model that learns the underlying distribution of a dataset to generate new data that resembles the training data.

generative model

Learn boundaries

Discriminative

Learn how data is structure

Generative

Discriminante Model

Generative Model

is a type of neural network used to compress and then reconstruct input data.

autoencoder

Compresses input into a latent representation

Encoder

Reconstructs input from this compressed form

Decoder

Goal of autoencoder is to learn efficient _________

representations

autoencoder

Autoencoder Architecture

_________ → ____________ → ____________ → ___________ → ____________

Input
Encoder
Bottleneck
Decoder
Output

Learns to reconstruct clean input from noisy version

Denoising AE

Introduces sparsity constraints in hidden layers

Sparse AE

Concept only Learns probability distributions in the latent space Enables generation of new data by sampling from that space

Variational Autoencoder (VAE)

Measures how similar output is to input (e.g., MSE)

Loss function

Use Cases of Autoencoders

Image compression

Anomaly detection

Data denoising

Pretraining for deep networks

Autoencoders aren 't just for _________—they help the network understand data better.

compression

consists of two neural networks—the Generator and the Discriminator—that compete in a game-theoretic setup.

GAN

GAN

Generative Adversarial Network

tries to produce realistic data

Generator

tries to detect if the data is real or fake

Discriminator

In GAN, The goal is for the generator to eventually “_______” the discriminator

fool

z→G(z)→xfake

Generator

x→D(x)→[real/fake]

Discriminator

Generator wants Discriminator to be ___________

wrong

Discriminator wants to be _________

correct

Generator outputs similar images

Mode collapse

Generator and discriminator don ’t balance

Unstable training

Discriminator becomes too good too fast

Vanishing gradients

DCGAN

Deep Convolutional GAN

DCGAN

Uses __________ layers instead of fully connected layers

convolutional

Introduced architectural guidelines (ReLU in G, LeakyReLU in D)

DCGAN

____________ was the first GAN architecture to produce high-quality images at scale.

DCGAN

Goal - Reconstruct input

autoencoder

Goal - Generate realistic new data

GAN

Architecture - Encoder-Decoder

autoencoder

Architecture - Generator - Discriminator

GAN

Output - Quality Often blurry

autoencoder

Output - Sharper, More realistic

GAN

Stability - Easier to train Difficult to converge

autoencoder

Stability - Difficult to converge

GAN

Use Case - Denoising, Compression

autoencoder

Use Case - Data synthesis, Image generation

GAN