Autoencoders and their applications Flashcards
What is dimensionality reduction?
A technique for representing multi-dimensional data in a lower-dimensional space while preserving key information.
Why is dimensionality reduction important?
It helps address the curse of dimensionality, where increased dimensions make data sparse and models prone to overfitting.
What causes the curse of dimensionality?
Adding features without increasing training samples makes the feature space sparse, leading to overfitting.
What is an autoencoder?
An unsupervised artificial neural network that learns to encode input data into a lower-dimensional representation and reconstructs the original input.
What are the key components of an autoencoder?
Encoder, Decoder, Latent Space, and Objective Function.
What does the encoder do?
Compresses input data into a lower-dimensional representation.
What does the decoder do?
Reconstructs the original input from the latent representation.
What is the latent space?
The compressed representation (code) learned by the encoder.
What is the objective function of an autoencoder?
To minimize the reconstruction error, often using Mean Squared Error (MSE).
What loss function is typically used in autoencoders?
Mean Squared Error (MSE).
What is a sample structure of a simple autoencoder?
Input → Dense (hidden) → Dense (code) → Dense (hidden) → Output.
What activation functions are commonly used in AE layers?
ReLU for hidden/code layers, and Sigmoid for the output layer.
What does training an autoencoder involve?
Using input data as both input and target for training.
How do we visualize an autoencoder’s performance?
By comparing the input images with the reconstructed outputs.
What is a denoising autoencoder?
An autoencoder trained to reconstruct clean data from noisy input.
What is a convolutional autoencoder?
An autoencoder that uses convolutional layers for encoding and decoding images.
What does the encoder in a CAE do?
Applies Conv2D layers with activation (e.g., ReLU) followed by MaxPooling to compress spatial dimensions.
What does the decoder in a CAE do?
Applies Conv2D and UpSampling (or Conv2DTranspose for learnable upsampling) layers to reconstruct the image.
What is the benefit of using Conv2DTranspose over UpSampling?
Conv2DTranspose performs learnable upsampling, allowing the model to better reconstruct spatial details.
How is the convolutional autoencoder trained?
By minimizing the MSE between input and output images using optimizers like Adam.
What are typical training parameters for CAEs?
Examples include 200 epochs, batch size 2, learning rate decay, and MSE loss.
Name some applications of autoencoders.
Denoising images
Image colorization
Anomaly detection
Dimensionality reduction
Feature learning
How are autoencoders used in denoising?
By training on noisy inputs and clean targets, the model learns to remove noise.
How are autoencoders used in colorization?
Train the model to convert grayscale images into color versions.
