https://github.com/wilberquito/mnist-autoencoder-classification

Autoencoder Feature Extraction for Classification. SSL
https://github.com/wilberquito/mnist-autoencoder-classification

Last synced: 4 days ago
JSON representation

Autoencoder Feature Extraction for Classification. SSL

• Host: GitHub
• URL: https://github.com/wilberquito/mnist-autoencoder-classification
• Owner: wilberquito
• Created: 2024-04-30T19:32:59.000Z (7 months ago)
• Default Branch: main
• Last Pushed: 2024-05-01T21:16:06.000Z (7 months ago)
• Last Synced: 2024-05-02T13:27:07.564Z (7 months ago)
• Language: Jupyter Notebook
• Homepage:
• Size: 2 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Autoencoder Feature Extraction for Classification

Autoencoders are a type of neural network which generates an “n-layer” coding

of the given input and attempts to reconstruct the input using the code

generated.

The Autoencoder architecture architecture is divided into the encoder

structure, the decoder structure, and the latent space, also known as the

“bottleneck”.

## Encoder

$$h = E(x)$$

## Decoder

$$x' = D(h)$$

## Latence space

This is the data representation or the low-level, compressed representation of

the model’s input. The decoder structure uses this low-dimensional form of data

to reconstruct the input. It is represented by $h$.

## Self Supervised Learning

### pretext task

Now for both model to learn, we need a metric. This metric of loss $L$ which

should mesure how good the decoder $D$ does reconstructing the original data

from the encoder $E$.

$$L = Loss(x, x')$$

### downstream task

The learned representation by encoder $E$ can be used and fine tunned. As

encoder $E$ "knows" important features from the SSL problem, we can use it for

transfer learning in a classification or regression task.