Thread by Harpreet Sahota 🥑

twitter.com
1 Recommender
1 Mention
Recommend
Post
Save
Complete
Collect
Thread
Unlock the power of latent space and latent vectors: the key to understanding autoencoders.

A brief guide for busy learners 👇🏽🧵

#deeplearning
Autoencoders are a type of neural network that take in some input data and try to output the exact same data.

They do this by first "encoding" the input data into a lower-dimensional space (called the "latent space") and then "decoding" it back into the original data.
The latent space is like a "shortcut" the encoder takes to go from the input data to the output data.

Imagine you're going to your friend's house, but there's construction on your usual route.

You want to find an alternate path that will get you there fast.
You'd want a path that has fewer lights, shorter distance, and less left turns.

In essense, a simple path and not have to go around the entire city.

The latent space would be that alternate path that cuts through the construction and gets you to your friend's house, quickly.
The lower-dimensional representation in the latent space is called a "latent vector," and it's useful because it's a compressed version of the input data.
You can think of the latent vector as a "map" of the shortcut that shows you exactly where to turn and how far to go.

The latent vector shows the encoder how to navigate the latent space to get from the input data to the output data.
Autoencoders can be used for a variety of tasks, such as compression, dimensionality reduction, data denoising (removing noise from images), data augmentation, and anomaly detection.
Autoencoders are typically trained using a reconstruction loss, which measures how well the output data matches the input data.
Imagine you have a jumbled puzzle, and you're trying to put the pieces together.

The puzzle pieces represent the input data, and the completed puzzle represents the output data.

The reconstruction loss is like a measure of how well you're doing at putting the puzzle together.
It's a way of comparing the output data (the completed puzzle) to the input data (the puzzle pieces) and seeing how well they match up.
If the reconstruction loss is small, you've almost perfectly recreated the picture.

If the reconstruction loss is large, you have more work to do to get the puzzle to match the picture.
There are various ways to do this, but a common approach is to use a loss function such as the L2 distance between the output data and the input data.
The L2 distance is calculated by taking the difference between each element of the output data and the corresponding element of the input data, squaring the differences, and then taking the square root of the sum of all the squared differences.
This gives us a single scalar value that represents the overall difference between the output data and the input data.
That's it for autoencoders.

If you want more deep learning fundamentals in your feed, for free, be sure to follow me @DataScienceHarp
Mentions
See All