What does hidden (latent) space hide underneath?
Basic Concepts
Encoder in machine learning – it is the part of the model that transforms the input data into another representation.
Decoder in machine learning is a model component that takes a hidden representation of data (such as that received from an encoder) and transforms it back into the original data.
Latent space – Latent space, also known as hidden object space or embedding space, is an embedding of a set of elements into a manifold in which elements that are similar to each other are located closer to each other.
To better understand this idea, let's compare it to how people perceive the world around them.
For example, when we are walking through the forest and see a tree, instead of remembering every little detail of a given plant, we retain only a general idea of its shape, type and size. This allows us to quickly identify a tree without being bogged down with unnecessary details.
Similarly, latent space attempts to provide the computer with a compressed understanding of the tree. It highlights key characteristics of a tree, such as crown shape, trunk structure, and the presence of branches, without explicitly indicating each of them.
That is, it is simply a representation of compressed data in which similar data points are located closer together in space.
We can understand patterns or structural similarities between data points by analyzing the data in latent space, be it through manifolds, clustering, etc.
Working with Latent Spaces
Latent space is useful for exploring data features and finding simpler representations of data for analysis.
As we can see in the picture, such concepts as encoder and decoder are touched upon. But how do they work?
The encoder shapes vectors in latent space such that similar objects are placed close to each other, which helps create clusters, while different objects are placed further apart.
Let's say we have a well-trained encoder working with images of trees. If we feed it an image of a spruce, it will place the vector of this image close to the vectors corresponding to coniferous trees, while it will place the vector of the image of an oak in another part of the latent space, next to the vectors of deciduous trees.
After training, the decoder is able to reconstruct the original objects from a high-dimensional latent vector. It is important to note that in addition to restoring the original objects, the decoder can also be used to create completely new data. To do this, it is enough to provide it with a latent representation of objects that were not in the training data set.
How are latent spaces used in the eXplain-NNs library?
Visualization of latent spaces: This method allows hidden features or patterns learned by the neural network to be mapped into these latent spaces. This can be useful for understanding how the model organizes data and what internal representations it uses to make decisions.
Homology analysis of latent spaces: Another method provided by the eXplain-NNs library is latent space homology analysis. Homology analysis is used to study the structure and connections between these latent representations. This helps to understand how information is organized within the model and influences its ability to make decisions.
