latent space
In plain wordsThe hidden internal map a model builds of its data, where each item becomes a point and similar items sit near each other. Moving smoothly through it gradually changes what the model represents or generates.
The vector space of a model's internal representations, especially the code space of an autoencoderautoencoder, n.: A neural network trained to squeeze its input into a compact code and then rebuild the original from that code. Learning to compress this way forces it to capture what matters in the data., variational autoencodervariational autoencoder, n.: A neural network that learns to squeeze data into a compact, slightly fuzzy code and rebuild it, arranged so that decoding random codes produces new, plausible examples., GANgenerative adversarial network, n.: Two neural networks trained as rivals: one forges fake images, the other tries to spot the fakes. As both improve, the forger's output becomes realistic enough to fool people. or other generative model, in which each point (a latent vector) corresponds to a possible input or output. 'Latent' signals that the coordinates are not observed in the data but inferred or learned. Because training places similar items near each other, operations in latent space can be meaningful: interpolating between two latent vectors yields gradual transitions, and vector arithmetic can add or remove attributes such as a smile. Latent diffusion modelslatent diffusion model, n.: A diffusion image generator that works on a shrunken, compressed version of each picture instead of the full-size pixels, which makes it far cheaper to train and run. run their generative process in a compressed latent space instead of pixel space to save computation. In discussions of language models the term is used loosely for activation spaceactivation space, n.: The imaginary space of all possible internal states of one layer of a neural network, where each input becomes a point and related meanings tend to share places or directions.. Individual latent dimensions are usually not interpretable without additional constraints (disentangled representationdisentangled representation, n.: An internal representation in which each part tracks one separate, meaningful factor of the data, such as one number for a face's angle, another for its lighting and another for hair color.).
Interpolating in latent space between the two faces produced a smooth morph.
Origin
From 'latent variables' in statistics: unobserved quantities inferred from observed ones, as in factor analysis (Charles Spearman, 1904). The phrase spread in machine learning with deep generative models, notably Diederik Kingma and Max Welling's variational autoencoder (2013) and Alec Radford, Luke Metz and Soumith Chintala's DCGAN (2015), which demonstrated latent interpolation and vector arithmetic on faces.
Usage note
Each model, and each layer, has its own latent space. 'Latent space' and 'embedding space' are often used interchangeably; the former is favored for generative models, the latter for retrieval and similarity.
Related
- embeddingembedding, n.: A list of numbers that stands for a word, picture or other item, arranged so that similar items get similar lists. It lets a computer measure similarity of meaning as distance.
- autoencoderautoencoder, n.: A neural network trained to squeeze its input into a compact code and then rebuild the original from that code. Learning to compress this way forces it to capture what matters in the data.
- variational autoencodervariational autoencoder, n.: A neural network that learns to squeeze data into a compact, slightly fuzzy code and rebuild it, arranged so that decoding random codes produces new, plausible examples.
- latent diffusion modellatent diffusion model, n.: A diffusion image generator that works on a shrunken, compressed version of each picture instead of the full-size pixels, which makes it far cheaper to train and run.
- activation spaceactivation space, n.: The imaginary space of all possible internal states of one layer of a neural network, where each input becomes a point and related meanings tend to share places or directions.
- disentangled representationdisentangled representation, n.: An internal representation in which each part tracks one separate, meaningful factor of the data, such as one number for a face's angle, another for its lighting and another for hair color.
- manifold hypothesismanifold hypothesis, n.: The idea that real-world data, such as photos of faces, fills only a thin, smoothly curved sliver of the vast space of all possible data — which is why learning from examples works at all.
- representation learningrepresentation learning, n.: Letting a computer work out for itself how best to describe its data, such as the edges and shapes in images or the meanings of words, instead of relying on descriptions people design by hand.
- activation spaceactivation space, n.: The imaginary space of all possible internal states of one layer of a neural network, where each input becomes a point and related meanings tend to share places or directions.
Mentioned in
embeddingembedding, n.: A list of numbers that stands for a word, picture or other item, arranged so that similar items get similar lists. It lets a computer measure similarity of meaning as distance., autoencoderautoencoder, n.: A neural network trained to squeeze its input into a compact code and then rebuild the original from that code. Learning to compress this way forces it to capture what matters in the data., encoder–decoder architectureencoder–decoder architecture, n.: A two-part model design: one part, the encoder, reads the input and condenses it into an internal summary; the other, the decoder, uses that summary to produce the output., generative modelgenerative model, n.: A model that learns what a kind of data looks like — faces, sentences, songs — well enough to produce new examples that resemble the originals without copying them., hidden layerhidden layer, n.: Any layer of a neural network between the input and the output. Its values are never shown to the user; they are the network's internal working-out., latent diffusion modellatent diffusion model, n.: A diffusion image generator that works on a shrunken, compressed version of each picture instead of the full-size pixels, which makes it far cheaper to train and run., representation learningrepresentation learning, n.: Letting a computer work out for itself how best to describe its data, such as the edges and shapes in images or the meanings of words, instead of relying on descriptions people design by hand., StyleGANStyleGAN, prop. n.: An image-generating network from NVIDIA, famous for photorealistic faces of people who do not exist, that sets coarse traits such as head angle and face shape separately from fine details such as coloring and skin texture., variational autoencodervariational autoencoder, n.: A neural network that learns to squeeze data into a compact, slightly fuzzy code and rebuild it, arranged so that decoding random codes produces new, plausible examples., activation spaceactivation space, n.: The imaginary space of all possible internal states of one layer of a neural network, where each input becomes a point and related meanings tend to share places or directions., disentangled representationdisentangled representation, n.: An internal representation in which each part tracks one separate, meaningful factor of the data, such as one number for a face's angle, another for its lighting and another for hair color., manifold hypothesismanifold hypothesis, n.: The idea that real-world data, such as photos of faces, fills only a thin, smoothly curved sliver of the vast space of all possible data — which is why learning from examples works at all.