Get the latest tech news
The Structure of Neural Embeddings
Another place for thought infusion
A small collection of insights on the structure of embeddings (latent spaces) produced by deep neural networks. Adversarial Vulnerability: Small changes in input space can cause large shifts in embeddings and therefore also in predictions made from them, suggesting the learned manifolds have irregular geometric properties. Within-class variation becomes minimal compared to between-class differences, effectively creating distinct, well-separated clusters for each class.
Or read this on Hacker News