Superposed Latent Autoencoder
The Superposed Latent Autoencoder (SLAE) is introduced, which preserves high-capacity latent representations while sharing storage through learned superposition, and suggests a new principle for representation compression: instead of making every latent smaller, keep representations wide and let them share memory.
Quan-Ling Zhao, Jia-Ying Yang, Tian-Qi Zhang et al.
· 0 citations