Superposed Latent Autoencoder improves representation compression with shared memory

Researchers have introduced the Superposed Latent Autoencoder (SLAE), a novel approach to representation compression that allows multiple wider latent representations to share storage through learned superposition. Unlike traditional autoencoders that reduce latent size, SLAE transforms latents into storage-friendly codes, binds them with randomized keys, and superposes them into a single memory tensor. This method significantly improves the reconstruction-memory tradeoff, reducing reconstruction error by up to 56% and enhancing downstream classification performance by up to 16.79 percentage points under identical memory budgets. AI

IMPACT Introduces a new principle for representation compression that could lead to more efficient AI models.

RANK_REASON The cluster describes a new method presented in an academic paper on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• CIFAR-10
• CIFAR-100
• SLAE
• STL-10
• Superposed Latent Autoencoder
• The Street View House Numbers Dataset
• Tiny-ImageNet

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →