Neural network toy model demonstrates computation in superposition

Researchers have developed a toy model to explore computation in superposition within neural networks. By training a single-hidden-layer ReLU network with 50 neurons to compute 100 sparse input features under an L4 loss function, they observed a solution that appears to perform all computations in superposition. The study reverse-engineered this solution, finding that the network assigns sparse binary codewords to each feature and decodes them using a pseudoinverse of the encoder. Further analysis revealed that a description with just three scalars could recover most of the network's performance, which was validated by constructing equivalent networks with hand-designed codes. AI

IMPACT This research offers a new theoretical framework for understanding and potentially designing neural networks capable of more efficient computation.

RANK_REASON The cluster contains a research paper detailing a novel toy model for computation in superposition within neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Francisco Ferreira da Silva
• Gotit.pub
• Hugging Face
• IArxiv
• L4
• ReLU
• ScienceCast

• paper
• model release

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