A new arXiv paper explores how the resolution at which neural computation is observed affects its apparent structure and computational capacity. The research introduces a scale-indexed decomposition to differentiate predictive information from direct dependencies versus residual information. It demonstrates that projection can transform higher-order dependencies into pairwise structures, potentially increasing the fraction of information captured by direct dependencies without necessarily increasing the total predictive information. This framework suggests that models using point neurons or direct population approaches establish descriptions at a specific scale, but do not fully identify the finer computational processes that generated the observed relationships. AI
IMPACT Provides a theoretical framework for understanding how observational scale impacts the interpretation of neural network computations.
RANK_REASON The item is an academic paper published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
Read on arXiv cs.NE (Neural & Evolutionary) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
