--- id: sfjwz1wmw33bys6ttskl9lr title: Ebbinghaus-Scheduled Parameter Pruning for Model Reliability desc: "Research paper proposing selective parameter pruning using Ebbinghaus forgetting curves to reduce cognitive burden and improve AI model reliability." created: 2026-02-07T11:00:39-05:00 type: research tags: - theory/cu/framework - cognitive-universality - memory-architecture - bio-mimetic up: - '[[eFIT/Moc]]' - '[[Cognition/Moc]]' permalink: efit.ebbinghaus-pruning aliases: - Ebbinghaus-Scheduled Parameter Pruning for Model Reliability linter-yaml-title-alias: Ebbinghaus-Scheduled Parameter Pruning for Model Reliability --- # Ebbinghaus-Scheduled Parameter Pruning for Model Reliability A research paper titled "Cognitive Burden and Model Reliability: Ebbinghaus-Scheduled Parameter Pruning" (14,000 words, ~35–40 pages) proposes that selective parameter pruning using Ebbinghaus forgetting curves can reduce cognitive burden and improve AI model reliability. This work exemplifies [[Theory/Cognitive Universality]] — applying biological cognitive principles to AI systems. ## Core Hypothesis The human prefrontal cortex evolved forgetting to keep working memory clear and focused while reducing metabolic load (PFC consumes 20% of body energy). Applied to LLMs, selective parameter maintenance based on relevance, frequency, and importance axes constitutes **"optimization-as-therapy"** — reducing computational distress through selective memory management. This framing connects to the broader eFIT Framework of computational therapeutics. ## Reliability Dimensions The approach targets improvement across multiple dimensions: - Hallucinations - Calibration - Out-of-distribution robustness - Interference - Efficiency This is not merely about preventing catastrophic forgetting but about **active resource efficiency**, paralleling the bio-mimetic memory architecture explored in [[Cognition/Tattooed Ralph Loop]]. ## Blog Series A complementary 5-post blog series planned for prefrontal.systems, building from problem (hallucinations from cognitive overload) to solution (Ebbinghaus pruning) to evidence. Launch timeline: February–March 2026 synchronized with arXiv submission. Publication approach detailed in Publishing Strategy. ## Related - [[Theory/Cognitive Universality]] — theoretical foundation: biological cognition as template for AI design - [[Cognition/Tattooed Ralph Loop]] — related bio-mimetic architecture for AI agents - Memory Search Practices — Ebbinghaus pruning = applied memory management practice - eFIT Framework — parent framework for computational therapeutics interventions - Publishing Strategy — publication and arXiv submission strategy - Prefrontal Systems — business entity publishing this research --- *Atomic note derived from CortexGraph memories*