CoLLAs 2026 talk: memorization may be unavoidable — curation, unlearning, pruning as strategies
gkdziugaite · x · 2026-09-16
Karolina Dziugaite's talk at CoLLAs 2026 in Bucharest examined the complex dynamics of memorization in deep learning. Her key claim: since eliminating memorization entirely may be impossible, she outlined practical strategies around data curation, unlearning, and pruning. The post also reflects on engaged Q&A, panel, and mentorship sessions at the conference.
More from Research
- New paper lifts LLM memory retention 28x to 34.9% across 100 sequential tasks — DanielKhashabi · 2026-09-16
- ArtManip: category-level in-hand manipulation of articulated objects — siyuanhuang95 · 2026-09-16
- smolperfbenchmark: a public leaderboard for small open models on 8GB-class devices — East-Muffin-6472 · 2026-09-16
- Economist challenges AI labor paper: diff-in-diff can't support wage-loss claims — paulnovosad · 2026-09-16
- Microsoft paper: unaligned small models 'launder' capabilities via frontier model consultation — dair_ai · 2026-09-16
- New Paper Shows SAE and Probe Steering Outperform Prompts for LLM Social Simulation Agents — daveholtz · 2026-09-16