Risks from Learned Optimization in Advanced Machine Learning Systems
arXiv:1906.01820
Abstract
We analyze the type of learned optimization that occurs when a learned model (such as a neural network) is itself an optimizer - a situation we refer to as mesa-optimization, a neologism we introduce in this paper. We believe that the possibility of mesa-optimization raises two important questions for the safety and transparency of advanced machine learning systems. First, under what circumstances will learned models be optimizers, including when they should not be? Second, when a learned model is an optimizer, what will its objective be - how will it differ from the loss function it was trained under - and how can it be aligned? In this paper, we provide an in-depth analysis of these two primary questions and provide an overview of topics for future research.
References in corpus (5)
- Scalable agent alignment via reward modeling: a research direction
- Deep learning generalizes because the parameter-function map is biased towards simple functions
- Universal Planning Networks
- Occam's razor is insufficient to infer the preferences of irrational agents
- TreeQN and ATreeC: Differentiable Tree-Structured Models for Deep Reinforcement Learning
Cited by in corpus (11)
- On the Opportunities and Risks of Foundation Models
- Unsolved Problems in ML Safety
- Alignment of Language Agents
- Tasks, stability, architecture, and compute: Training more effective learned optimizers, and using them to train themselves
- Extending Environments To Measure Self-Reflection In Reinforcement Learning
- Truthful AI: Developing and governing AI that does not lie
- On Controllability of AI
- REALab: An Embedded Perspective on Tampering
- Avoiding Tampering Incentives in Deep RL via Decoupled Approval
- Alignment Problems With Current Forecasting Platforms
- Achilles Heels for AGI/ASI via Decision Theoretic Adversaries