Distribution Shift
A phenomenon in machine learning where a significant discrepancy exists between the distribution of training data and the distribution of data encountered in the deployment environment, leading to systematic degradation of model performance.
Core Facts
Distribution shift refers to the statistical distribution difference between the training environment and the deployment environment, and is the direct mechanism by which spurious correlations cause performance collapse.
Distribution shift refers to the systematic difference between the statistical distribution of data seen during model training and the distribution of data in the actual deployment environment.
Timeline (last 90 days)
When a neural network trained offline is deployed to a closed-loop control system, its actions change the system's state distribution, causing the data distribution at actual runtime to deviate from the offline data distribution at training time, that is, distribution shift.
Models based on causal mechanisms capture stable structures in the data-generating process and are theoretically more robust to distributional shifts.
The positive correlation between base task performance and generalization ability weakens under significant distribution shift.
GAP-DPO mitigates distribution shift caused by off-policy sampling through an epoch-wise regeneration mechanism that regenerates samples at each epoch.
Even with palm-size normalization, differences in hand structure, pinching habits, and finger-length ratios across individuals mean a single static threshold cannot simultaneously work for multiple users, which is fundamentally a distribution shift problem.
The real world is non-stationary, and a policy's performance during training may deteriorate sharply after the environment undergoes a distribution shift.
Distribution shift refers to the statistical distribution difference between the training environment and the deployment environment, and is the direct mechanism by which spurious correlations cause performance collapse.
All Facts (10)
Distribution shift refers to the statistical distribution difference between the training environment and the deployment environment, and is the direct mechanism by which spurious correlations cause performance collapse.
75%VerifiedDistribution shift refers to the systematic difference between the statistical distribution of data seen during model training and the distribution of data in the actual deployment environment.
75%UnverifiedWhen a neural network trained offline is deployed to a closed-loop control system, its actions change the system's state distribution, causing the data distribution at actual runtime to deviate from the offline data distribution at training time, that is, distribution shift.
50%UnverifiedModels based on causal mechanisms capture stable structures in the data-generating process and are theoretically more robust to distributional shifts.
50%UnverifiedThe positive correlation between base task performance and generalization ability weakens under significant distribution shift.
50%UnverifiedGAP-DPO mitigates distribution shift caused by off-policy sampling through an epoch-wise regeneration mechanism that regenerates samples at each epoch.
50%UnverifiedEven with palm-size normalization, differences in hand structure, pinching habits, and finger-length ratios across individuals mean a single static threshold cannot simultaneously work for multiple users, which is fundamentally a distribution shift problem.
50%UnverifiedThe real world is non-stationary, and a policy's performance during training may deteriorate sharply after the environment undergoes a distribution shift.
50%Unverified离线RL的核心挑战是分布偏移，可能导致外推误差，使学到的策略在部署时表现远不如预期
50%Unverified域偏移是指训练数据分布与实际部署场景数据分布之间的系统性差异，是迁移学习和AI工程落地的核心挑战之一
50%