Short Definition
Importance sampling is a technique that weights samples to correct or emphasize parts of a data distribution.
Definition
Importance sampling is a statistical method in which samples are drawn from one distribution but weighted to reflect another target distribution. In machine learning, it is used to focus learning or estimation on more informative, rare, or high-impact samples while maintaining unbiased estimates under certain conditions.
Importance sampling changes how much each sample matters, not which samples exist.
Why It Matters
Uniform sampling treats all data points equally, even when some contribute more to learning or evaluation objectives. Importance sampling can improve efficiency, reduce variance, and correct sampling mismatches—especially under class imbalance or distribution shift.
When applied correctly, it accelerates learning without changing the underlying objective.
How Importance Sampling Works
Each sample is assigned a weight proportional to the ratio between the target distribution and the sampling distribution. These weights adjust the contribution of each sample during training or estimation.
Weights compensate for non-uniform sampling.
Minimal Conceptual Example
# conceptual importance weightingloss = weight * compute_loss(sample)Common Use Cases
Importance sampling is commonly applied in:
- handling class imbalance
- rare event detection
- off-policy evaluation and reinforcement learning
- Monte Carlo estimation
- correcting covariate shift
- accelerating convergence during training
Its flexibility makes it widely applicable.
Importance Sampling vs Resampling
- Importance sampling: reweights samples without duplication
- Resampling: changes sample frequency via duplication or removal
Both alter the effective data distribution but in different ways.
Benefits and Trade-offs
Benefits include:
- improved sample efficiency
- better focus on informative regions
- unbiased estimation under correct weighting
Trade-offs include:
- increased variance from extreme weights
- sensitivity to weight estimation errors
- numerical instability if weights are poorly controlled
Weight management is critical.
Common Pitfalls
- using unbounded or highly skewed weights
- applying importance sampling without validating assumptions
- confusing importance weights with class weights
- ignoring variance inflation effects
Importance sampling amplifies both signal and noise.
Relationship to Generalization
Importance sampling can improve generalization by aligning training data with deployment distributions. However, if target distributions are misspecified, it can degrade performance and introduce bias.
Relationship to Training Dynamics
By emphasizing certain samples, importance sampling influences gradient magnitudes and optimization behavior. This can accelerate convergence but may destabilize training if not carefully tuned.
Neural Network Lexicon 