DataTerminology.comNEW
A set of training examples processed together in a single model update step.
DESCRIPTION
In machine learning, a batch groups multiple examples to compute gradients and update parameters efficiently. Mini-batches balance gradient estimate quality and compute efficiency, enabling GPU parallelism. Full-batch computes exact gradients but can be slow; stochastic uses single examples with noisy updates. Batch selection influences convergence, generalization, and stability. Practical training typically uses mini-batches tuned alongside learning rate and optimizer.
ALTERNATIVE DEFINITIONS
Google Developers ML Glossary
A set of examples used in one training iteration.
Concise definition focused on training iteration semantics.
Stanford CS231n
In practice, we use a minibatch of examples to estimate the gradient at each step, which makes the updates less noisy and more efficient.
Emphasizes mini-batches for efficiency and reduced gradient noise.
RELATED TERMS
Batch Size Mini-Batch Epoch Stochastic Gradient Descent Full-Batch Gradient Descent Learning Rate Optimizer Convergence
COMPETING TERMS
Online Learning Streaming Training
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Stochastic gradient descent
- Stanford CS231n - Optimization basics with mini-batches
- Google Developers - Glossary entry for batch
- PyTorch - Data loading and batching utilities
TOP QUESTIONS
Why use batches in training?
Batches enable efficient parallel computation, stabilize gradient estimates, and reduce memory overhead compared to full-batch training.
Category: Relevancy
How do batches relate to epochs?
An epoch is one pass over the dataset; it consists of multiple batches processed sequentially.
Category: Concept
Do larger batches always train faster?
They improve throughput but may hurt generalization or require learning rate scaling; beyond hardware limits, returns diminish.
Category: Tradeoff
What is mini-batch training?
Processing a small subset of examples per update, balancing gradient quality and computational efficiency.
Category: Definition
How to choose batch composition?
Shuffle data, optionally stratify by labels, and consider padding or bucketing for sequence lengths in NLP.
Category: Practice
Does batch order matter?
Shuffling each epoch reduces order effects and helps avoid systematic bias during training.
Category: Best practices
What is gradient accumulation?
Accumulating gradients across several small batches simulates a larger effective batch size within memory limits.
Category: Technique
How do batches affect Batch Normalization?
Small batches can yield noisy batch statistics, degrading BatchNorm performance; consider alternatives like LayerNorm.
Category: Interaction
Are batches relevant for inference?
Yes, batching improves throughput, but may add latency; tune batch sizes to meet SLAs.
Category: Operations