Exponential Moving Average Explained Briefly | Generated by AI

Home 2026.07

EMA stands for Exponential Moving Average.

It’s a type of weighted average that gives more weight to recent observations and less weight to older ones, decaying exponentially over time.


In the context of your training loss discussion

When you said “the EMA window is very small”, you meant that the smoothing factor used to average your loss values over time is set to look at only a very short history.


Mathematically (simple version)

\[ \text{EMA}_t = \alpha \times \text{value}t + (1 - \alpha) \times \text{EMA}{t-1} \]


Where else is EMA used in ML/LLM training?

  1. Model parameter averaging (e.g., EMA of weights) – often used to produce a more stable final model for inference.
  2. Optimizer states – some optimizers (like Adam) use EMA-like logic for gradient moments.
  3. Batch norm statistics – often updated with EMA during training.

In your case, you were using EMA on the loss curve to decide whether training is stable. Your point was: don’t be fooled by a smooth EMA – if the window is too small, the noise is still there, and if you look at raw 10-step intervals, you’ll see the jumps clearly.

Does that clear it up? Want me to explain how to choose the right EMA window for your use case?


Back Donate