Is online learning the same thing as "incremental learning"?

They're closely related and often used interchangeably. Some practitioners reserve "incremental learning" for any regime that updates on new data without retraining from scratch (which could still happen in scheduled chunks) and "online learning" specifically for updating example-by-example or mini-batch-by-mini-batch in real time as data streams in. In casual use across most ML content, the terms overlap heavily.

Does online learning mean the model never sees historical data again?

Not necessarily — many online systems still start from a batch-trained base model (trained on historical data) and then apply online updates on top of that foundation. Pure "learn from scratch, online-only, no historical batch training" setups exist but are less common in production than hybrid approaches.

What is catastrophic forgetting, concretely?

It's when a model's weights shift enough in response to recent updates that performance on older, still-relevant patterns degrades — for example, an online fraud model that sees mostly one fraud pattern for a few weeks can lose sensitivity to a different pattern it learned months ago and hasn't seen reinforced since, even though that older pattern could resurface at any time.

Can a neural network be trained in an online fashion?

Yes — SGD-based training is naturally compatible with online updates, and streaming/online variants of neural network training are common in production recommender systems and ad-tech. The main engineering challenge is usually infrastructure (low-latency feature pipelines, safe weight-update mechanisms) rather than the algorithm itself.

How do teams get the adaptiveness of online learning without the instability risk?

Common guardrails include shadow-mode evaluation (the online model runs alongside the production model without serving live predictions until it proves out), bounded or decaying learning rates so any single bad example can't swing the weights too far, periodic rollback checkpoints to a known-good state, and automated drift/anomaly monitoring that pauses online updates if incoming data looks unusual.

← AI and Data Science Studio

Concept Explainer · AI & Data Science

Batch vs. Online Learning

Not a question of algorithm — a question of when the model is allowed to change. One retrains on a fixed snapshot; the other never really stops learning.

"Batch" here doesn't mean mini-batch gradient descent (that's an optimization detail — how many examples contribute to one weight update). It means the training regime: does the model see a full, fixed dataset before it's allowed to make a single production prediction, or does it update incrementally, example-by-example or mini-batch-by-mini-batch, as new data streams in after deployment? That distinction — fixed snapshot vs. continuous stream — determines how a model behaves when the world underneath it changes.