Batch, Online, and Continual Learning

Three regimes, one objective. That objective can be trained under very different schedules and controls.

Batch retraining collects a dataset, retrains periodically, validates, and deploys a version. Online updating adjusts the model incrementally as examples or feedback arrive. Continual learning works across a sequence of tasks or distributions while trying to preserve earlier capabilities. The rest of this lesson is about what each of those choices costs when it is wrong, measured in cases where somebody counted.

An online learner may receive labels, rewards, pseudo-labels, or another signal. The term describes incremental updating, not a separate target format.

A batch model can serve real-time predictions. An online model can update only after delayed outcomes mature. Serving latency and learning cadence are separate system properties.

Faster adaptation increases the burden of validation, traceability, and rollback. Periodic batch releases buy clear version boundaries, broad offline testing, and simple rollback. They pay for it by adapting more slowly. Incremental online updates respond quickly to change, but they need stable feedback and guards, can amplify transient noise, and are harder to reproduce exactly. A continual strategy meets recurring distribution change with replay or parameter protection and evaluates across old and new tasks. Forgetting stays its central risk.

The price of the slow cell has been counted. Google Flu Trends kept predicting flu after the world had moved. The algorithm was rebuilt in 2009 and then left alone: “GFT engineers updated the algorithm in 2009, and this model has run ever since, with a few changes announced in October 2013.” Measured against the outcome it was built to track, that model reported overly high flu prevalence in 100 out of 108 weeks between 21 August 2011 and 1 September 2013. It overshot the actual 2011–2012 level by more than 50%. Lazer and three colleagues reported this in Science in 2014; epidemiologists had already reassessed the same system at three geographic scales the year before. On 20 August 2015 Google's Flu Trends Team announced it would stop maintaining the Flu Trends website and hand the signal directly to Columbia University's Mailman School of Public Health, Boston Children's Hospital/Harvard and the CDC Influenza Division.

Nothing about that failure required a bad model. It required a good one and a cadence that never came round again. 100 wrong weeks out of 108 is what “adapts more slowly” costs when the world moves and the release does not.

Fraud patterns change quickly. The label does not arrive when the fraud does. For US consumer debit card fraud, how long a confirmed label can legally take to exist is not a modelling decision at all. It is fixed by Regulation E, 12 CFR 1005.11, “Procedures for resolving errors”.

The update mechanism should specify evidence, gates, and recovery, not only code. In one regulated sector that is no longer advice. It is a filing.

The US Food and Drug Administration finalised guidance on predetermined change control plans for AI-enabled device software on 4 December 2024, and reissued it on 18 August 2025. It covers devices whose model modifications are implemented automatically, and it gives that arrangement a name: “continuous learning”. A manufacturer has to establish three things in advance. The modifications it intends to make. The methodology to develop, validate and implement them. And an impact assessment. On the fastest cadence the agency writes: “To help facilitate FDA's review in this regard, it may be helpful for manufacturers to clearly establish boundaries or guardrails that define the range of automatic updates.”

The steps that follow are that document rewritten for a team with no filing deadline. Define feedback maturity, stating which outcomes are reliable enough to train on and when. Choose an update cadence matched to drift speed and review capacity. Preserve replay evidence covering old, rare, and safety-critical cases. Gate every version with offline, shadow, and slice checks before expanding exposure. Detect regressions in both recent performance and retained capabilities. And support rollback with reproducible artifacts, thresholds, and deployment controls.

If behavior immediately shapes the next model, then coordinated manipulation, accidental feedback, and unequal exposure can all change the system. Users may not know which of their actions become training signals. Both halves of that have already happened on the record.

The deliberate half took under a day. Microsoft's Tay chatbot learned from the conversations users had with it. Peter Lee, then Corporate Vice President of Microsoft Research, wrote on The Official Microsoft Blog on 25 March 2016: “Unfortunately, in the first 24 hours of coming online, a coordinated attack by a subset of people exploited a vulnerability in Tay.” Tay was taken offline. Lee wrote that it would return only once Microsoft could better anticipate malicious intent. The learning loop was the attack surface. The product had no other.

The accidental half is slower and much harder to see, because nobody has to attack anything. In 2016 Lum and Isaac simulated PredPol on Oakland drug-crime records. Black people would be targeted at roughly twice the rate of whites, while estimated drug use was far more evenly spread. Drug arrests in the underlying records concentrated in two Oakland neighbourhoods by a factor of about 200. Five researchers then proved the mechanism formally in 2018: “Such systems have been shown susceptible to runaway feedback loops, where police are repeatedly sent back to the same neighborhoods regardless of the true crime rate.” The model was retrained on the data its own deployment generated, and the loop closed by itself.

Define consent, privacy, abuse controls, and monitoring before online adaptation. In many products, periodic reviewed updates are the more responsible choice.

Each deployed version needs a record. It should preserve enough evidence to explain what changed and to restore earlier behavior.

Three supervisors now put that obligation in their own words. On 17 April 2026 the Federal Reserve Board, the OCC and the FDIC jointly issued revised guidance on model risk management, as Federal Reserve SR 26-2 and OCC Bulletin 2026-13. It supersedes SR 11-7 of 4 April 2011 and OCC Bulletin 2011-12, and the OCC's issuance also rescinds Bulletins 1997-24 and 2021-19. It is expected to be most relevant to banking organizations with over $30 billion in total assets. The guidance defines the standing question a ledger exists to answer: “Ongoing model monitoring involves an evaluation of the extent to which a model is performing as expected given potential changes in products, exposures, activities, clients, data relevance, or market conditions.” Note the date. The guidance it replaced had stood since 2011.