Go deeper · Advanced track

How can models learn with less data, less sharing or without forgetting?

By changing how training is set up, not only how big it is. A model can reuse what another model already learned, ask people to label just the examples it is least sure about, learn something new while protecting old skills, or train across many devices without collecting the raw data in one place. Automated search can even help choose the model itself.

Grades 9+25 minutesAverages onlyOptional: noise and privacy

This is an optional advanced page. You do not need it for the main foundations route.

What you will be able to do

Match transfer, active, continual and federated learning, evolutionary search and AutoML to the problem each one solves; spend a small label budget wisely; compute a federated average by hand; and explain why “federated” does not automatically mean “private.”

Before you start

Read first: Learning now, or using what it learned?, Do you always need a neural network? and Can you build and test a model yourself?

1 · You have already seen this

What you bring to a new job

Most people have started something new without starting from nothing.

On a first day at a new job, you bring skills from the last one, so you learn faster than someone who never did similar work.

Models can be trained in similar ways. This page covers transfer learning, active learning, continual learning, federated learning and automated search: ways for a model to learn with less data, less sharing or without forgetting. It shows how they work, lets you try one, and shows where each goes wrong.

2 · What happens, step by step

Training, one choice at a time

Training a model means making five choices, and each method on this page changes one of them. The next section sets them side by side for one garden app, and the one after goes deeper into step 1. The lab lets you try step 2, and the three sections after it go deeper into steps 5, 3 and 4.

How a model gets trained

Five choices

  1. 1 · Where to startTraining from scratch needs a lot of labeled data. Transfer learning starts from a model already trained on another task and adjusts it with your few examples, which is called fine-tuning.
  2. 2 · Which examples to labelEach label costs a person’s time. In active learning the model asks for labels on the examples it is least sure about.
  3. 3 · Where the data livesUsually all the data is collected in one place. In federated learning each device or site trains on its own data and sends back only model updates, which a server averages.
  4. 4 · Which model and settingsSomeone has to pick the kind of model and settings such as the learning rate. Evolutionary methods and AutoML search for them: try, score, keep the best, change, repeat.
  5. 5 · What happens next timeNew data keeps arriving after training. Continual learning teaches the model new things while protecting what it learned before.

Five problems, one app

One garden app, five problems

Before any new words, here is a situation that needs all of them. The garden and its numbers are made up.

A school garden club wants a phone app that spots ripe tomatoes in photos. They have 60 labeled photos, far too few to train an image model from scratch, plus 2,000 unlabeled ones, and volunteers can label perhaps 100. Two other schools want to help, but their photos show students in the background and must not leave the school. Next spring the app should learn peppers without getting worse at tomatoes. And nobody is sure which model or settings to use.

Each problem has a family of methods. None is magic: each moves the work somewhere else, and the table says where.

Five problems, six setups (the garden is made up)
The problemTraining setupCore ideaMain catch
Too few labeled photosTransfer learning, then fine-tuningStart from a model already trained on millions of other images and adjust it with your few photosHelps most when the old and new tasks look alike
Labeling takes people’s timeActive learningThe model asks for labels on the examples it is least sure aboutThe chosen examples are not a fair sample, so test on separate random data
New lessons push out old skillsContinual learningLearn new data while protecting what was learned beforeStill an open research problem for large models
The data must stay where it isFederated learningEach site trains locally and sends only model updates, which a server averagesUpdates can still leak information: not automatically private
Which model and settings?Evolutionary methods and AutoMLAutomated search: try, score, keep the best, change, repeatPeople still choose the goal and the score, and check the data

Going deeper · step 1

Don’t start from zero

A model that has already learned edges, colors and shapes from millions of photos knows most of what a tomato spotter needs.

Transfer learning means reusing what a model learned on one task for another. Take an image model that was pretrained on a large, general photo collection. Keep its early layers, which respond to edges, textures and shapes, and replace its last layer with a new one for your own labels: “ripe” and “not ripe.”

Then comes fine-tuning: a short extra round of training on your small dataset, usually with a small learning rate so the useful old knowledge is nudged rather than overwritten. Sometimes only the new last layer is trained and everything else stays frozen. PyTorch’s official tutorial builds an ants-versus-bees classifier this way from about 120 training photos of each.

Chatbots are built this way too (pretrained, then fine-tuned). Transfer helps most when the old and new tasks share structure, and the old model’s blind spots and biases come along with its skills.

3 · Try it

Ask for the labels that teach the most

Labels cost people’s time. If you can afford only a few, which examples should you ask about?

In active learning the model helps choose its own training data. It looks at the unlabeled pool, finds the examples it is least sure about, asks a person to label those, retrains and asks again. Burr Settles’s 2009 survey describes this strategy, uncertainty sampling, as the simplest and most commonly used.

Why would the unsure examples be better? A photo of a deep red tomato teaches the model little it does not already know. A tomato right on the border between orange and red pins down where the line goes.

Predict first

Twelve tomatoes have a redness score from 0 to 100. You know the palest one (6) is not ripe and the reddest (93) is ripe. You may ask about four more. Should you spread your four questions evenly, or keep asking about the tomato in the middle of what you still don’t know?

Show the answer

Keep asking in the middle. Each answer cuts the zone where the hidden boundary could be, like the number-guessing game “higher or lower.” In the lab, four well-placed questions shrink the zone from 87 points wide (6 to 93) to 4 points (50 to 54). Questions far from the boundary spend the budget confirming what the model already guessed.

Lab

Spend a budget of four labels

Each label you ask for costs a volunteer’s time. Choose which tomatoes to ask about and watch the model’s rule, its unsure zone and how many of the 12 it sorts correctly (checked against hidden answers).

Simulated in your browser with 12 made-up tomatoes. Nothing is learned from you or stored.

Or ask about a tomato yourself (each unlabeled tomato shows the model’s current guess):

Questions asked so far:

    Three challenges

    1. Press “next photo in order” four times. How wide is the unsure zone, and how many tomatoes does the rule get wrong?
    2. Reset and press “where the model is least sure” four times. Compare. Why is each question worth more?
    3. Reset and choose four tomatoes yourself. Can you beat the model’s strategy? Can you think of a case where asking only near the rule would mislead it?

    Twelve made-up tomatoes have redness scores 6, 12, 19, 27, 35, 46, 50, 54, 58, 61, 78 and 93. The hidden rule is “ripe from 52.” The model starts knowing only that 6 is not ripe and 93 is ripe, and may ask about four more. Its rule is “ripe from the midpoint between the reddest tomato known to be unripe and the palest tomato known to be ripe.”

    Strategy 1: ask in the order the photos were taken (19, 61, 12, 27, …)
    QuestionAsked aboutAnswerModel’s ruleUnsure zoneSorted correctly
    119not riperipe from 5619 to 93 (74 wide)11 of 12
    261riperipe from 4019 to 61 (42 wide)10 of 12
    312not riperipe from 4019 to 61 (42 wide)10 of 12
    427not riperipe from 4427 to 61 (34 wide)10 of 12
    Strategy 2: always ask about the tomato closest to the current rule
    QuestionAsked aboutAnswerModel’s ruleUnsure zoneSorted correctly
    150not riperipe from 71.550 to 93 (43 wide)9 of 12
    278riperipe from 6450 to 78 (28 wide)9 of 12
    361riperipe from 55.550 to 61 (11 wide)11 of 12
    454riperipe from 5250 to 54 (4 wide)12 of 12

    The same budget of four labels leaves a zone 34 points wide and 2 tomatoes wrong in photo order, but only 4 points wide and none wrong when the model asks where it is least sure. Labeling all ten would give the same rule as the four well-chosen questions.

    The catch. The examples active learning picks are deliberately unusual, so they are not a fair sample of the world. A model that only asks near its current rule can also miss a second region it never thought to doubt. Always test on a separate, randomly chosen set, as on Build a small model.

    Continual learning

    Learn peppers, keep tomatoes

    Next spring the club adds peppers. The obvious move has a surprising side effect.

    Predict first

    You fine-tune the tomato model on 200 pepper photos and nothing else. What happens to how well it spots ripe tomatoes?

    Show the answer

    It can get much worse, sometimes suddenly. Training adjusts the same parameters that held the tomato knowledge, and nothing in pepper-only data tells the model to keep them. This is catastrophic forgetting: a standard neural network trained on one task after another often loses the earlier one.

    Continual learning studies how to learn from a stream of new tasks or data without losing old skills. Three families of fixes are common. Replay mixes a sample of old examples into each new round of training. Protecting important weights makes the parameters that mattered most for the old task harder to change; Kirkpatrick and colleagues proposed this in 2017 as elastic weight consolidation. Adding new parts gives each new task some fresh parameters while freezing the old ones. Each one costs memory, computing or flexibility, and doing this well for very large models is still an open research question.

    It matters whenever the world drifts (see after a model is trained). A chatbot remembering your earlier messages is different: that text sits in its context window, and its parameters do not change.

    Federated learning

    Train together without pooling the photos

    Three schools want one better model, but each wants its photos to stay on its own computers.

    In federated learning, a server sends the current model to many devices or sites. Each one trains on its own data for a short while and sends back only its updated model numbers, which the server combines into a new shared model. McMahan and colleagues at Google proposed the standard recipe, federated averaging, in 2016, with phones in mind: their data is often private, large, or both. Sites with more examples count more, because the average is weighted by how many examples each one trained on.

    Worked example

    Federated averaging for three school gardens

    1. The setupEach school trains the tiny tomato model from the lab on its own photos. That model has one number: the redness score where “ripe” starts. School A has 20 labeled tomatoes and finds 54; School B has 50 and finds 50; School C has 30 and finds 53. Only these numbers and the counts are sent. No photo leaves a school.
    2. Weight by examples20 × 54 = 1,080. 50 × 50 = 2,500. 30 × 53 = 1,590. Sum: 1,080 + 2,500 + 1,590 = 5,170.
    3. Divide by all examples20 + 50 + 30 = 100 tomatoes. 5,170 ÷ 100 = 51.7. The shared model now says “ripe from 51.7.”
    4. Compare a plain average(54 + 50 + 53) ÷ 3 = 157 ÷ 3 ≈ 52.3. That would let School A’s 20 tomatoes count as much as School B’s 50.
    5. Next roundThe server sends 51.7 back, each school trains a little more from there, and the averaging repeats. A real model has millions of numbers; each is averaged the same way.
    6. Optional: add noiseEach school adds a small random number first (made-up draws: +0.4, −0.3, +0.2) and sends 54.4, 49.7 and 53.2. (20 × 54.4 + 50 × 49.7 + 30 × 53.2) ÷ 100 = (1,088 + 2,485 + 1,596) ÷ 100 = 51.69. The shared answer barely moves, but no school sent its exact result.

    All numbers are made up. Real differential privacy also limits how much one example can change an update and chooses the noise size mathematically; this only shows why averaging many noisy updates can still work.

    Federated is not automatically private. Keeping raw photos at each school helps, but an update is computed from the data, so it can still carry information about it, especially from a site with few examples. Differential privacy limits how much any single example can change an update and then adds carefully sized random noise, so the result reveals little about any one example. It costs some accuracy, and the noise size is a trade-off someone has to choose and justify.

    In practice

    An application, a limit and a connection

    Real application

    Phones and small projects

    Federated learning was designed for data that stays on phones, such as typed text. Fine-tuning a pretrained model is a common start for small image and language projects.

    Limitation

    The work moves, it doesn’t vanish

    Transfer needs a related model, active learning a person ready to answer, federated learning online devices and privacy protection, AutoML a clear goal and clean data.

    Connection

    Asking is a kind of exploring

    Active learning echoes the explore-or-exploit trade-off: both decide what to try next to learn the most. Evolution is a cousin of search and planning.

    4 · Common mix-ups

    What people get wrong about these methods

    Four ideas that trip people up, and what to think instead.

    • Mix-up: Federated learning is private by design.

      Better: It keeps raw data on the device, which helps. Updates can still carry information about the data, so real systems add protections such as differential privacy, and someone still has to check them.

    • Mix-up: Fine-tuning teaches a model from scratch.

      Better: Fine-tuning starts from what a pretrained model already learned and adjusts it. That is why it needs far less data, and why the old model’s gaps and biases can carry over.

    • Mix-up: AutoML means no people are needed.

      Better: It automates the search. People still define the goal and the score, check the data and labels, and decide whether the result is fit for use.

    • Mix-up: A chatbot that remembers our chat is doing continual learning.

      Better: Remembered text sits in the context window or a saved note. Continual learning changes the model’s parameters.

    5 · Check yourself

    Three quick questions

    Answer in your head or on paper before you open each one.

    1. A wildlife group has 5,000 unlabeled trail-camera photos and one volunteer who can label 200. Which two methods from this page would you combine, and how?

      Check your answer

      Transfer learning: start from a model pretrained on general photos and fine-tune it. Active learning: let that model pick the photos it is least sure about for the volunteer, retraining as labels come in. Keep a separate random set of photos for testing, because the chosen 200 are not a fair sample.

    2. Two sites train a shared model with federated averaging. Site X trained on 10 examples and sends 4.0; site Y trained on 30 and sends 8.0. What is the shared value?

      Check your answer

      (10 × 4.0 + 30 × 8.0) ÷ (10 + 30) = (40 + 240) ÷ 40 = 280 ÷ 40 = 7.0. A plain average, 6.0, would ignore that Y had three times as many examples.

    3. After fine-tuning only on pepper photos, the tomato spotter gets much worse at tomatoes. Name the problem and one fix.

      Check your answer

      Catastrophic forgetting. Fixes: replay some tomato photos while training on peppers, protect the weights that mattered most for tomatoes (as in elastic weight consolidation), or add new parameters for peppers and freeze the old ones.

    For younger learners

    A grades 5–8 version

    This page is for grades 9 and up. Younger classes can still feel the main idea of active learning.

    15 minutes · teacher-led

    Higher or lower?

    1. Write a secret number from 1 to 100 where students can’t see it.
    2. Team 1 must ask in a fixed order: “Is it 10 or more? 20 or more? 30 or more?”
    3. Team 2 may ask anything that has a yes or no answer. Both teams get five questions.
    4. Compare how close each team got.

    Talk about it

    Which question teaches the most?

    Good answers say Team 2 asked in the middle of what was still unknown, so every answer cut the possibilities in half. That is what an active-learning model does when it asks for a label.

    Honest limits

    What this page leaves out

    What this page leaves out

    Semi-supervised and self-supervised learning (learning from unlabeled data directly), meta-learning, the math of differential privacy, secure aggregation, neural architecture search in detail, and how large language models are updated over time. Each section here is a first look; the sources go further.

    Sources

    Where this comes from

    1. H. Brendan McMahan et al. (2016), “Communication-Efficient Learning of Deep Networks from Decentralized Data”: federated averaging, weighted by each device’s number of examples.
    2. Martín Abadi et al. (2016), “Deep Learning with Differential Privacy”: limiting each example’s influence and adding noise during training.
    3. Burr Settles (2009), “Active Learning Literature Survey,” Computer Sciences Technical Report 1648, University of Wisconsin–Madison: uncertainty sampling and other query strategies.
    4. James Kirkpatrick et al. (2017), “Overcoming catastrophic forgetting in neural networks” (PNAS): elastic weight consolidation.
    5. PyTorch tutorials, “Transfer Learning for Computer Vision”: fine-tuning a pretrained network on a small ants-and-bees dataset.
    6. scikit-learn user guide, “Tuning the hyper-parameters of an estimator”: grid search and randomized search with cross-validation.
    7. Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th edition, contents: evolutionary algorithms within search and optimization.

    AI words on this page

    Transfer learning, Pretraining, Fine-tuning, Active learning, Continual learning, Federated learning, Differential privacy, Evolutionary algorithm, AutoML, Hyperparameter, Label, Validation set, Distribution shift

    All definitions live in one place: AI words in plain language →

    Review status

    Written by ARQIV. Last checked October 2026. Independent subject review and classroom pilots are still to come, so treat the activities as drafts and tell us about anything that is wrong or unclear.