What just happened
Connect your own training run to examples, features, learned numbers, and prediction.
Follow one photograph through the process
In the last reading you pressed a few buttons and got answers. This reading names what happened at each step, using words you will meet again in blocks and in Python. Your trained model should still be loaded. If you refreshed the page since then, press Open model in Test & export and choose the model file you downloaded, or simply press Train model again.
Imagine one photo of a green mug on a wooden desk. To the computer, it is a square grid of tiny dots called pixels, and each pixel is stored as a few numbers describing its colour. You put the photo in the mug class. The photo and its label together make one training example. The photo is the input. The label is the target: the answer we want the model to give for that input. A collection of examples is a dataset.
Build with me · 1
Find the input and the target in your own project
The model never sees the word "handle". It receives the grid of numbers and the target you attached. If you accidentally attach mug to a glass photo, training receives a glass image with a mug target; it cannot tell that you pressed the wrong button.
When you look at a mug photo, you might explain its label by pointing at the handle. The model is free to use anything in the numbers that lines up with the label: the handle, but also texture, position, lighting, or the background. That is why consistent labels are necessary but not enough. Varied photos are what stop an accidental detail from deciding the class.
Input: a 224 by 224 grid of colour numbersTarget: mugTogether: one training exampleAll of your Training photos: the training datasetOpen Photos, keep Training selected, and look at one photo from each class. Say out loud what the input is and what the target is. Check that neither photo shows one of your reserved objects.
What to look for
You can point to the photo as the input, the selected class as the target, and the whole Training collection as the dataset, without saying the model was given a rule about handles.
Make it yours
Describe a photo that is correctly labelled but carries little evidence, such as a mug with its handle cropped away. Decide whether your label rule allows it, and write the decision in Notes.
The model had a head start
Fifty photos per class is a tiny dataset. The model still learned something useful from it, because it did not start from nothing.
The image lab begins with a pretrained network. A network (short for neural network) is a kind of model built from layers of learned numbers. Pretrained means someone else already trained it, on a very large collection of everyday photos, before your project began. That training taught it to turn any photo into a list of numbers describing what is in it: edges, curves, textures, and combinations of them. Each of those numbers is called a feature.
Your photos did not change that part. When you pressed Train model, the lab trained only a small extra piece on top, called the classifier. The classifier learned how the features of your mugs differ from the features of your glasses. Reusing a model trained for one job as the starting point for a new job is called transfer learning. It is why a project with a few dozen photos can work at all.
Build with me · 2
Draw the path from photo to answer
A feature does not have to match a word a person would use. We should not pretend that one number means "has a handle" unless we have evidence. And your photos are not the model's whole visual education: most of its ability to see came from the pretrained part.
In Level 4 you will train a small network for handwritten digits that starts from random numbers instead of a pretrained one. The setup is different, but the roles are the same: inputs, targets, learned numbers adjusted by training, and an honest test.
224 by 224 photo → pretrained network (learned before your project) → features: a list of numbers describing the photo → classifier (trained on your Training photos) → a score for mug and a score for glassIn Notes, write the arrows from the example in your own words. Under the pretrained network, write 'learned before my project'. Under the classifier, write 'trained on my Training photos'.
What to look for
Your notes separate the part that arrived ready-made from the part trained on your photos. Development and final-test photos appear in neither.
Make it yours
Explain in one sentence why the head start makes a small project practical. Then name one way it could still fail on your objects, such as a kind of glass unlike any in your training photos.
Guess, compare, adjust
Level 1 followed one training step by hand: make a guess, measure the error with a loss, adjust a number. Training your image model did the same thing many times over.
For each training photo, the classifier produced a score for mug and a score for glass. A loss turned the gap between those scores and the target into a penalty. A mug photo scored as 70% glass gets a bigger penalty than one scored as 95% mug. The training procedure then nudged the learned numbers inside the classifier, which Level 1 called parameters (people also call them weights), in the direction that makes the penalty smaller. One nudge rarely fixes much, so this repeats over all your photos, round after round. The loss shown in the status line during training is that penalty, averaged over the training photos.
What you end up with is a set of numbers, not a list of definitions. The model may rely on shape, background, colour, or several signals at once. Its answer alone does not tell you which one it used.
Build with me · 3
Name the action: train, predict, or evaluate
Three actions are easy to blur together, so give each one its own name.
Train changes the model. It happens only when you press Train model, and it uses only Training photos.
Predict uses the model as it is. A photo goes through the pretrained network and the classifier, and scores come out. No correct answer is needed, and the model does not change. Prediction is also called inference.
Evaluate compares predictions with answers you already know, and counts how often the model was wrong. Without known answers you can still predict, but you cannot evaluate.
Here is a test case. A friend shows your model a cup, gets mug, and never presses Train model. The friend has predicted. The model has learned nothing about that cup. To evaluate the answer, you would also need the correct label under your rule.
Train Train model: the learned numbers changePredict Predict collection: scores and an answer; nothing changesEvaluate compare answers with known labels: correct or incorrectIn Test & export, predict your Development collection again. For each result, say which action produced the answer and which produced the correct or incorrect mark. Then open Train: the counts have not changed.
What to look for
You can say that predicting the Development collection is prediction followed by evaluation (development photos carry known answers), and that neither step changed the model.
Make it yours
If you added an Unlabelled targets photo earlier, predict that collection. You get a prediction but no correct or incorrect mark. Explain why in one sentence.
Why the first test is limited
Your first development photos were probably taken in the same room, with the same camera and light, as your training photos. Doing well on them is encouraging, but it is weak evidence about someone else's mug in someone else's kitchen.
The first predictions do not have to be perfect either. A badly framed photo, a label that broke your rule, too little variety, or a failed run can all cause early mistakes. Write down what you saw either way. The next module builds a proper test.
Full reference solution
The whole Level 2 workflow on one page. Each reading covers part of it; this list shows where that part fits. The numbers in the readings are examples, and your own results will differ.
1. Name the two classes and write a one-sentence rule for each.2. Reserve objects or sessions for Development and Final test before collecting.3. In Photos, choose Training and a label, then add varied photos of both classes.4. Add reserved photos under Development and Final test, never under Training.5. In Notes, record label rules, objects, conditions, the split, and a run name.6. In Train, check both counts, keep 20 rounds for the first run, and press Train model.7. Download project and Download model, with matching run names.8. In Test & export, predict Development and read every ID, actual label, prediction, and score.9. Work out accuracy yourself, read the confusion matrix, and compare with the Always- baseline.10. Write one hypothesis. Change only the Training photos (or one stated setting), retrain, and predict the same Development collection.11. Choose a version using Development results. Open its saved model if needed.12. Predict Final test once, after choosing, and report its fraction, conditions, and limits.13. For photos with no known answer, use Unlabelled targets, predict, and Download predictions.csv.14. Keep project, model, notes, and predictions file together under one run name. None of them replaces another.Compare this with your version. Different names and personal choices are fine when the program follows the same logic.
Keep your progress
Sign in and every reading, quiz, and exercise you finish is saved.