Image classification with transfer learning and CNNs.
ML4Schools image labs support a trainable head on a frozen feature extractor and a small convolutional neural network built from scratch. The lesson supplies the image pack and the experiment; the student investigates the model.
Reuse features or learn from prepared pixels.
Image classification assigns an image to a category. The lab supports two ways to build that decision, with different choices and computational costs.
| Approach | What is trained | What students can investigate |
|---|---|---|
| Frozen base + trainable head | A pretrained feature extractor stays fixed. Dense and dropout layers in the head learn from its feature vectors. | How changing the head changes the decision without retraining the whole feature extractor. |
| Small CNN from scratch | A classroom-sized convolutional network learns from the prepared image pixels. | How convolution, pooling and the final classifier work together, and how architecture affects training. |
The CNN layer palette includes convolution, max pooling, global average pooling, flattening, dense layers and dropout. The lab checks shapes and layer order before training. It is a bounded teaching environment, rather than an unrestricted architecture editor.
A supplied-model activity can instead focus on inspection and evaluation. The activity determines which building and training choices students actually receive.
Compare runs, not just the final label.
Image training runs use TensorFlow.js in the browser. Students can change an allowed build and training settings, then compare the run's training and validation results. Inspect reads the chosen run's model rather than a stock illustration of a network.
A frozen feature extractor lets a small head train without learning every image feature again. A CNN from scratch has more work to do. Training speed depends on the device and the build; CPU-only CNN training can take minutes per epoch.
Browser training does not mean that all learning records stay only on the device. The portal saves learning work, and model storage is described separately in the student privacy explanation.
Validation images are not an independent test.
The current image Inference Lab evaluates a supplied catalogue model on the image pack's validation cohort. The pack's reserved test images are not served to students for browser evaluation.
It does not treat images already used to validate a student's training run as an independent test of that run. In that situation, image inference uses the supplied model rather than presenting the student's validation results as new test evidence.
This distinction matters: a validation result, a reserved-test result and a prediction on an unlabelled image support different conclusions. An image prediction alone does not reveal whether it is correct.
Change an image and observe the answer.
Where enabled by the lesson, the Inference Lab can change brightness, contrast, rotation, blur, noise or crop, then score the changed image. Students can investigate whether a prediction is sensitive to changes that seem unimportant to a person.
For supported models with a frozen feature extractor and a compatible classification head, an explanation can include a Grad-CAM-style map, feature activations and model structure. A heatmap highlights sensitivity in that model; it is not proof that the system understands the object as a person would.
The image explanation is not available for every architecture. In particular, the current explanation path does not provide the same Grad-CAM view for a student's from-scratch CNN. The page should not be read as a promise of a universal explanation method.
Which changes should matter to this task?
- What stays fixed when training a head on a frozen feature extractor?
- How does the cost of training a CNN compare with training the head?
- Which image changes alter the prediction, and is that sensitivity desirable?
- What does a validation result establish, and what evidence is still missing?
Read more about thresholds, model explanations and evaluation limits. Image activities are teaching experiments; the Brain MRI example is teaching data, not a diagnostic tool.
Compare the experiment as well as browser training.
Free model-building tools can also offer no-code classification and browser training, so browser training alone is a shared capability. In ML4Schools, evaluate how the offered head/CNN choices, image changes and interpretation tasks fit a prepared lesson. The scope and evaluation limits of the chosen activity matter more than a generic claim of advanced AI.
Read the sourced comparison with free AI teaching options and see when each approach may fit.