01 Can each class probability be interpreted independently?
Description
Softmax converts class scores to probabilities that sum to 1, so changing one score also affects the other probabilities.
Multiclass Classification
English translation
Understanding the relationship between class scores, probabilities, and ground truth labels helps you interpret multiclass classification errors and incorrect labels.
Before practice
02 How does one incorrect label affect training and evaluation?
Description
Training follows the modified labels, but errors against the original labels may increase.
Practice
Preview practice notebookDownload practice notebook1. Comparing multiclass logits and predictions
Starting conditions
Generate 6 samples per class, 18 in total, around the provided centers (-2,0),(2,0),(0,2). Use seed=7.
Procedure
- Check the input and label shapes for
Linear(2,3)and CrossEntropy. - Record the 3 logits, 3 probabilities, and prediction for each of the first 6 samples.
- Plot a scatter plot of the data and a loss curve.
Results to retain
- A 6-row output table and plots of the data and loss.
Items to record and points to check
- Samples
- Answer
- 3 logits
- 3 probabilities
- Prediction
Points to check
- Do probabilities sum to 1 for each sample, and do logits and probabilities have the same argmax?
- Did you use integer indices for class labels?
2. Results by class and ambiguous inputs
Starting conditions
Pass query=[(-2,0),(2,0),(0,2),(0,0)] to the trained model.
Procedure
- Create a 3×3 confusion matrix for the 18 training samples.
- Record the probabilities and predictions for the 4 query points in a table.
- Compare the probability distributions for the cluster centers and (0,0).
Results to retain
- A confusion matrix and a query table.
Items to record and points to check
- query
- P0
- P1
- P2
- Prediction
- Interpretation
Points to check
- Does the confusion matrix sum to 18?
- Did you avoid interpreting the maximum probability as an absolute guarantee of the model's reliability?
3. Data with one changed label
Starting conditions
Change only sample 0's label from 0 to 1, and retrain under the same conditions.
Procedure
- First mark the changed label and its position.
- Compare the baseline and modified models in terms of loss, errors against original labels, and errors against modified labels.
- Explain the training difficulty by distinguishing issues with the iteration count from issues with the data.
Results to retain
- A comparison table of the two models and a brief interpretation.
Items to record and points to check
- Model
- Final loss
- Error count against original labels
- Error count against modified labels
Points to check
- Did you keep the input data and initialization unchanged?
- Did you specify which labels were used as the evaluation reference?