In plain language
Imagine sorting observations into two groups. Training finds that reduce mistakes on examples whose are already known.
01 / Classical machine learning
Patterns, not instructions.
Give a model examples, and it learns a rule. Change the problem and see where a simple boundary succeeds or fails.
Train a classical classifier
Run, pause and inspect each epoch0epoch
—train loss
—validation accuracy
Inspect a point. Color = predicted probability; shape = observed class.
The initial surface is untrained. Run the experiment to compute the full trajectory.
○ Class 0◇ Class 1 / train lossValidation loss
Ready to trainThe nine, side by side
Pick one, then watch it train belowNine classifiers, one plane
Pick a family, then watch it train100%training accuracy
98%validation accuracy
0.124log loss
Step through the training
Class 1Class 0Validation
Epoch 150 / 150The technical idea
This experiment fits logistic regression to two numerical inputs using and descent. Its probability boundary at 0.5 is a straight line. requires a nonlinear representation.
From data to proteins
A classical protein model might use charge, length or amino-acid composition as . A random forest or a kernel SVM can also model nonlinear relationships without being a deep network.
Look deeper: mathematics, methods & limitations
p(y=1|x) = σ(w·x + b)
For each example, compute a probability, measure against its , differentiate the and update the . One epoch here is one pass over the entire fixed training set.
The visual uses 100 synthetic training points and 60 independently generated validation points. Both are sampled from a chosen rule. The boundary heatmap shows predicted class probability; the symbols show training . Validation accuracy is separate from training .
A random split of highly similar proteins can leak family information between sets. For biological studies, consider sequence-identity or family-aware splits, an untouched test set, class imbalance and appropriate baselines. Accuracy alone can be misleading.