Learning to look beyond a classifier’s accuracy

Adapt the model to the dataset

Transfer learning provided an existing image representation to build on. I used rotation, flips, and color jitter to vary training images, along with dropout regularization to address overfitting.

The task still required evaluation on the target classes. K-fold cross-validation and confusion-matrix analysis helped examine performance rather than treating a successful training run as the end of the project.

Read the errors as part of the result

The classifier achieved 57.35% accuracy across the 17 classes. That number describes the overall outcome, but the confusion matrix and misclassified examples provide more detail about which distinctions were difficult.

This was a closed-set classification project: choosing among known classes in a particular dataset. Its result is useful in that context. It does not describe performance on new identities or establish a general identity-verification system.

The project expanded my experience beyond tabular models into image classification. The lasting lesson was to connect the reported metric to the task and inspect the errors before deciding what the model can do.