◆ QUEST LOG / GENDER CLASSIFICATION

Three CNNs walk into a bake-off

2026Computer visionTensorFlow / KerasTransfer learningCelebA

Same data, same training budget, three architectures — and a decision framework instead of a single accuracy number.

01The task

Binary gender classification on the CelebA celebrity faces dataset — a classic benchmark task, which is exactly why it makes a good controlled experiment. The question wasn’t “can a CNN do this” (it can) but which architecture earns its compute when you hold everything else constant.

02The contenders

Three ImageNet-pretrained architectures, each fine-tuned the same way:

VGG-19 — the heavyweight classic: simple, deep, and expensive.
InceptionV3 — multi-scale filters designed for efficiency.
ResNet50 — skip connections, the default choice of its era.

Experiment pipeline: CelebA preprocessing and augmentation feeding three transfer-learning paths (VGG-19, InceptionV3, ResNet50) into a shared training configuration and evaluation — One pipeline, three transfer-learning paths. Identical preprocessing, augmentation, and training configuration, so the architecture is the only variable.

03Results

InceptionV3 led at ~91.5% accuracy, but the ranking is the least interesting part. The comparison covered precision and recall per class, confusion matrices, and inference time — because a model that misclassifies one class disproportionately, or takes twice as long per image, can lose in production even when it wins on the leaderboard.

Confusion matrices for VGG-19, InceptionV3, and ResNet50 on the CelebA test set — Test-set confusion matrices. InceptionV3 leads at ~91.5% (343 of 375 correct), and each architecture fails in a different direction.

Example VGG-19 predictions: a true positive, true negative, false positive, and false negative with confidence scores — What the errors look like: VGG-19 was 99.9% confident on a false negative. High confidence and correctness are different things — a theme that stuck with me.

Read the full report (PDF) ↗

04What I learned

Transfer learning makes the first 90% almost free — the craft is in the evaluation. Designing the comparison (same splits, same augmentation, same budget) mattered more than any architectural choice, and reading confusion matrices taught me more about the models than the accuracy scores did. This project is also where my habit of benchmarking cost alongside quality started — it shows up again in my NLP and LLM work.