🧠 CIFAR-10 Image Classification

Tip: These models were trained on 32×32 CIFAR-10 thumbnails. They work best on simple, centred images of a single object — use the example images below for reliable results.

On real-world high-resolution photos the Custom CNN (trained on 32×32) often looks more plausible because its 32×32 resize pipeline matches its training distribution, while MobileNetV2 sees a 224×224 upscale it was never trained on. This is the classic domain gap — not evidence that the smaller model is better.

Example images — click to load

Compare All Deployed Models

Load an image above, then click the button to classify it with Custom CNN, MobileNetV2, ResNet-18 simultaneously.

Architecture Benchmark

All models were evaluated under a consistent training budget on the full CIFAR-10 dataset, using Adam, lr=0.001, 15 epochs, and batch size 128, with architecture-specific adaptations where required.

Deployed in this demo: Custom CNN, MobileNetV2, ResNet-18.
EfficientNet-B0 and ViT-Small are included in the table for study comparison only.

Key Findings

• ResNet-18 — 87.48% accuracy with only 5,130 trainable parameters (fewest of any deployed model).
• MobileNetV2 — 86.91% accuracy with 12,810 trainable parameters (frozen backbone + linear head).
• Custom CNN — 48.40% accuracy with 2.46M trainable parameters (trained from scratch — demonstrates the transfer-learning gap).
• ResNet-18 achieves +39.1 percentage points over the Custom CNN with ~480× fewer trainable parameters.

Top-5 Hardest Confusion Pairs

Bidirectional misclassification counts on the full 10,000-image test set, comparing the Custom CNN (before transfer learning) to MobileNetV2 (after).

CIFAR-10 Image Classification

How much does a pretrained backbone actually help compared to training from scratch — when every model shares the same dataset, optimiser, and epoch budget?

This project answers that question through a controlled, end-to-end deep learning experiment. Five architectures are trained and evaluated on the CIFAR-10 benchmark. Three are deployed here as a live demo. The goal is not simply to report accuracy — it is to surface the real trade-offs between model complexity, training strategy, parameter efficiency, latency, and generalisation.

Why I Built It

Transfer learning is often presented as an obvious win, but the degree of that advantage is rarely quantified under fair conditions. I wanted to run a structured experiment — identical training budget, same dataset, same optimiser family — and let the numbers speak. The result is a reproducible comparison that is useful both as a portfolio piece and as a practical reference for model selection decisions.

What This Project Demonstrates

• End-to-end ML pipeline — data loading, augmentation (RandomCrop, CutOut, MixUp, CutMix), training with cosine annealing, evaluation, and deployment
• Transfer learning vs from-scratch training — with measurable, empirical results
• Parameter efficiency — ResNet-18 achieves top accuracy with only 5,130 trainable parameters
• Deployment engineering — lazy model loading, device-aware inference, Gradio UI on HF Spaces
• Interpretability — Grad-CAM visualisations to understand model attention
• Benchmarking discipline — latency, throughput, model size, and confusion analysis alongside accuracy

Key Findings

• ResNet-18 reaches 87.48% accuracy with just 5,130 trainable parameters — a frozen ImageNet backbone with a single linear head.
• MobileNetV2 matches it closely at 86.91%, with a slightly larger head and a smaller total model footprint.
• Custom CNN, trained entirely from scratch, tops out at 48.40% over 15 epochs — a +39 percentage-point gap that makes the transfer learning advantage concrete and measurable.
• The transfer models converge in 1–3 epochs. The Custom CNN is still climbing at epoch 15.

⚠️ Limitations — The Domain Gap

All models were trained on 32×32 CIFAR-10 thumbnails. This has a direct impact on real-world performance:

• Uploading a high-resolution photo forces a resize that the models were never trained on.
• The Custom CNN is resized to 32×32, which matches its training pipeline — so it may look more confident on real images, but this is misleading.
• MobileNetV2 and ResNet-18 upscale to 224×224, which is outside their fine-tuning distribution.

For reliable results, use the provided example images or simple, centred, single-object photos against a plain background.

Tech Stack

About the Developer

Pouya Alavi Naeini is a final-year Information Technology student at Macquarie University, Sydney, specialising in Artificial Intelligence and Web/App Development. He builds practical software and machine learning projects with a focus on clean engineering, reproducibility, and deployment.

· GitHub · LinkedIn · Portfolio · Source Code