← AI Engineering / Case study 07
CIFAR-10 CNNs.
Object recognition with convolutional neural networks: ten deliberate iterations, each one testing a single hypothesis, from a 55.7% baseline to a 93.2% ensemble.
- Team
- Pavitra, Rima, Mohamed
- Context
- Ironhack AI Engineering
- Year
- 2026
- Best result
- 93.21% test accuracy
Project presentation · use ← → or swipe
Iterations
This is where my biochemistry training shows: change one thing, measure, keep what works. We logged every model version in a tracking sheet.
| Model | What changed | Test acc. |
|---|---|---|
| #1 | Baseline, 2 conv layers | 55.70% |
| #2 | + BatchNorm & Dropout | 62.63% |
| #3 | Optimised for 32×32 input + BN | 82.15% |
| #4 | 6 conv layers + L2 regularisation | 81.00% |
| #6 | 13 conv layers + Global Avg Pooling | 93.03% |
| #7 | Contrast augmentation + AdamW (best single) | 93.12% |
| #9 | EfficientNetB0 transfer learning | 91.35% |
| #10 | Final ensemble | 93.21% |
Cats and dogs stayed the hardest pair to tell apart, as the confusion matrices show. Interestingly, transfer learning did not beat our own well-regularised CNN on these tiny 32×32 images.