Step 5 · Classification
AlexNet
The baseline, used to understand how a network extracts features from this data.
The baseline, built to study feature extraction
From the paper
What we built
The AlexNet-based model was used to examine feature extraction and understand how the model operates. It incorporated DepthwiseConv2D layers, adaptive learning-rate scheduling and early stopping, mixed-precision training and the AdamW optimizer.
- DepthwiseConv2D
- Mixed precision
- AdamW
- Label smoothing
- Adaptive learning rate
- Early stopping
- Eye photo in
- Conv 11×11, 96 filtersstride 4
- Batch norm + max pool
- Depthwise conv 5×5
- Batch norm + max pool
- Conv 3×3 ×3384 · 384 · 256
- Max pool → global average pool
- Dense 4096 → dropout
- Dense 4096 → dropout
- Softmax, 5 classes
- EsotropiaExotropiaHypertropiaHypotropiaNormal
Published results
How AlexNet did on the 81 test photos
- Accuracy
- 59.26%
- Precision
- 58.88%
- Recall
- 59.26%
- F1 score
- 58.71%
| Predicted class → | |||||
|---|---|---|---|---|---|
| True ↓ | Esotropia | Exotropia | Hypertropia | Hypotropia | Normal |
| Esotropia | 0.73 | 0.07 | 0.07 | 0.00 | 0.13 |
| Exotropia | 0.00 | 0.65 | 0.06 | 0.18 | 0.12 |
| Hypertropia | 0.19 | 0.19 | 0.44 | 0.19 | 0.00 |
| Hypotropia | 0.12 | 0.12 | 0.12 | 0.44 | 0.19 |
| Normal | 0.06 | 0.06 | 0.06 | 0.12 | 0.71 |
Each row is a true class; the numbers show where its test photos ended up. The diagonal is the share classified correctly.
It was strongest on esotropia (73% correct) and weakest on hypertropia (44% correct).