Convolution Layers
Scan small image regions to detect edges, curves, and stroke fragments.
How CNNs Work Here
Think of the CNN as a smart pattern reader for handwriting. It looks at tiny local parts of the image first (small line segments and curves), then combines them step by step into larger patterns that represent full Greek letters.
Convolution layers are the pattern detectors. They use many small filters (also called kernels) that slide over the image and react to specific patterns like edges, corners, and curved strokes. In this project, these filters are 3×3, so each detector looks at a tiny local neighborhood and then the network stacks many such detectors to understand larger structures.
Pooling layers summarize nearby activations and make the model more robust to small shifts in position. In ML terms this is called translation invariance: the model can still recognize a letter even if it is drawn a bit left/right/up/down.
The final classifier converts these learned features into probabilities for the 24 small Greek letters. We test multiple architectures because there is no universal best model: smaller models are faster and need fewer resources, larger models can learn more complex details but may overfit. Comparing them helps find the best accuracy/stability tradeoff for this specific dataset.
Pooling Layers
Condense nearby pixels and add translation invariance, so tiny shifts matter less.
Dropout + Weight Decay
Regularize the model to reduce overfitting and improve robustness on unseen writing.
Small vs. Large Models
Small models are efficient; larger models can capture more detail. We benchmark both.
Why Compare Architectures
The best setup depends on your data, not theory alone. We rank candidates by validation performance.
Architecture Comparison
This plot is loaded from backend artifacts and updates after retraining.
Try It Yourself
Draw a small Greek letter and run model inference.
No prediction yet.