Epoch — one full pass through all 60,000 training images. Each epoch the network sees every digit once.
Batch — images are processed in groups of 128. Each batch triggers one round of weight updates.
Test Accuracy — measured on 10,000 images the network has never trained on. This is the real scorecard.
W1/W2/W3 Mean — the average weight value in each layer. These should stay small and stable. If they explode or collapse to zero, training has gone wrong. 🔊

Loss starts high, around 2.3, which is what you get from random guessing across 10 classes. It drops rapidly in the first epoch as the network learns the obvious patterns, then improves more slowly. Accuracy follows the mirror pattern: it jumps to around 83% after just one epoch, then climbs toward 95% over the remaining epochs. The steepest learning happens in the first few thousand batches — watch the curves flatten as the easy wins run out and the network fine-tunes the harder cases. 🔊

When you release the mouse, your 280×280 drawing is downsampled to 28×28 pixels, normalised to values between 0 and 1, then flattened into a 784-number vector — exactly the same format as the MNIST training images. That vector passes through Hidden Layer 1, then Hidden Layer 2, then the Output Layer. Softmax converts the 10 output scores into probabilities. The digit with the highest probability is the prediction, and the bar chart shows you how the confidence is distributed across all 10 options. 🔊

Draw centred and fairly large — the MNIST digits fill most of the 28×28 frame. A tiny digit in the corner will confuse the network because it's never seen that layout during training. Use a thick stroke; thin wispy lines lose detail when scaled down to 28 pixels. If the network gets it wrong, try again — sometimes a slightly different stroke style is all it takes. An untrained network will guess randomly, so train first for meaningful predictions. 🔊

🔊

Before training, every neuron starts with random weights — the grid looks like TV static. Nobody taught the network what to look for. But as training runs, each neuron self-organises to detect a specific visual feature: maybe a horizontal stroke, a curved edge, or a particular corner shape. You're watching the network invent its own visual alphabet from scratch, purely from seeing thousands of handwritten digits and being told whether it guessed right or wrong. No human designed these features — the maths discovered them. 🔊