A 784→60→10 neural network trained on 60,000 MNIST handwritten digits. Sigmoid activation, stochastic gradient descent, 30 epochs. Test accuracy: 97.2%.
Upload a handwritten digit — any filename works
Your image
28×28 translation
step 2 — inputs
2 784 pixel values, 0–255
Each pixel is converted to a brightness value, then inverted so that ink = high (up to 0.9) and white background = 0. The network receives all 784 of them at once — nothing else.
Input matrix — 28×28 (hover any cell to see its value)
step 3 — one neuron
3 How a hidden neuron is calculated
Each of the 60 hidden neurons looks at every pixel and computes a weighted sum: it multiplies each pixel value by a trained weight, then adds them all up. The weights were learned during training — a large positive weight means "this pixel being dark pushes me toward active"; a large negative weight means "this pixel being dark pushes me toward silent." Finally, the sum is passed through a sigmoid function, which squashes it to a value between 0 and 1.
Green terms push the neuron toward active (higher output). Red terms push it toward silent (lower output).
Neuron 1 — top 10 inputs by strength of contribution
step 4 — the full network firing
4 Hidden and output neurons
Color shows activation strength: white = silent (0), orange = fully active (1). Hover any hidden neuron to see its activation score. Hover an output neuron to highlight its 5 most influential hidden neurons. Click any neuron to see its full calculation.
Note: output scores are raw sigmoid values, not softmax probabilities — they don't sum to 1.
Hidden Neurons
Output Neurons
silent (0.0)fully active (1.0)
step 5 — verdict
5 The network's prediction
Architecture: 784→60→10, sigmoid activation, SGD, lr=3.0, batch=10, 30 epochs.
Training data: MNIST (LeCun et al., 1998) — 60,000 training images, 10,000 test images. Test accuracy: 97.2%.
Part of AI Literacy curriculum. Inside the Machine — DSC234, University of New England.