01204563 Advanced Machine Learning · Topic 7

10×10 Bitmap Autoencoder

100 input pixels → small hidden representation → 100 reconstructed pixels

100–2–100
Epoch 0 Training weighted BCE — Interactive controls ready

Training set: digits 0–9

One epoch processes all ten 10×10 bitmaps once. Plain mode: input = target.

Fully connected autoencoder

Encoder: hⱼ = sigmoid(zⱼ), where zⱼ = Σᵢwᵢⱼxᵢ + bⱼ
Decoder: x̂ₖ = sigmoid(Σⱼvⱼₖhⱼ + cₖ)
x
input pixel
w, v
connection weights
b, c
biases (adjustable offsets)
z
weighted sum before activation
h
hidden value after sigmoid
x̂
reconstructed output pixel
Hidden-feature weights
negative0positive

Input bitmap x

Choose Draw or Erase, then drag across cells. Reconstruction updates when you release.

Reconstruction x̂

Input MSE —

Closest training template: —

Normalized similarity —

Similarity histogram: digits 0–9
Normalized similarity probability—not a calibrated classifier probability. Labels are not used to train the autoencoder.