⚙️ Simulator: Full CNN Architecture
Input (12×12) → Conv → Pool → FC → Softmax
PIPELINE INPUT IMAGE (12×12)
🎯 Training (Forward Pass + Backpropagation)
Input ▸ Conv+ReLU ▸ Pool ▸ Flatten ▸ FC ▸ Softmax ▸ Prediction
Loss ◂ Optimizer ◂ Backpropagation ◂ (each step)
LOSS
ACCURACY
Loss: —
Accuracy: —
Training step: 0
The kernels and weights start random (no longer the fixed filters of the previous simulator). Each step, the network does the forward pass on the 3 images, computes the loss (cross-entropy) e a accuracy (how many of the 3 images are classified correctly), backpropagates the error, and adjusts all weights (including the kernels of the convolution) via gradient descent. ⚠️ Since the "training set" has only 3 examples, this demonstrates the mechanism of training (loss decreasing, accuracy increasing, weights changing) — not the ability to generalize to new images, which would require much more data.
Description: Figure 9.11: Interactive simulator of a CNN architecture: choose one of the 12×12 input images (house, happy face, or sad face), click on each block of the pipeline — Input, Conv+ReLU , Pooling , Flatten , FC , and Softmax — and run real training steps ( forward pass + backpropagation) to observe the loss and accuracy evolving, the kernels being adjusted, and the Softmax starting to point to the correct class.