LeNet – 5-History
•Gradient-Based Learning Applied to Document
Recognition
•Y.LeCun, L Bottou, Y.Bengio an dP.Haffner
Proceedings of the IEEE, 86(11):2278 – 2324,
November 1998
CNN General Formula
•Training phase = learn weights from many images.
• Prediction phase = use those learned weights to classify new input images.
7.
LeNet-5 Architecture twomain parts
Let’s break it into two phases:
1. Training Phase
• You give many labeled images (say handwritten digits 0–9).
• LeNet-5 does forward pass (conv → tanh → pooling → fully connected →
output).
• The output is compared with the ground truth label (like "this image is digit
5").
• Error (loss) is calculated.
• Backpropagation adjusts the weights and biases so that next time the
prediction is closer to the truth.
• After repeating this on thousands of images, the model learns good
weights.
8.
LeNet-5 Architecture twomain parts
2. Prediction (Inference) Phase
• Now training is finished, weights & biases are fixed.
• When you give a new input image (say a handwritten "3"), the
network runs forward only:
• input → conv → tanh → pooling → fully connected → output scores
• At the last layer (10 neurons), the highest score indicates the digit.