Understanding Convolutional Neural Networks: Theory, Architecture, and Practical Techniques

This article is the second part of a series on generative AI, focusing on Convolutional Neural Networks (CNN) and their application in image recognition.

Principles of CNN : A CNN mimics the way a child draws a stick figure—first capturing edges, then contours, then shapes, and finally combining them to recognize objects. The process consists of repeated convolution (feature extraction), pooling (feature simplification), and fully‑connected layers (feature synthesis).

Convolution uses filters (kernels) that slide over the input image, emphasizing edges and textures. Pooling reduces spatial dimensions, lowering computational cost while preserving important features. Multiple convolution‑pooling blocks increase abstraction depth.

Mathematically, a convolution kernel with size 3*3*1+1 has 10 parameters (weights plus bias). For a kernel of depth n , the parameter count is 3*3*n+1 . Example: a 2‑stack, 3×3 kernel has 3*3*2+1 = 19 parameters.

Implementation Example : The article demonstrates a CNN for American Sign Language (ASL) letter recognition. The dataset contains 784 columns per sample (1 label + 784 grayscale pixel values). After one‑hot encoding the labels and normalising pixel values, the data is reshaped to (samples, 28, 28, 1) .

Model architecture (illustrated in the original article):

Conv2D(75, (3, 3), stride=1, padding="same", activation="relu", input_shape=(28, 28, 1))

BatchNormalization()

MaxPool2D((2, 2), strides=2, padding="same")

Dropout(0.2)

Parameter calculation example: the first convolution layer has 75 * (3*3*1+1) = 750 trainable parameters; the second convolution layer (50 filters) has 50 * (3*3*75+1) = 33,800 parameters. Batch‑normalisation layers add non‑trainable statistics.

The trained model reaches about 92% accuracy on the validation set.

Data Augmentation : To further improve performance, the article introduces image data generators that apply random rotations (±10°), scaling (±10%), translations (±10%), and horizontal flips. Example code snippet:

train_datagen = ImageDataGenerator(rotation_range=10, width_shift_range=0.1, height_shift_range=0.1, zoom_range=0.1, horizontal_flip=True)

Training with the generator uses model.fit(train_generator, steps_per_epoch=len(x_train)/batch_size, epochs=20, validation_data=(x_valid, y_valid)) . This approach quickly boosts accuracy.

Transfer Learning & Fine‑Tuning : The article explains how to reuse a pre‑trained model (e.g., a Google image classifier) by removing its top layers ( include_top=False ), adding new dense layers, freezing the base ( trainable=False ), and training on the target dataset. If performance is insufficient, fine‑tuning is performed by unfreezing the base ( trainable=True ) and training with a very small learning rate.

Key take‑aways: CNNs excel at image tasks, pooling reduces over‑fitting, data augmentation enriches limited datasets, and transfer learning accelerates development when data are scarce.

The series will continue with Transformer fundamentals for natural‑language processing.