Deep Learning and CNNs: Introduction

One Million Cat Photos

In 2012, a research team at the University of Toronto fed a deep convolutional neural network roughly 1.2 million labeled images from ImageNet. Nobody programmed what a cat looks like. Nobody defined "fur," "whiskers," or "pointy ears." They simply showed the network millions of labeled photographs and let it figure out the rules itself.

The network AlexNet shattered every previous benchmark for image recognition. It did not just win the competition; it won by a margin that stunned the research community. The era of deep learning had arrived.

No rules were programmed. No domain experts were consulted. Just labeled data, a smart architecture, and compute. That is the deep learning paradigm: learn hierarchical representations directly from raw data at scale.

What Makes Deep Learning "Deep"?

Traditional machine learning requires humans to engineer features. For image classification, a human expert might manually extract edges, corners, and textures before handing them to a classifier.

Deep learning eliminates this step. The network learns features automatically, layer by layer, in a hierarchy:

• Early layers learn primitive patterns: edges, color gradients, corners.
• Middle layers combine primitives into shapes: circles, curves, textures.
• Deep layers combine shapes into semantic concepts: eyes, wheels, faces, cats.

The more layers, the more abstract and powerful the learned representations. A 150-layer network can represent incredibly complex features that no human would have thought to engineer.

Convolutional Neural Networks (CNNs)

A standard fully connected MLP applied to a 224×224 color image would require processing 224 × 224 × 3 = 150,528 input values in the first layer alone. Each neuron connects to every input. The parameter count explodes and training becomes infeasible.

CNNs solve this with three key components:

Convolution Layer

Instead of connecting each neuron to every pixel, a convolutional filter (kernel) slides over the image typically a small window like 3×3 or 5×5 pixels. At each position, it computes a dot product with the pixels underneath. This produces a feature map that highlights where a particular pattern (edge, curve, texture) appears.

Key advantages:

• Parameter sharing: The same filter is applied everywhere. A 3×3 filter has only 9 weights regardless of image size.
• Translation invariance: If a cat appears on the left or right side of the image, the same filter detects it.

Multiple filters per layer detect different features simultaneously.

Pooling Layer

Pooling reduces the spatial size of feature maps. Max pooling (most common) takes the maximum value in each small window (typically 2×2), reducing each dimension by half.

This provides approximate translation invariance and drastically reduces computation. A 224×224 feature map becomes 112×112 after one max pooling layer.

Fully Connected Layer

After several convolution + pooling layers, the resulting feature maps are flattened into a 1D vector and fed into standard fully connected layers. These layers combine the learned spatial features to make the final class prediction.

Why CNNs Work: Spatial Locality and Weight Sharing

Images have a fundamental property: nearby pixels are related. The color of one pixel tells you something about its neighbors. A 3×3 convolution exploits this it looks at local regions, not random pairs of pixels across the image.

Weight sharing means the same filter is reused across all positions. This reduces parameters by orders of magnitude. A VGG-16 network processes 224×224 images with 138 million parameters. Without weight sharing, the first layer alone would require billions.

Transfer Learning: Standing on the Shoulders of Giants

Training a deep CNN from scratch requires millions of labeled images and days of GPU time. Most projects have neither.

Transfer learning solves this: use a network already trained on a massive dataset (ImageNet, with 1.2 million images and 1,000 classes) and adapt it to your task.

Two approaches:

Feature extraction: Freeze all pretrained layers. Remove the final classification head. Pass your images through the frozen network and use the output features as input to a new, small classifier trained on your data. Works well with small datasets (<1,000 images).

Fine tuning: Load pretrained weights, replace the final layers, and continue training the entire network (or just the later layers) on your data with a small learning rate. Works better with moderate datasets (1,000–100,000 images).

Popular pretrained models: ResNet (deep residual networks, up to 152 layers), EfficientNet (balanced depth/width/resolution scaling), VGG (simple and widely understood), MobileNet (optimized for mobile devices).

Simple Keras CNN for MNIST

import numpy as np
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.datasets import mnist

# Load MNIST: 60,000 training images, 10,000 test images, 28×28 grayscale
(X_train, y_train), (X_test, y_test) = mnist.load_data()

# Reshape and normalize to [0, 1]
X_train = X_train.reshape(-1, 28, 28, 1).astype('float32') / 255.0
X_test  = X_test.reshape(-1, 28, 28, 1).astype('float32') / 255.0

# One-hot encode labels
y_train = keras.utils.to_categorical(y_train, 10)
y_test  = keras.utils.to_categorical(y_test, 10)

# Build CNN (8 lines)
model = keras.Sequential([
    layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)),
    layers.MaxPooling2D((2, 2)),
    layers.Conv2D(64, (3, 3), activation='relu'),
    layers.MaxPooling2D((2, 2)),
    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.Dropout(0.5),
    layers.Dense(10, activation='softmax')
])

model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
model.summary()
# Total params: 93,322

# Train for 5 epochs
history = model.fit(X_train, y_train,
                    epochs=5, batch_size=64,
                    validation_split=0.1, verbose=1)
# Epoch 5/5
# 844/844 — 8s — loss: 0.0521 — accuracy: 0.9843 — val_accuracy: 0.9912

# Evaluate on test set
test_loss, test_acc = model.evaluate(X_test, y_test, verbose=0)
print(f"\nTest Accuracy: {test_acc:.4f}")
# Output: Test Accuracy: 0.9912
# 99.12% accuracy with a simple 8-layer CNN trained in under 60 seconds

99.12% accuracy on handwritten digit classification. A human expert struggles to beat 99.5%. This small CNN, trained in under a minute on a laptop GPU, approaches human performance.

Applications of CNNs

• Image Classification: What is in this image? (ImageNet, medical imaging)
• Object Detection: Where are the objects, and what are they? (YOLO, Faster R CNN)
• Semantic Segmentation: Classify every pixel. (autonomous vehicles, satellite imagery)
• Medical Imaging: Detect tumors, diabetic retinopathy, COVID in chest X rays
• Face Recognition: Identify individuals from photos
• Video Analysis: Action recognition, anomaly detection in surveillance

Traditional ML vs Deep Learning

Key Takeaways

• Deep learning learns hierarchical features automatically no manual feature engineering needed.
• CNNs use convolution (local filters), pooling (spatial reduction), and fully connected layers for image tasks.
• Parameter sharing and spatial locality make CNNs efficient and effective on images.
• Transfer learning lets you use pretrained networks (ResNet, EfficientNet) for your own tasks with very little data.
• For images and audio, deep learning dominates. For tabular data with limited samples, traditional ML usually wins.
• Always start with the simplest model. Reach for deep learning only when simpler approaches fall short.