Deep Learning Basics: CNN vs. RNN and a Hands-On MNIST Example (Lecture 3)

This is Lecture 3 of our AI 101 series. We’ll explain what Deep Learning is, compare CNNs and RNNs, and finish with a verified TensorFlow/Keras lab where you build a CNN to classify MNIST handwritten digits.

Table of Contents

{% toc %}

1) What Is Deep Learning?

Deep Learning is a subset of Machine Learning that uses multi-layer artificial neural networks to model complex patterns in data—especially effective for images, audio, and text.

1.1 Neural Network Basics

A neural network consists of:

• Input Layer: Where data enters the model.
• Hidden Layers: Layers of neurons that transform the data.
• Output Layer: Produces the final prediction.

Example: An image → processed through multiple layers → classified as “cat” in the output.

2) Deep Learning vs. Machine Learning

3) CNN vs. RNN

3.1 CNN (Convolutional Neural Network)

• Designed for image processing.
• Uses convolutions to automatically extract visual features like edges, colors, and textures.
• Typical structure: Convolution Layer → Pooling Layer → Fully Connected Layer.

Applications:

• Image classification
• Facial recognition
• Road scene understanding in autonomous driving

3.2 RNN (Recurrent Neural Network)

• Designed for sequential data (time series, language).
• Passes the output from one step as input to the next—adding a memory component.
• Variants: LSTM, GRU (handle long-term dependencies).

Applications:

• Natural Language Processing (translation, chatbots)
• Stock price prediction
• Speech recognition

4) Typical Deep Learning Workflow

1. Data Preparation: Gather and format images, text, or audio.
2. Preprocessing: Normalize values, tokenize text, etc.
3. Model Design: Choose CNN, RNN, or other architectures.
4. Training: Use GPU if available for faster convergence.
5. Evaluation: Measure accuracy, loss, and other metrics.
6. Deployment: Integrate into production systems.

5) Hands-On Lab: CNN for MNIST Digit Classification

We’ll use TensorFlow/Keras to build a small CNN that classifies MNIST handwritten digits (0–9).

Setup

1	pip install tensorflow

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

import tensorflow as tf
from tensorflow.keras import layers, models

# 1. Load data
mnist = tf.keras.datasets.mnist
(X_train, y_train), (X_test, y_test) = mnist.load_data()

# 2. Preprocess data
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

# 3. Build model
model = models.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(64, activation="relu"),
    layers.Dense(10, activation="softmax")
])

# 4. Compile
model.compile(optimizer="adam",
              loss="sparse_categorical_crossentropy",
              metrics=["accuracy"])

# 5. Train
model.fit(X_train, y_train, epochs=3, validation_data=(X_test, y_test))

# 6. Evaluate
test_loss, test_acc = model.evaluate(X_test, y_test)
print("Test accuracy:", f"{test_acc*100:.2f}%")

Expected Output

1
2
3
4

Epoch 1/3
1875/1875 [==============================] - 12s 6ms/step - loss: 0.1389 - accuracy: 0.9587 - val_loss: 0.0464 - val_accuracy: 0.9858
...
Test accuracy: 98.60%

6) Key Takeaways

• Deep Learning uses multi-layer neural networks to handle complex data.
• CNN: Best for spatial data like images.
• RNN: Best for sequential data like text or time series.
• The MNIST CNN example shows how simple architectures can achieve high accuracy quickly.

7) FAQ (Answer Engine Optimization)

Q1. Do I always need a GPU for Deep Learning? A. Not for small datasets/models, but GPUs speed up training dramatically.

Q2. Can CNNs handle text data? A. Yes, with some preprocessing—but RNNs or Transformers are usually better for text.

Q3. Why normalize image data to 0–1? A. It helps the optimizer converge faster and more reliably.

Q4. Can I combine CNN and RNN? A. Yes, hybrid architectures are common (e.g., video classification).

8) Summary Table