Skip to content

README.md

Latest commit

 

History

History
209 lines (133 loc) · 3.38 KB

README.md

File metadata and controls

209 lines (133 loc) · 3.38 KB

✅ PROJECT 2: DNN — Handwritten Digit Classification (MNIST)

🎯 What You Are Learning (Big Picture)

Unlike ANN:

  • Input is not one number
  • Input is an image (28×28 pixels)
  • Output is 10 classes (0–9)

Here, depth + non-linearity matters.

🖼 What Is MNIST?

Image

Image

Image

  • Grayscale images
  • Size: 28 × 28
  • Pixel values: 0–255
  • Labels: digits 0 → 9

🧩 Step 1: Load & Prepare Data

import tensorflow as tf

# Load MNIST
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()

# Normalize (VERY IMPORTANT)
x_train = x_train / 255.0
x_test = x_test / 255.0

✅ Why normalize?

Pixels become 0–1, so learning is stable.

🧠 Step 2: Understand the Model (LINE BY LINE)

model = tf.keras.Sequential([
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])

Let’s break this fully.

🔹 1. Flatten()

Before:

28 × 28 image

After:

784 numbers → [x1, x2, x3, ... x784]

✅ Converts image → vector ✅ Required for Dense layers

🔹 2. Dense(128, activation='relu')

  • 128 neurons
  • Each neuron looks at ALL 784 pixels
  • Learns simple patterns

ReLU:

max(0, x)

✅ Adds non-linearity

🔹 3. Dense(64, activation='relu')

  • Combines features from layer 1
  • Learns more complex digit shapes

This is hierarchical learning ✅

🔹 4. Dense(10, activation='softmax')

  • 10 neurons → digits 0–9
  • Outputs probabilities

Example output:

[0.01, 0.02, 0.90, 0.01, ...]

✅ Highest probability = predicted digit

📊 Step 3: Compile Model (CRITICAL)

model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
)

✅ Why ADAM?

  • Smarter than SGD
  • Auto-adjusts learning rate

✅ Why this loss?

  • Integer labels (0–9)
  • Matches softmax output

🚂 Step 4: Train the DNN

model.fit(
    x_train, y_train,
    epochs=5,
    validation_split=0.1
)

Expected accuracy:

  • Train: ~98%
  • Validation: ~97%

🧪 Step 5: Test the Model

model.evaluate(x_test, y_test)

Accuracy ~97–98% ✅

🧠 Step 6: WHAT JUST HAPPENED?

Layer Learned
Flatten pixels
Dense 128 edges, curves
Dense 64 digit parts
Output digit class

This is deep feature learning 🧠🔥

⚠️ Important Concept: Overfitting

If you increase layers too much:

  • Train accuracy ↑
  • Test accuracy ↓ ❌

Fix using:

  • Dropout
  • Regularization
  • Less depth

✅ What You NOW Understand

✔ Difference between ANN & DNN ✔ Why depth matters ✔ Non-linearity ✔ Multi-class classification ✔ Softmax + crossentropy